We learn from mistakes, and we accelerate progress through making more smart, risk-adjusted moves. My claim today: AI works fine, you’re just not pushing it far enough.

Most AI advice is too shallow. It focuses on tools, prompts, and isolated use cases. This piece is longer because the real issue is deeper: AI only works when leaders change how work happens.

We already know that AI can help you rewrite an email, build a spreadsheet, or automate the production of social media posts. But leaders need to stop using it like a productivity toy and start using it as an operating model redesign lever.

The gap between companies that are getting returns on AI and those that are writing off failed AI projects isn’t about the model capability. It’s about ambition, integration into the business’s daily operations, and leadership.

There’s a lot of talk at the moment about companies struggling to find the return on spend for their AI subscriptions and AI API usage. I think that’s because the focus has not been enough on the people and the change management side. Too much attention has been given to treating the rollout of AI like a tool provision exercise, not an operating model redesign exercise.

Subscribe now

Take smarter risks with AI use cases

Time is the dominant factor in gambling. Risk and time are opposite sides of the same coin, for if there were no tomorrow there would be no risk. Time transforms risk, and the nature of risk is shaped by the time horizon: the future is the playing field. - Peter Bernstein, Against the Gods

AI adoption shouldn’t be reckless. It can’t be too timid either. A 2024 Gartner study predicted that by the end of 2025 at least 30% of AI projects would be abandoned after the proof of concept because of poor data quality, weak risk controls, increasing costs, or unclear business value. They weren’t wrong.

At the same time, the HBS and BCG Jagged Frontier study found that consultants completed 12.2% more tasks and worked 25.1% faster on tasks inside the AI capability frontier. Consultants who over-relied on AI for tasks beyond what it was capable of struggled.

I know from my own AI testing and work over the last 18 months, the lesson is very clear: AI works, and it works well. However, this is only when you’re working on the right use case with the right model, the right harness, and the right context.

The way you take smart risks with AI use cases is to make sure that every use case has a clear owner, a measurable outcome you are targeting, guardrails around what cannot be done, and a due date.

You want to be asking, “Where are we wasting time?” or “Where are we doing things that are repetitive, context-dense, too slow, or inconsistent where we’re using manual processes when we could be automating with AI?”

When it comes to getting AI to work, I hope that you’ve learned that the real risk isn’t trying out AI and giving your people licences to use ChatGPT or Claude. The risk is running vague pilots and proofs of concept that don’t teach the business anything new, and don’t give you lessons to be learned and folded into a wider portfolio of AI bets.

Some of these bets will fail. Some of these will deliver core workflow enhancements that save a lot of time and money. More ambitious ones will start to shift the economics of your business. We’re getting AI agents to do tasks at scale and in volume. Where they are better, faster, and cheaper at doing them, it shifts the economics of your operating model and starts to push those return on spend numbers in the right direction.

Waste less time on manual processes

Manual work hides in plain sight: customer support that could be self-serviced, meeting notes that never become actions, spreadsheets moved between systems, documentation reviewed manually, and emails that exist only to coordinate work that a workflow could have handled.

AI gives leaders permission to start attacking a lot of the repetitive and low-value work that exists in every team in their business. The first generation of AI value delivery is shrinking or compressing waste, cutting the time between getting a lead and sending out a response.

Holding a meeting and taking action, getting a question and issuing the answer (all of this stuff isn’t glamorous). But inside every operating model, there are operating processes that are currently far too manual. System one doesn’t talk to system two. Even though an API key exists, no one in the business knew what it was. So, people will be downloading information into a spreadsheet, reformatting it, and then uploading it into another system.

When I think about wasting less time on manual processes, using AI in your business is partly about discovering all of this waste and inefficiency. Instead of having to wait months to do something, you can start removing waste in an afternoon.

We can redesign and reimagine how the whole end-to-end process can work. We do not need to make broken processes faster. We can start drilling into why the processes exist at all. In the example I mentioned above, just by automating data transfer between systems, some businesses could free up entire teams that are currently just doing back-and-forth data reconciliation.

Take advantage of AI leverage to create value

I think a lot of the AI conversation is focused too much on reduction: saving time, saving money, and letting you reduce headcount. A bigger opportunity is creating new value, improving product quality, improving response times, closing customer requests faster, and adding product features. All of these new value creation ideas are easier than they’ve ever been.

Saving time isn’t the finish line for an AI project, it’s an opportunity for teams that have that freedom to start running more experiments. An example in my own business has been doing A/B testing and quantitative analysis of what’s working and what’s not.

Leaders using AI well shouldn’t just have much nicer-looking dashboards, spreadsheets, and well-formatted emails with perfect grammar.

The real value comes from shortening that decision cycle, where problems get surfaced faster with richer information to help support better decisions.

For example, self-service data and analytics was a challenge for many years. Yet the lack of technical skill on the part of a lot of people meant that they were still hamstrung by what a data and analytics team could produce for them.

You can now use natural language to query and get answers to questions that previously might have taken someone in the data and analytics team a couple of days to make sure everything was right.

You obviously need a lot of testing, quality control, and data governance around it, but you can just start and learn. You can find out what works and what doesn’t, and where the data quality is good enough and where it isn’t.

When it comes to the value you can get from using AI tools properly, the last place you want to be is in more meetings and emails going back and forth. You want more outputs and outcomes. You want to be moving the dial, not creating a whole new architecture of governance and bureaucracy around AI transformation.

Become more ambitious with operating models

Most businesses are still trying to roll out AI inside their operating model and their org chart exactly as it stands today. There’s a leadership-employee perception gap where leaders think that managers are creating space for AI experimentation, trial, and error, but fewer employees think that’s actually what’s happening.

The barriers around change management and organisational culture dictate how people try things, and what the organisational response is whether things work or don’t. If there’s a blame culture that does not let people make mistakes, you just are not going to get the results, compared to an organisation with a learn-from-mistakes culture that lets people learn and move on.

The real operating model redesign opportunity isn’t about giving everyone ChatGPT access. It’s thinking about how work flows across the business.

In this newsletter and in my videos, I’ve spoken a lot about questions which aren’t really being asked yet. Why do functions exist for some value streams? Why do teams exist? If you think about the task-by-task flow for some value streams, if you have processes that move back and forth across teams with no clear gateways and quality control checkpoints, you can probably redesign these workflows with the help of AI agents. Instead of having three or four teams touching a work product, you can have one flow handled by one team. You fold the contributions from each team into the workflow itself, complete with the evals, controls, and guardrails that you desire.

Then, the specialist experts act as the quality assurance and performance improvement people, owning and remaining accountable for the delivery of that new workflow. We’ve got an opportunity to redesign everything.

If you think about realising the benefits from all of this spend, it’s obvious to see why some companies will never realise $1 of net benefit from AI spend. They have not changed how they work at all. The entire point is a complete rethink of how we do everything.

The future operating model isn’t humans versus AI. It’s a collaboration of people, AI, processes, unique IP, and knowledge, sitting inside whatever guardrails the board or company owners want to put around it.

PS: I’m running a limited offer until midnight 1 July 2026. Upgrade your Getting AI To Work subscription to an annual paid one and you’ll lock in a 30% discount today and keep that 30% for the life of your subscription. Lock in your exclusive discount now to get access to all paid articles.

5 Tips To Implement This Today

Read more

If your AI rollout has failed or is failing, this is actually a good thing, not something to shy away from.

In this article, I’ll share a playbook for how to turn your change program around.

PS - if you missed this video, you should check it out:

Subscribe now

In today’s newsletter:

1. Why mistakes build valuable knowledge

2. How teams deal with complex change

3. The playbook for successful AI change

1 - Why mistakes build valuable knowledge

When you make mistakes with AI, you are building process knowledge. You learn what does and does not work with the tools and the environment you are working in.

Mistakes build valuable knowledge for business owners and leaders because they start to understand what needs to change. Everyone can drive a successful change program that leverages AI to do things better, faster, and cheaper.

The difference between tacit knowledge, living in our heads, and explicit knowledge, written down, matters. Here is where the intellectual property supporting a business’s ability to do things for its customers in a predictable, reliable, and high-quality way lives.

If you try to do something with AI and it works, you’ve learned something. If you try to do something with AI and it doesn’t work, you’ve learned something.

Because of the cultural impulse towards more risk aversion over the last few decades, an obsession with risk management has often killed off these important entrepreneurial trial-and-error processes.

This means that many companies have spent more time and effort worrying about AI governance and AI risk management than actually figuring out how to use the tools themselves.

They need an accurate fact base on which to reason about the risks involved and how to govern them. You could not dream up a more ironic example of putting the cart before the horse.

Some companies have spent months worrying about all of these concerns before they have even let their users have access to one license to conduct experiments.

This is why if you get to the point where you have had a failed AI initiative, you have accumulated valuable information. Some companies haven’t started the trial-and-error phase of AI adoption yet.

Every little step that you have taken in the direction of making your company more AI-first is a step where you have learned and improved. You have updated information.

For example, you will know who the natural AI leaders are in your business. You will know who is advancing change and who is blocking change.

When it comes time to think about organisational design, you will have a better sense of who leaned into this opportunity and who threw up roadblocks.

Action Point: If you still haven’t let your team do some trial-and-error and experiments, with the goal of making improvements to your operating model - get that sorted this week. There are ways to reduce the risk from light AI experimentation. You should be able to find one use case and get it into production every week.

2- How teams deal with complex change

What we know about change management from the literature, and what I know from my own direct experience in the space, is that a lot of teams struggle to deal with change. The more things that are changing at the same time, the harder it is for a team to process the change. There is such a thing as change fatigue or change exhaustion.

This is what happens when there are so many little things changing across how someone does their job every day that cumulatively, after many years of this, having something as threatening and emotional as AI potentially automating you out of a job is almost too much to process.

The rhetoric in the corporate world is about AI job layoffs. Reducing headcount is one way to realise benefits from AI investment. It is disingenuous for people to claim it is all about augmentation; some spend is intended to reduce payroll expenditure.

We need to be open and honest about all this AI change and what it means for teams in reality because they have lives to live and you owe it to them to be upfront. That openness, transparency, and communication is so much more important than I think a lot of corporate leaders realise.

When you work at big companies, you know that a lot of things are scripted and rehearsed. Communication is planned. Everything feels like it has this flavour of a Hollywood production.

People see through that. They know when you are talking absolute nonsense as a corporate leader or a business owner.

What you are doing when you choose that communication style is effectively saying that your team is incapable of handling the truth. It is a paternalistic and arrogant style which does not really work in the 21st century when constant streaming information is the default life experience.

When someone in your workplace picks up their smartphone, they are extremely exposed to the algorithms and everything they see and process. If they are coming into a work environment and your style is stuck in the 1990s, that means that everything you say is being filtered as if it is curated nonsense.

You don’t have trust. You have an environment where free and frank conversation can’t happen. There are other highlights of this which make adopting AI harder than it has to be. When there are things like restricted information flows, or instructions to “please cascade this information to your team”.

Everyone gets a different story about what the agenda is, why we’re doing it, how we’re going to do it, and when we need to be finished by. This is a common theme in failed projects- what you want isn’t what the team hears, and confusion makes it harder to get the change understood let alone adopted.

This means all these programs in the AI space face an uphill battle. You are going into an information ecosystem that is already anti-AI. How do you get cut through? That’s what I’ll talk about in the next section.

PS: I’m running a limited offer until midnight 1 July 2026. Upgrade your Getting AI To Work subscription to an annual paid one and you’ll save 30% today and 30% for the life of your subscription. Lock in your exclusive discount today.

3- The playbook for successful AI change

Read more

Today I had an article planned about what AI models like Anthropic’s just-classified-as-export-controlled-products Fable and Mythos mean for AI transformation and change work in our businesses.

So, I won’t be writing about that. But look at what OpenRouter published over the weekend:

When we think about how we do things better, faster, and cheaper in our businesses, we don’t want to get caught up in “which model is the best” because this is going to keep changing every few weeks for many years to come.

Source: OpenRouter

Is this just an “Anthropic thing”, or is it a new normal which changes the vendor decision calculus around the world?

If you are a US business, you are winning as you will be able to sustain a competitive advantage over your global competitors. But what are the consequences of the lower level details of what this means? None of your overseas employees or operations can benefit from this tech if it’s export controlled, can it?

For the Project Glasswing aligned logos below, what % of their US based staff are US citizens? How will their overseas operations use frontier AI under restrictions like this? What does that operating model look like if this is the new normal?

Source: Anthropic list of Project Glasswing initial participants (there are more now)

Subscribe now

If you are a non-US business or a non-US person, you are losing, as you are now blocked from accessing the frontier as defined by most benchmarks. There is now an added challenge of - what happens to OpenAI’s next model? Is it different for OpenAI or is this now the new normal?

There are a lot of reasonable AI risk related concerns that mean some sort of argument can be constructed for why this has happened. There are also a lot of awkward diplomatic and commercial conversations that will be had around the world today if this is not resolved promptly.

I write and comment often that testing open source models and understanding how to use the tools is important. If you have been doing this, it’s not nearly as confronting a challenge as if you are locked into being an “Anthropic shop” and find yourself tools-down while you figure out how you cope with being downgraded to the obviously-worse Opus 4.8!

Share

This is why people are more important than the technology still. If your team doesn’t have opinion and playbooks around this “risk event” - why not? Are they experimenting and innovating or waiting for top-down instruction on how to respond?

You likely don’t need Fable and Mythos - in fact, they are over-powered for the majority of business tasks. But they delivered higher capability and lower error rates for a few days - so many people now feel “loss” if they run a task on a “worse” model.

I’ll keep it short - and want to hear your reaction in the comments - after seeing the elevation of AI into “national security”, and considering the implications for both US citizens and foreign nationals alike, what is the path forward for frontier models?

Where does this go and what does it mean from a corporate governance perspective where your AI-first operating model now needs to elevate independent open-source intelligence as the core to protect your business value from disruption?

Leave a comment

Regards,

Brennan

I think it would be an understatement to say that the focus in the corporate world on doing an AI transformation or adopting AI has reached a fever pitch.

The idea that you can do things better, faster, and cheaper in your operating model and achieve your corporate goals through just adopting a new technology is quite an enticing prospect.

I fear that a lot of leaders have forgotten many of the lessons we’ve learned in the realm of change management. These lessons apply to an AI transformation just the same as they do to any other project or program of work.

Tailoring beats the default playbooks

AI transformation requires a tailored change approach. It needs to take into consideration how your business works. It also needs to consider how your people feel about change and all of the different characteristics of your culture. This means how you approach an AI rollout is going to be very different.

A cookie-cutter approach where we do a pilot, a proof of concept, then roll things out with the help of a big consulting firm or a business process outsourcing firm - it just isn’t going to work for many businesses. The default isn’t going to move the dial here.

A lot of smaller businesses and startups will find that because they don’t have the baggage of bureaucracy and process, they can move faster. This means there are a lot of risks they’re introducing. These require careful management and deliberate decision-making about where you strike the balance of human and AI in a new workflow.

Rebuild fast, protect trial-and-error

The pace of change is faster than it’s ever been before. Every few weeks, there’s something new that is worth exploring. We sanity check whether it could add value, or dismiss the latest tool as mere hype. The pace of this change is far beyond what most businesses are capable of dealing with properly, let alone implementing, doing all of the testing and controls, and getting it into production use.

This new normal presents enormous challenges. Leaders want to be able to demonstrate that they’re operating close to the frontier. Unless they’ve been operating their business in this way for the last decade, they need to entirely rebuild their operating model from the ground up. It’s not that they were too slow in the past, it’s that the speed of today dissolves the quarterly cadence of yesteryear into mist.

If they want to evolve in this direction, that presents enormous risk to any business. Being able to run business operations from a position of stability is critical. You can’t be constantly chopping and changing what tools you’re using and what workflows you’re doing. Clients won’t tolerate it, and your people will get very frustrated if they have to change how they do their work every five minutes.

Imposing top-down change is hard in this environment. A lot of the value is coming from bottom-up experimentation. Your best people are figuring out, through using the tools themselves, how to apply AI in new ways to do things better, faster, and cheaper. You need to listen to them, incorporate their new solutions, and reward them.

This entrepreneurial trial-and-error process is how we get all the cool things and innovations in our world today. Keeping that experimentation and openness to novelty is one of the most important cultural things that a company going into this AI era with eyes open is going to want to retain.

People, not technology, constrain adoption

Read more

If you’re new here, I’m Brennan McDonald and I write about the people side of AI transformation. This newsletter grows through word of mouth and your recommendations. If you enjoy this, please share it with a friend today. It’s always appreciated. If you have any feedback, you can reply to this email. You can also find me over on YouTube. I’ve also started a new newsletter called AI Risk & Response where I’ll be sharing my thinking and research on AI risks and policy responses. This newsletter will continue normal service!

- Brennan

In today’s newsletter:

• What does this mean for change?

• So what about my AI transformation?

• So how do you respond?

• What do I do this week?

I’ve been thinking a lot about how attitudes in wider society about AI spill over into the workplace. What change playbooks can we run to drive better adoption and return on spend?

When you’re thinking about how you want to implement AI to do things better, faster, and cheaper in your business, part of the landscape you’re operating in is one where many of your employees spent the morning commute scrolling through a social media feed with reels where people are ranting about AI data centre water use or other sorts of politically charged anti-AI commentary.

This means when you make any AI-related announcement at work, you’re already operating inside a polluted thought space.

Share

What does this mean for change?

I think it means that if you don’t have a clear idea of what level of AI use you want in your future state operating model, and you haven’t clearly articulated how you will bring your people on that journey, you’re starting on the back foot.

A lot of the layoffs that have been labelled as AI layoffs haven’t helped this year because it’s flooded the zone where a lot of different issues are being conflated into one.

Wanting to cut jobs because of “overhiring”. Wanting to cut jobs onshore so that you can do things “cheaper” offshore. Wanting to cut jobs because you actually did deploy some AI workflows or previous automation projects that have some similar impact but no AI attached.

All three options could be sitting beneath something labelled as an “AI job layoff”. When you read or listen to what corporate executives say on earnings calls, press releases, or interviews about how they are thinking about AI, it is clear that a lot of the rhetoric in the public domain is split between two things at the opposite end of the spectrum.

On one side, you have people claiming that they are hiring more. They will need more skilled people. AI is going to create more new jobs. It is just like any other normal technology.

At the opposite end of the spectrum, AI is helping cut costs. AI is helping reduce headcount. AI is going to enable the business to scale up headcount reductions over the next couple of years.

When companies update earnings guidance, those statements carry regulatory consequences if they are materially misleading.

This all feels like “flooding the zone” where there are so many new announcements and new perspectives every day that it’s hard to know what’s going on.

So what about my AI transformation?

Read more

If you’re new here, I’m Brennan McDonald and I write about the people side of AI transformation. This newsletter grows through word of mouth and your recommendations. If you enjoy this, please share it with a friend today, it’s always appreciated. If you have any feedback, you can reply to this email. You can also find me over on YouTube.

- Brennan

In today’s newsletter:

• Frontier time vs friction time

• Why more thinking time matters

• The Mistake Mining Playbook

Frontier time vs friction time

The firms that win with AI will not be the ones that move recklessly in frontier time or defensively in friction time. They will be the ones that use structured thinking time to close the gap safely.

There’s frontier time, the fast pace, something new every day, something better, something shinier, something loudly heralded or criticised on X in volume.

Then there’s friction time. Months to years behind where the frontier is operating, subject to a slower pace of adoption, the constraints of people, process and platforms, and considerations around governance and risk management that only those who have lived through the pain appreciate.

I’ve spent most of my career working in firms where friction time rules the day. There are always reasons not to do something. Not the time, not the people, not the budget, not the right compatibility with existing operating model capabilities.

At regulated entities in particular, we have things we care about where we want humans to be accountable for decisions. A lot of this friction time in isolation makes total sense. Do we want a vibe-coded core banking platform that doesn’t get proper testing? No way!

This tension between being an early adopter and not accidentally blowing up businesses that customers depend on is a very real part of what’s going through the minds of boards and executives when it comes to AI adoption.

At the moment, I pretty much live in frontier time. I literally am using AI every day for as much as I possibly can.

I spend a decent amount each month on making sure that I test out whatever I can, burning AI API usage so you don’t have to. When I work with people, I have as one of my goals to expand their mindset for what is possible with AI.

This is why Getting AI To Work has focused on the people side of AI change. Because of the pain I’ve experienced at the frontier, I know that for people living in friction time, there is so much pain relief that is possible if you clearly think about an AI-first operating model in your business.

There’s a reason why I say you need to start shipping AI use cases in production that your people can use to do things better, faster, and cheaper every week. You need to do whatever you can to reduce that gap between frontier time and friction time.

Does this mean you need to deploy a model to your people the day it comes out? No, but it does mean you should have someone in your business scanning that horizon. Scanning what is happening in real time, checking, testing, evaluating and recommending what the next pivot in your business’s AI adoption journey is going to be.

Share

Why more thinking time matters

I’m not referring to thinking time when an AI model is generating its response to your query. I’m thinking about thinking time. You step away from your computer. You step away from the emails. You step away from the meetings. You actually reflect on what you want your AI-first operating model to look like.

The idea of path dependence where choices you make today have long-running impacts has never been more real to my mind. How you choose to respond to AI today as a challenge and opportunity could mean life or death for your business in a few years’ time.

These sorts of decisions can be expensive to reverse if you make the wrong one. If you picked the wrong mix of people, process and platforms and don’t have the right talent helping you achieve your transformation, you could be permanently locking in structural disadvantages.

There are clearly risks in AI that you need to manage. A lot of complexity in the AI space is about deciding where you want to automate and where you want to keep humans in the loop to avoid outages or mistakes leading to reputational damage or fines.

Thinking through all of these things and figuring out what you really want that operating model to look like, especially as an intelligent firm where you’re pushing more decisions and day-to-day activity down to AI agents, takes time and effort.

So some leaders can be forgiven if they’ve been moving slowly. The best ones have already figured out how critical this moment actually is. I’ve often referred to AI acting as a ray of sunlight. It reveals the cans you kicked down the road.

You’re now at the point in many boardrooms and executive meetings where AI pilots start revealing the consequences of some of these cans that have been kicked as far as you can get away with in the AI era.

You need to put in the thinking time to design your operating model in an AI-first way. Thinking time is about going from frontier time to friction time and not ending up in the wake of AI-first startups in your niche.

In the next section, I’ll share how you can run this playbook in your business today.

The Mistake Mining Playbook

Read more

Hi there,

If you’re new here, I’m Brennan McDonald and I write about the people side of AI transformation. This newsletter grows through word of mouth and your recommendations. If you enjoy this, please share it with a friend today, it’s always appreciated. If you have any feedback, you can reply to this email.

- Brennan

Google just released their new model, Gemini 3.5, and it’s a good model.

I like the fact that there are more than just two US competitors in the AI realm.

It’s clear that in the scientific and multimodal space, Google DeepMind have real strengths.

Do you need to care about Gemini 3.5 though?

Only if you want to spend the time and only if you have already implemented a number of AI use cases in your business.

I published a video yesterday on my YouTube channel where I spoke about reasons why AI projects fail. You can watch it over there:

One of the things I spoke about was the need for people who are leading AI transformation or sitting in roles like “Head of AI transformation” need to be testing things themselves and developing taste and opinions about which models suit which tasks best.

Subscribe now

Why getting use cases shipped matters more

I think that getting use cases shipped matters more than knowing what every model does.

I’ve started to realise how my own technical optimism makes me overestimate how interested leaders and business owners are in the low-level technical details.

You definitely need to become more technical and have an awareness of all of this stuff. I realise that a lot of people are not going to be running their own private evals every time a new model gets released.

That’s why one of the easiest ways to test out a model if you want to is to try and build something and ship it.

The relaunch of Google’s agent tool called Antigravity is one opportunity to test this out. The move away from CLIs for prosumers and enterprise users towards the desktop apps like Claude Desktop, Codex and Antigravity is definitely a trend at the moment.

One improvement here is the addition of scheduled tasks, and better use of subagents.

Share

So what do I actually do with this?

1. Go check out Gemini

2. Share a business problem you face with it

3. Ask for an action plan to solve it

4. Assess whether it is any good at all based on your subjective opinion

5. Go back to work on transforming your operating model to get more out of AI

There’s really not much to it. You could, if you’ve been working with Claude Desktop or Codex, download Antigravity and give it a go for some agentic tasks or setting up automations on a schedule, but chopping and changing models is probably not the right thing. Especially if you are working in a small business where it was hard enough to get your people to accept that they even need to be thinking about using AI. Let alone learning a whole new tool.

I’m looking forward to seeing what 3.5 Pro looks like. I think that for businesses that are already in the Google ecosystem, more of them are going to get more than enough of what they need from the Gemini model family. That is good from a competition perspective.

The reality is that for almost all business use cases today, if you break them down clearly and coherently into their component tasks and subtasks, a model like Gemini 3.5 Flash is more than good enough.

I think this point isn’t well enough understood, and it’s one of the reasons why I changed this newsletter’s focus to really be about the people side and change management side of AI transformation.

If you break down your processes clearly, you can get them done to a required level of quality and cost with a model like 3.5 Flash. We have almost been gifted the luxury of not having to worry about clearly articulating exactly what we want out of some of our AI workflows because of the capability of models like Claude Opus 4.7 and ChatGPT 5.5!

The most important thing for you to do with new things coming out every week is making sure you are getting at least one use case into a production environment. Leverage AI every week - your delivery matters more than the never-ending hype cycle.

Leave a comment

Regards,

Brennan

PS: If you liked this shorter format, drop a comment below or send me a DM.

If you’re new here, I’m Brennan McDonald and I write about the people side of AI transformation. This newsletter grows through word of mouth and your recommendations. If you enjoy this, please share it with a friend today, it’s always appreciated. If you have any feedback, you can reply to this email.

- Brennan

In today’s newsletter:

• The buy-in problem is misdiagnosed

• Evidence creates buy-in, clever words don’t

• How to apply this to your business

The buy-in problem is misdiagnosed

There’s a standard change management playbook. It’s tempting to roll it out when you’re doing an AI transformation.

However, if there’s one thing I’ve learned over the last year and throughout my corporate career, it’s that you need to tailor your approach to the situation you’re actually facing.

A lot of the standard change management collateral can be easily generated with the help of AI. You can make a stakeholder map. Prepare a comms plan, prepare the training course content, draft emails and communications with stakeholders and sponsors.

It doesn’t mean you should. In fact, one trend I’m starting to see is a lot of AI-generated content working its way into corporate change program collateral and being exposed as soon as the problems begin.

As a discipline, change management has learned an awful lot and incorporates a lot of the lessons of program management, project management, systems engineering, and general business practice.

It was built to solve the challenge of getting people to do new things in new ways when they would much prefer to keep doing things the old way. When it comes to AI transformation, I think the buy-in problem is getting misdiagnosed. That is one reason why I worked on the 5C framework and shared it in this newsletter.

All of this change around AI in the workplace is happening when we have unresolved policy problems to solve in the personal domain. If your worker doesn’t want an AI data centre built in their neighbourhood and buys into a lot of anti-AI rhetoric, they’re not going to lean into working on an AI transformation of how they do their daily tasks, are they?

A lot of the pushback around AI in the workplace is in regard to things like compliance with regulations, the quality of outputs, or what things mean for job security. A lot of these are reasonable issues for employees to raise.

It doesn’t help when you have a global group of corporate leaders who have turned AI job layoffs into a mantra. Whether the reason is because an AI automation workflow was deployed or because they hired too many people during the pandemic doesn’t really matter.

There’s an enormous amount of fear. When you add it to the fact that a lot of this is going to impact knowledge work jobs that have already suffered from enormous amounts of both disclosed and undisclosed offshoring and de-skilling, the scale of the challenge from the change management side is clearly apparent.

There is a difference between irrational resistance to change and rational threat detection to your own well-being and safety. In this article, I’m going to argue that you have to show how AI works and improves your team’s experience in the workplace instead of rerunning played-out persuasion and communications playbooks from the pre-AI era.

Evidence creates buy-in, clever words don’t

Read more

If you’re new here, I’m Brennan McDonald and I write about the people side of AI transformation. This newsletter grows through word of mouth and your recommendations. If you enjoy this, share it with a friend today. It’s always appreciated. If you have any feedback for me, you can reply to this email.

- Brennan

In today’s newsletter:

• Everyone is being asked for an AI story

• What AI does to the work you already do

• Where AI change comes unstuck

• Where to start, and what to do next

Everyone is being asked for an AI story

All your stakeholders have an opinion about AI. Give everyone access to Claude, get a consulting firm in to map out all of your existing processes, outsource your operations overseas, or vibe code a new CRM.

In 2026, you definitely need an AI story. It can’t be about what other people are doing. It has to be what suits you and your business and the industry you operate in, inside all of the regulatory constraints and industry practices you have to get right.

There’s a gap between what you’re doing today and what is possible with the help of AI. I often see comments like “I can’t think of ways to use AI in my business” or the like.

This sort of thinking is a tragedy. If you’re struggling to come up with use cases, all you need to do is turn on the voice command. Dictate into your favourite AI tool, tell it about your business and ask about basic use cases.

It really is that simple. For example, one client did not know how to connect one application with another. They were still copying and pasting. It turns out all they needed to do was ask for developer access, get given an API key, and then suddenly they could easily query and access their data which had previously been locked up.

There are efficiencies lying on the floor in every business in the world at the moment. A lot of hardworking and clever business owners just don’t have the time to even think about which questions will get the most out of an AI agent if they want it to partner with them.

AI is a people problem, not a technology problem. What this means is that if we’re leaders and we’re business owners, we can’t just delegate and outsource the thinking around what AI change needs to look like in our business. We need to put in the work.

The risk when you have the pressure to tell a story about how you’re using AI in your business is that you fall prey to what I’ve called cosmetic AI. Everyone gets a ChatGPT Enterprise licence.

Maybe you put in some workflows which reduce manual effort on repetitive tasks. You might even connect system one to system two, which means you don’t need a team which used to manually reconcile the two anymore. A lot of this stuff could have been done even before the AI era.

I think a lot of this AI washing is partly people finally realising maybe we could use technology to help us solve problems. The intelligent firm is somewhere a lot of companies are quite far from on their journey.

They need to do a lot of deep thinking about their operating model as it stands today and where it could be with the fullest application of the multiplier effect of AI before they are in a position to really break through.

What is your AI story going to be if vast parts of your operating model as they stand today have no reason to exist in the AI era?

Share

What AI does to the work you already do

An operating model is people, processes and platforms. As we move into the AI era, firms where more of the thinking and execution is done by AI agents are building on top of the foundations of what has gone before.

If you remember the cloud migration era, there were two ways that businesses dealt with this. The first approach was called lift and shift, where you’d take an application and just deploy it in the cloud with limited reflection on what refactoring or redesign was required to optimise things like cost, security and efficiency.

The second was by doing a process to make what was being deployed to the cloud more suitable. This involved reflecting on the overall architecture of the business to make sure that you didn’t have absurd situations that existed in on-premise data centres being perpetuated into the cloud era.

For those of you who’ve worked on projects like this, yes, I’m simplifying. For the purposes of this general business audience, I think it is a fair reflection of what the two main approaches were.

The first approach of lift and shift led to outcomes like security breaches. Bill shock. And a level of complexity which did not use a lot of the native functions of the cloud platforms like AWS.

The second approach of redesign led to some of the first movers achieving much more efficient and cost-effective operating models. These were not only more resilient, but were in a position to take advantage of all the incremental additions to cloud platform functionality as they were introduced.

We can think of AI as a multiplier on the operating model you already have today. You already have strengths and weaknesses, where the problems are in your business, and where the pain points are.

You need to be open with yourself if you’re thinking about AI transformation in your business. If you have a weakness, maybe you never put enough into cybersecurity, or maybe you never put enough into data governance and data quality.

If your operating model is already this brittle, AI will widen every crack you have been ignoring. Where to start matters more than which tool you pick.

Where AI change comes unstuck

Read more

In today’s newsletter:

• The Anthropic and SpaceXAI deal

• Why the theory of the firm needs an upgrade

• The playbook for building an intelligent firm

The Anthropic and SpaceXAI deal

If you’ve been using Claude, Claude Code or Claude Cowork over the last few months, you know that at times it just doesn’t work.

The rapid growth in use of Claude has led to enormous pressure on Anthropic’s ability to service that demand with the supply of compute it has access to.

As you can see below, in an industry where multiple nines is considered good uptime, Anthropic’s product suite has not had good uptime.

Dario Amodei revealed that in the early part of 2026 they were experiencing 80x growth when they had planned for 10x growth.

Times like this are completely insane. You have to give them some indulgence. OpenAI is not in the same position. Nor is Google. From the perspective of enterprise customers, uptime at that level would immediately disqualify most vendors from even being in contention or getting past procurement hurdles.

Fortunately for Anthropic, they are selling one of the best implementations available of general intelligence. They can be consumed across a wide range of products or via API endpoints. They charge a premium. This is especially true for their top model, Opus 4.7. Their recent marketing of the Claude Mythos model has caused a panic in governments, corporates, and regulators around the world.

I think a lot of people forget that if you go beyond the headline, a lot of the stories that get written about AI miss the wood for the trees. ChatGPT 5.5 is already benchmarked to have similar or greater capability on cyber than Claude Mythos.

For enterprise customers, especially big ones, despite the preference of developers and non-technical users alike for Claude, what we’re seeing here is a need for a multi-vendor strategy.

How would you feel if the chart below was powering most of your operations? You would be pretty filthy. This is especially true if you are in a regulated industry. Any outage at all needs to be reported to the regulator. It also comes with a lot of issues. This includes the consequences of making those sorts of notifications.

All this background is important to understand what’s actually going on when Anthropic does a deal with SpaceXAI. They are growing so fast that they need compute and they do not care where they get the compute from. Almost all of the deals they have announced so far in 2026 are for capacity. That capacity is yet to come online with other partners.

SpaceXAI effectively had an under-utilised capital asset that they had spent a fortune building and using, sitting ready and as a “real option”. The deal Anthropic struck gives them access to the entire Colossus 1 data centre with something like 220,000 GPUs of various types originally focused on training earlier Grok models.

I was astonished to see how bad so many of the takes were on X around what this means for SpaceX AI, what this means for Anthropic, and what this means for Grok. Colossus 2 is an even bigger and better data centre where more Grok models are still being trained. Some analysts have estimated that the $5 to $6 billion per year in GPU-rental revenue that Anthropic is likely to be paying SpaceX AI will roughly offset the annual costs of SpaceX AI itself.

So what’s actually happened? Anthropic gets a short-term immediate boost to their compute capacity. This is why they are able to slightly ease the limits. They are still facing limits, though, due to the enormous continued growth of their business.

SpaceXAI is part of SpaceX. It is going to IPO shortly. It will get billions of dollars in high-margin revenue. We will soon know through the SEC filings roughly what Anthropic are paying. SpaceX will have to disclose a whole new category of hyperscaler revenue.

This will attract a different multiple from other parts of SpaceX revenue. It will also give more debt servicing capacity to build things like the TerraFab. Or more data centres for SpaceXAI.

What does this mean for AI transformation? You need to understand what is happening at the top of the supply chain. This is important if you are going to build a business that is sitting on top of Frontier Labs building the AI models and an ecosystem of data centres serving the compute that you rely on.

Just like you would expect your chief technology officer to be well-versed in every single little thing that is critical and moves the dial for your AWS, GCP, or Azure reliance, leaders in business need to know what is happening under the hood of whatever API endpoint you are relying on. Who is serving their model, and where are the dependencies?

Why the theory of the firm needs an upgrade

Understanding the physical reality of this supply chain is the first step, the next is restructuring your firm to actually capitalise on it.

An operating model is just people, processes and platforms. The coordination of all these factors used to be a heavily human task. The theory of the firm is a concept that I have written about extensively over the last year.

As part of that research, some of the thinking I have done is about what changes as AI capability rises. The need for verification and accountability rises, and a firm faces a choice between allocating tasks to humans, humans using AI, AI agents themselves, or external suppliers.

There are a number of economists doing good work in this space. One important thing to remember is that there is a lag between how fast AI is changing and the pace at which widely held beliefs in a profession like economics update through the process of journal articles, peer review, and conferences.

The intelligent firm, where more and more decisions are removed from human oversight to decision-making by AI agents inside guardrails placed around them, presents a real dilemma.

All of this might seem a bit over the top and not worth thinking about if you’re running a function, running a business or an AI project. Yet, I think what is being revealed when people have simplistic reactions to the SpaceXAI-Anthropic thing, saying things like “Grok is over” or “Anthropic shouldn’t have given money to SpaceX,” is the next level down of thinking about how the ecosystem of your supply chain works.

We’re all building on AI API endpoints, and as part of the duties we owe, we need to be going deep. We cannot be cosmetic in our thinking and application of AI.

In the next section, I am going to share some more of my thinking. This thinking is around all of the different levers that a firm has in its operating model. The goal is to become more intelligent. It is moving that decision-making and strategic layer from human to machine, and getting the absolute most value from AI and technology.

The playbook for building an intelligent firm

Read more

There is a mood I see on social media at the moment. You can see it in the defensive reactions to Coinbase’s 14% headcount reduction yesterday.

It’s an attitude which I’m calling, the desperation for AI to not be real.

The CEO of Coinbase, Brian Armstrong, announced layoffs of 14% of Coinbase employees. It’s not pleasant for those impacted. Hopefully they are able to find something suitable and soon.

In the X post below, he outlined some of his thinking. I have not seen this level of attempted pushback to an AI announcement. Even the Block AI layoff announcements that were made a while back did not get this level of shade.

Why simple stories about AI change don’t work

As humans, we love to tell simple stories. CEOs of publicly listed companies have incentives to tell simple stories that are easy for investors to understand. That doesn’t mean they’re wrong. It doesn’t mean that when they talk about how they want to change their operating model, that they’re not serious.

One of the more simplistic stories going around today is this is all just another case of company underperformance and over-hiring in the low interest rate era. They’re jumping onto using AI just to cut costs and look better.

I think the story is getting played out by events in the real world as they unfold. When you take a step back and actually think about how things work, you know that many different things can be true at the same time.

The simple story is often glossing over lots of other components which need to be at least considered and evaluated on an equal footing with your preferred simple story before you really start drawing conclusions about what’s really going on.

When I read what Brian Armstrong is saying, what I am hearing is a lot of focus on centring the people strategy around people who are already or can use AI to do things better, faster and cheaper.

It also reads as if they’re doing quite serious experimentation with exactly what their future operating model and organisation chart is going to look like. That’s exactly what you want to be doing right now.

One aspect of his post that has had a lot of pushback is the no “pure manager” comment. I think he is being reasonable. Why? As I’ve written about many times over the last year, the operating model of the future is moving away from human coordination to AI agent coordination.

You need your management layer to be technical. They should take advantage of, guide, and apply their taste and experience to the work being done. They also need to curate and coach the remaining human workforce.

This means people who just want to do their one-on-ones, send emails and sit in on endless meetings don’t really have much of a role to play. When CEOs are merging PRs, we should be expecting everyone to be much more hands-on.

The other part of the post which is getting a lot of pushback is the comment around non-technical people now vibe coding things to production. Coinbase is a publicly listed company. It is also regulated.

I would expect what is happening under the hood to be much more controlled: a non-technical person can create a PR, but it only ships after CI/CD, testing, security and code review gates.

Brian Armstrong has already responded to some of the comments with this:

Why do people want to put AI back in the box?

People don’t like change. One of the central lessons of change management is that if people don’t feel like they have a clear sense of where they fit in and how they benefit from a change, they’re going to fight it.

I can’t help but notice that a lot of people still have a reflexive desire to minimise what AI is changing. Any news item or announcement or improvement in the state of the capability of these tools is met with so much negativity.

The reality is that our world moves forward through the actions of live players. The CEOs of public companies have chosen primarily to lean into AI and use it for work and for achieving their other capital markets goals.

In the game they’re playing, it doesn’t really matter what people on the timeline think. They are choosing to take action. This is based on up to dozens of underlying reasons or results from their AI experiments.

Many public company CEOs, especially founder CEOs, have realised something. They need an AI story. This story must align with how they use AI. It should show how AI makes things better, faster, and cheaper.

If they don’t have a story that aligns to how they are using AI to do things better, faster and cheaper, their access to and support from the capital markets is at risk.

Public company restructures rarely happen for no reason at all. A restructure like this would almost certainly have been tested, planned, workshopped, brought up at board level, and thought through carefully.

This is just the latest instalment in something that we’re going to see more and more. What I find interesting is that there’s even an argument to be made that the 14% headcount cut versus Block’s 40% headcount cut is perceived as not aggressive enough by the street.

What do people strongly rejecting this framing and claiming that it’s the end of Coinbase think is going to play out over the next few years?

There is still a need for skilled, capable people. The leverage of AI tools means they can do more things. They can do them better, faster, and cheaper. It also means that anyone not pulling their weight is more exposed.

There are definitely some potential downside risks here. Things need to be managed and mitigated. Every business can’t just cut without making sure critical tasks are transferred, or automated, or removed from the operating model.

What I would caution is thinking you can go slow and implement some form of gradual and staggered incremental change. You need to start off with the experimentation. Learn from mistakes. Bring people along on the journey.

At some point, you are going to hit a wall. I think this is where some of the stories that experts tell around new jobs created and old jobs destroyed start to come off the rails.

Many big companies likely still have excess headcount, but AI changes the calculation around what can be automated, simplified, transferred, or removed. You might be telling a story that AI isn’t the reason for the layoffs, it’s just excess hiring from the pandemic times. If you take into account AI’s capability to improve efficiency further, even more excess headcount could realistically be cut.

This is why the simple story is a real challenge. We can’t undo what is already happening in a hyper-competitive marketplace. If we aren’t thinking about how to radically reorganise and redesign our operating model to take the maximum leverage from AI, we’re at risk.

We also need to leverage our skilled and experienced people. They bring taste, curation, relationships, knowledge, and expertise. There is still a transition period. A lot of these people are not who you want to be including in your “AI layoffs”.

Below, I break down Armstrong’s memo line by line and translate it into operating-model questions and actions for your business.

How should I think about this in my business?

Read more

If you’re new here I’m Brennan McDonald and Getting AI To Work is about the people side of AI transformation. Because it’s not a technology problem, it’s a people one. Paid subscribers get access to a library of diagnostics and playbooks to apply to their own business and deeper insights.

In today’s newsletter:

• Enterprise AI spend will keep rising and needs to pay off

• Why change management matters more than ever

• So what’s stopping you from getting AI to work?

Reply to this email or drop a comment or DM with feedback.

Share

Enterprise AI spend will keep rising and needs to pay off

How much is your business spending on AI each month? What is your AI budget for 2026?

Last week I shared the story of how Uber has already “blown” the budget it allocated for 2026 because of the insatiable demand for AI tokens especially from engineers.

When you add that to the enormous numbers of CapEx guidance that the Big Tech hyperscalers are planning, over USD700 billion for 2026, it’s clear that an enormous amount of money is going into enterprise AI spend across the whole stack.

Revenue from providing AI via the cloud platforms to enterprise clients is growing rapidly. The change in the details of the agreement between Microsoft and OpenAI has led to AWS Bedrock now offering the OpenAI models.

You can see the demand management strain in premium-model pricing, tighter subscription limits, and providers policing heavy or non-compliant use of flat-rate plans.

All of this combined is why my claim is that this isn’t a bubble. This is a massive infrastructure build-out. If you read any of the Wall Street research reports or independent research it goes into immense detail. It’s not just the AI labs. It’s the data centres, the power suppliers, and the chip manufacturers. It’s all of the associated products and services with servicing one goal: more compute.

This is all progressing much faster than any other technology. What this means is that some of the mental models we had around previous tech waves are probably wrong.

There’s another aspect of the cost of AI use which I think is massively underexplored. And that is the use of open source models that are much cheaper than the closed source models. My own experimentation has enabled me to build workflows that use cheaper models and get better outcomes through more structured prompts and context management.

That ability to much more granularly allocate a specific model for a specific subtask in a process means that the opposite of this enterprise AI spend is that a lot of people are going to be able to do more granular cost management to be able to keep up.

Another example is my recent testing of DeepSeek V4 Pro. It’s nearly at the frontier. If you read anything in the news claiming that Chinese open source models are years behind US models, you’re getting an outdated picture. The most powerful US models are still the best, but that margin is shrinking. For most daily business tasks, you probably don’t need Opus 4.7 or ChatGPT 5.5 Pro.

Again, what is your AI monthly spend or AI budget? Think about that number.

Why change management matters more than ever

Read more

This week’s article wraps up a 12-week focus on building out the 5C framework. There’s definitely a gap in much of the AI content I see and read around the human side of AI. That’s one of the reasons why I refocused this newsletter last year around the people side of AI transformation and change.

There’s more to an AI transformation than getting all the ChatGPT licenses and controls setup or making sure that everybody does their e-learning modules on time. The longer I’ve been in this space, the more I realised that there’s another area of content or education I need to share with my audience around what is possible with AI from an operating model perspective.

This will build on the change management focus here. It will give more usable diagnostics, playbooks and templates for you to unlock the next chapter of your AI transformation.

Start here if you are new to this series: Your AI Training Isn’t Working. Here’s Why.

Share

The Pattern Behind the Quiet Failures

Somewhere in most organisations, an AI initiative is stalling. Not in a dramatic way. Not with a public failure. It is just slowing: launch energy gone, timeline slipped three times, the team meant to own it pulled onto other things, and nobody willing to say out loud that it might not work.

Most of the time, technology is not the part that breaks. The technical capability is increasingly available. Something simpler is missing: people have not agreed on why this matters, what should happen next, who gets to decide, or what success would look like.

The ones that succeed run adoption as a system rather than a project. Over twelve weeks, this series has mapped that system through the 5C Adoption Friction Model: Clarity, Capability, Credibility, Control, Consequences. Ignore any one of them and it pulls the initiative off course in a way the dashboard will not show.

Find the friction. Remove it. Move to the next. This week, the system.

The Five Friction Points Condensed

Clarity is about whether everyone is working from the same idea of what this is for. Your CFO thinks it will save headcount. Your Chief People Officer thinks it will lift the team’s skills. Your product team thinks it will ship new features. Everybody below them gets a different story and picks whichever version keeps them safe. You end up with a pile of projects pulling in different directions.

Capability comes down to skills. You have an accountant who has spent ten years in Excel, and you cannot expect them to start building advanced AI workflows tomorrow. Not without training. Not without coaching. Not without some permission to slow down while they learn. People try anyway, fail, and blame the tools. Adoption stalls, confidence drops, and your would-be advocates stop volunteering.

Credibility runs on trust, which is fast to lose. Leaders talk transformation but never put a budget behind it, set deadlines nobody can meet, and quietly pull someone off the project the second things get hard. Each move tells the team this does not really matter, and eventually your people stop believing anything you say about change.

Control sits at the level of power. People worry about their jobs, role changes, how their data is used, whether anyone will ask how this should work. They will not say any of it out loud. So if nobody brings it up, the pushback shows up later, slow and quiet, and your most influential people end up blocking the work.

Consequences turn on the stakes for people. When adoption is optional, early adopters get no recognition, and there is no real cost to waiting or reward for moving, the rational thing is to do nothing. Six months on, you cannot point to any results, justify what you spent, or hold onto the budget for what comes next.

What All Five Have in Common

Even the skill problem is not only a skill problem. People are not refusing the tools. They are protecting who they are at work: their role, their authority, their standing. Underneath, you are looking at identity rather than just competence. Train through it and the training does not stick. Mandate around it and you make the resistance worse.

Most leaders need shared purpose before anything else holds, not because Clarity is always the deepest friction but because the other four cannot work without it.

The Diagnostic Sequence Principle

Most leaders go for the loudest friction, not the one that is actually in the way, and that is how things get worse.

Fix Capability before you have Clarity, and you train people for work that does not matter. Fix Credibility with more communication and you make the trust problem worse. Throw incentives at Control and people hear a threat. Fix Clarity without Consequences and everyone nods, then goes back to what they were doing. In the wrong order, the problem grows. In the right order, each fix makes the next one easier.

The Condensed Diagnostic

This is not enough to fix the initiative on its own, but it will show you where to start. For each friction point below, ask whether the description fits your situation.

Clarity

• Can three different people in your organisation explain the purpose of your AI initiative in the same way?

• Do your frontline teams know what they are expected to do differently, or are they waiting for clearer direction?

• Has anyone said explicitly what problem this initiative solves, and for whom?

Two no answers means this is an active friction point.

Capability

• Do the people expected to use AI tools have time and structured support to learn how to use them?

• Has anyone measured the skill gap between what people have now and what they need?

• Are people learning through trial and error, or through formal training and coaching?

Two no answers means this is an active friction point.

Credibility

• Have leaders protected this initiative when it became inconvenient, or only supported it when it was easy?

• When people express doubt, do leaders address it or dismiss it?

• Have the people driving change made themselves available to listen rather than only to broadcast?

Two no answers means this is an active friction point.

Control

• Have you explicitly named the decisions that will change, and who will make them?

• Are the people whose authority is shifting involved in designing how that shift happens?

• Are people saying “this is a good idea” in meetings but failing to change behaviour afterward?

Two no answers means this is an active friction point.

Consequences

• Are there immediate consequences for adoption, and is it optional or required?

• Are early adopters recognised and rewarded, or does the organisation treat adoption as invisible?

• Is there a real cost to waiting, or can someone do nothing and face no consequence?

Two no answers means this is an active friction point.

Run the diagnostic and write your primary and secondary down. Whichever friction is blocking the others is your primary. The rest follow in sequence. Diagnosis matters. Get the order wrong and it does not work.

From Diagnosis to Action

By now you can see the pattern, and what happens next depends on whether it makes it into your next planning meeting or just sits in a document while another quarter passes.

Below is the Complete 5C Adoption System: the sequencing logic, the 90-day reset, and the conversation framework for the hardest friction.

Read more

If you’re interested in AI, it’s important to remember that most people on the planet don’t care about it. They actively hate it. The people who do care are about to learn the next lesson, and it isn’t a technology lesson.

Last year I wrote a number of pieces on this newsletter about lessons from economics, the theory of the firm, and how they might apply in the AI era. One of those was on operating model compression. The thinking at the time was that a lot of companies would realise the human coordination cost could be replaced through AI agents making better, more efficient decisions.

How is that turning out so far? It turns out many barriers to getting results from AI have nothing to do with the technology itself. It is all about the people and culture within the organisation trying to roll out an AI transformation. Many processes, controls, and ways of doing things that worked well in the pre-AI era are anti-patterns or obstacles in the moment we now find ourselves in. This means a proportion of AI projects are essentially doomed from inception.

Subscribe now

Take Uber.

In April 2026, Uber’s Chief Technology Officer Praveen Neppalli Naga said publicly that the company had exhausted its entire 2026 AI usage budget. Not because of GPUs, hiring, or new data centres. Almost entirely because of inference costs on AI coding tools, mainly Anthropic’s Claude Code and Cursor.

Uber gave around 5,000 engineers access to these tools in December 2025. By March, 95% of engineers were using them monthly. 70% of committed code was AI-generated. 84% were using agentic features. Per-engineer monthly costs ran from $500 to $2,000 in API tokens alone. The CTO described the situation as being “back to the drawing board” on AI spending plans. Nvidia executives have made parallel comments about teams where compute costs now exceed employee salaries. Uber became the poster child because it was honest about it, but it is not alone.

Forget the overspend. The point is what it proves. When the tools are useful, the consumption curve does not match any planning artefact you have. The bill is large because the savings are larger. The budget was the visible symptom. The org chart was the actual problem.

Why? When you think about what a business is and why a business exists, it exists because it is better to put together a bundled operating model of people, process, and platforms in one firm than to contract for every single output. The bundle is worth more than the sum of its parts because of coordination overhead. That overhead is the variable AI eats.

Read more

Start here if you are new to this series: Your AI Training Isn’t Working. Here’s Why.

The hardest part of a reset isn’t the planning. It’s the moment of saying out loud, in a room where it matters, that the current approach isn’t working.

That moment is uncomfortable for a reason. The training programme has a budget owner. The platform has a steering committee. The roadmap has been signed off. Naming that the direction is wrong puts earlier decisions on the table, including some you made. It feels personal in a way most business decisions do not.

Name the discomfort. It is what keeps most initiatives running long after they should have changed course.

The cost of not resetting is rarely visible on a dashboard. It shows up as another quarter of the same plateau, another round of training that does not stick, and the people closest to the work going quiet. You lose momentum, energy, and the trust of the people whose adoption was supposed to make this real.

The Cost of Continuing

The word “reset” carries weight. In technology it usually means something broke and had to be turned off. In leadership it can feel like an admission of error.

A reset is not an admission that the AI initiative failed. It is an admission that the response has been aimed at the wrong friction point.

A reset is the decision to stop investing in a friction point that is not in the way, and to start investing in the one that is.

Training does not fix cultural friction. Tools do not fix a credibility problem. Mandates do not fix a clarity gap. When the response and the friction are mismatched, the response can grow forever without moving anything. What it costs is the time, money and attention that could have gone somewhere else.

Resets that work tend to start in a similar way. A leader stops defending the current approach in the rooms where decisions are made. They name what is actually getting in the way. They redirect resources at it. The political cost of that moment is real, and it is finite. The political cost of avoiding it grows for as long as the wrong response keeps being funded.

What to Say in the Room

The reset needs a clean sentence, not a speech.

“We have been investing as if this is a training problem. The evidence suggests the friction is somewhere else. Before we spend another quarter expanding the current approach, we are going to diagnose what is actually blocking adoption.”

A sentence like that names the mismatch, avoids blaming the people who built the current plan, and creates permission to look at the problem again without pretending the previous work was worthless.

Do not over-explain. The more a leader defends the reset, the more it sounds like a failure. Say what has been learned. Say what will change. Move the conversation to evidence.

Phase One. Diagnosis. The First Four Weeks.

A reset starts with clarity. Not planning. Not solutions. Clarity about which of the five friction points is the one in the way.

The 5C framework names them: Clarity, Capability, Credibility, Control, Consequences. They ask whether people understand the change, can do the work, believe the priority, feel agency over it, and have a reason to move. The work of Phase One is finding which one matters most for this organisation, right now.

Week one. Map the signals.

There is more data already in the organisation than most leaders realise. Which teams turned the tools on. Which groups finished the training. Which departments responded to the strategy email and which did not. Where AI is being used informally, even in pockets the central programme has not noticed.

The work of week one is collecting, not interpreting. The signal is more specific than “people do not care.” The sales team is using the assistant. The operations team never logged in. One product squad has built a workflow nobody asked for. That texture is what matters.

Week two. Interview the friction.

Talk to people who adopted and people who did not. Not a focus group. A short, specific conversation. What would have made this easier. What got in the way. What was tried and abandoned.

Adopters describe what removed friction. Non-adopters describe what created it. Both are useful. The pattern to listen for is which of the five Cs surfaces in almost every conversation, even when nobody uses the word.

Week three. Analyse the pattern.

After eight to twelve conversations and a map of where adoption sits, a pattern emerges. It will not be “everyone struggles with X.” It will be more specific. The teams that adopted are high-agency teams in early-stage work. The teams that did not are in compliance-heavy functions where there is no cost to waiting and no reward for moving. The friction is Consequences, not Capability.

Or a different pattern. The people using AI tools have had a conversation with their manager about it. The people who have not, have not. The friction is Credibility, not tooling.

There is usually one primary friction point. Secondary ones exist. Moving the primary one tends to create space for the others to move with it.

Week four. Confirm and commit.

Take the diagnosis to the leadership team. Show the evidence. Do not propose the solution yet. The work of week four is agreement on the friction point itself. Once that agreement is on the table, the conversation about response becomes much easier, because the question has narrowed.

At the end of Phase One there is a named friction point, evidence behind it, and a decision from leadership that this is what the next phase will tackle. Everything after this depends on that foundation.

The full 90-Day Reset Sequence continues behind the paywall: weekly activities, deliverables, leadership prompts, and checkpoints for Phases Two and Three. Subscribe.

Read more

I used to waste a lot of time with lots of different browser tabs open, trying to stay across everything. Now I just get custom briefings on whatever I am interested in, delivered on a schedule. When I need to go deeper on something, I go off and run a deep research tool to produce a detailed report, drawing on a wide variety of sources including non-English ones. It is a much more targeted use of time than spending hours browsing the internet.

What is happening here is not a productivity hack. It is a shift in how information moves. For most of my corporate career, work was push-based. Slack pings, email notifications, CC lists, forwarded articles, the endless stream of things-possibly-worth-reading. Personal agents invert this. You pull what matters when you want it, and you are interrupted only when something actually requires a decision or an action, and get pushed what you built for your own needs.

Even a year ago this was hard to set up. Now it much easier with tools like OpenClaw or Hermes Agent. But pretty much no one I know in corporate roles actually works this way yet. They are still drowning in the push stream, still manually copying and pasting across systems, still doing work an AI agent could have done better, faster and cheaper twelve months ago.

The reasons are not technical.

Subscribe now

The gap is not where most people think it is

The standard narrative about corporate AI adoption is that companies are behind because the technology is moving too fast, or because they lack the right tooling, or because they have not figured out their data strategy. I do not think any of that is the real story.

The real story is that AI adoption is bottlenecked on people. Specifically, on how people think about technology, what they are economically free to say and do inside their organisations, and what reference points they have for what is actually possible. Many decision-makers in large companies are still forming their views of AI through locked-down, restrictive corporate deployments. Even when those tools are using current models, the experience is often several steps behind the frontier. Their frame of reference is limited.

This matters because the gap between what is possible right now and what most organisations are actually doing is enormous, and it widens every month. Meanwhile, mainstream media coverage still oscillates between bubble framing, labour panic, and tool-of-the-week commentary, which misses the operational change already underway.

The symptom of all this is what I call cosmetic adoption: AI implemented visibly, but without changing how work is actually allocated, reviewed, staffed, or governed. Real adoption changes the work. Dashboards only record activity. The 5C Adoption Friction Model I have been developing on diagnoses why organisations end up here, and which of the five friction points, Clarity, Capability, Credibility, Control, or Consequences, is actually blocking progress in a given team.

The uncomfortable part is that many organisations stall not because nobody knows what to do, but because the people with enough authority to change things often have strong personal reasons not to take visible risks. There are three structural reasons the gap persists, and they compound.

Share

Reason one: the technical middle ground almost everyone skips

One of the more persistent myths in the AI discourse is that you do not need to be technical to get value from these tools. The counter-myth, equally unhelpful, is that unless you can code professionally you are going to be left behind.

My take is more nuanced, and it matters for how you invest your time.

You do not need a software engineering degree to work effectively with AI tools. But you do need a specific set of working behaviours that most people in corporate roles have not built yet. You need to be able to inspect a failure when something goes wrong, read error messages and logs, understand what context the model has and does not have, break a task into testable steps, check outputs against reality, and recognise when a tool is producing plausible-looking output that is actually wrong. Terminal comfort is a useful proxy for most of these skills. A lot of them are easier to build if you are not afraid of a command prompt.

I know this because I lived it. When I was a student I ran a Linux laptop for a few years. That period built a comfort with the command line that I did not recognise as a strategic asset until much later. When I started using Claude Code, I was immediately comfortable experimenting with it. The learning curve was not steep for me because I had already paid the tax years before.

The people getting the most leverage out of AI tools today are not the non-technical people, whatever the “anyone can use it” marketing says. They are people who can form a hypothesis, run an experiment, read the output critically, debug when something goes wrong, and iterate quickly. That is a technical posture, even if it is not a software engineering posture.

If you do not have this posture, you hit a ceiling quickly. You will build a few impressive-looking demos, automate a couple of tasks, and then plateau. The compounding benefits go to people who can go deeper, and the gap between the two groups is widening every quarter.

The fastest way to build the middle-ground posture, realistically, is to experiment on your personal devices, in your personal time, on your own side projects. Corporate AI deployments are locked down for good reasons, but they give you a misleadingly limited frame of reference. You have to go wider on your own.

Reason two: the personal risk calculus behind corporate caution

This next point is the one I think is most under-discussed in the AI adoption conversation, and it needs to be raised carefully because it applies to a lot of folks.

A lot of what shows up as “we need to be careful about AI” or “let us wait and see” or “compliance will need to look at this” is not purely about risk management. It also reflects the personal risk calculus of the people making the calls. Senior corporate roles often come with asymmetric downside: the personal cost of backing the wrong transformation can be much higher than the personal reward for backing the right one early. Caution in that environment is not irrational. But some of what gets labelled corporate risk management is career-risk management wearing the same clothes.

Executives who are otherwise experienced and capable can find themselves more cautious about sponsoring uncertain work than their formal authority would suggest. When the personal cost of being wrong in public is high and the personal reward for being right early is modest, the rational move is often to wait for consensus.

This matters for AI adoption because substantive transformation requires sponsorship from people willing to advocate for work that might not succeed. That is a harder ask of anyone whose personal circumstances make public failure costly. I do not think this is a character failing. It is a structural feature of how senior corporate careers are constructed, and it is one of the quiet reasons progress tends to stall at visible, defensible activity rather than substantive change. The standard brakes on new work, cost discipline and compliance review, can then provide cover for inaction without anyone having to name what is actually going on.

The honest move is to separate two types of caution. Caution that reflects a considered view of the specific risks to your organisation is strategic, and it deserves respect. Caution that is more about protecting your own position is a different thing. Both are understandable. Only one is actually strategy, and the distinction is worth naming to yourself even if you never name it out loud.

Reason three: what actually blocks AI adoption

The 5C Adoption Friction Model I have been writing about in this newsletter identifies five friction points that block AI adoption inside real organisations. Each one is a question your people are silently asking:

• Clarity: What am I supposed to do differently tomorrow?

• Capability: Do I know how to do it?

• Credibility: Has leadership actually changed how they work, or are they just endorsing?

• Control: Do I have agency over how AI changes my role?

• Consequences: What happens if I simply do not engage?

Until those questions have real answers, the initiative does not change how work gets done, no matter how much is spent on tools and training.

The order matters. If Clarity is missing, training will not help. If Credibility is missing, communications will not help. If people feel no Control, adoption stays performative. If there are no Consequences, the old operating model wins by default. Fix them out of order and you compound the problem. Fix them in order and each one makes the next easier.

The pattern I keep seeing is that leaders diagnose a Capability gap, prescribe training, and grow frustrated when nothing shifts. Wrong problem. Technical leaders default to capability because capability is what their role trains them to see. Communications leaders default to clarity. The actual blocker is usually Credibility, Control, or Consequences, and those are harder to address because they require leaders to examine their own behaviour rather than commission another programme.

When those friction points sit unaddressed, what you get is cosmetic adoption. The dashboards show usage. Training completion ticks up. The leadership team reviews the numbers and approves the next phase. Meanwhile, the actual deliverables look the same as they did six months ago. The organisation has spent real money to produce theatrical return, and worse, is about to scale the next phase on a foundation of compliance rather than change.

A lot of the so-called “AI layoffs” reported in the press fit this picture. When you look closely at what the company has actually done on AI, the answer is often “rolled out Copilot, integrated one SaaS tool, reduced headcount, attributed the change to AI.” Calling that an AI transformation is a stretch. It reads closer to a cost exercise presented through an AI lens.

The pressure that produces cosmetic adoption is real. Leaders are getting analyst questions on every earnings call. Institutional investors come in and ask what the AI strategy is, how costs are going to be cut, how margins are going to improve. Shareholders expect movement. Stakeholders expect direction. The path of least resistance is to produce visible motion without restructuring anything important.

Cosmetic adoption provides exactly enough cover to delay the real work, which is addressing the 5C friction points in the right sequence. Operating model compression, the idea I have written about before, where an organisation becomes systematically faster, better and cheaper because AI agents coordinate work that human middle managers used to coordinate, does not happen by accident. It requires sustained, uncomfortable effort from sponsors willing to restructure their own organisations and, in many cases, their own roles.

Most leaders know they need to move. Very few have worked through the 5C friction points in the right order. And cosmetic activity buys time without closing the distance.

The counterintuitive bit: large firms can win

The consensus narrative in tech is that AI-first startups will eat incumbents in every sector, the way cloud-native startups ate on-premises incumbents in the 2010s. Parts of this will be true. AI-first entrants will take products and services that incumbents thought were locked down, and they will do it faster than most incumbents realise.

But there is a counterintuitive flip side, and I think it is underappreciated.

Some very large companies are already doing an enormous amount of serious work in this space, leveraging the economies of scale and resources they have in ways a startup cannot match. If you are a truly large company and you commit properly to AI transformation, the return on being first in your sector with the best capability, the cleanest data, the most automated workflows and the deepest tooling leverage is much bigger than the equivalent return to a startup. You have more processes to compress, more data to clean, more workflows to automate, and more people to redeploy. The arithmetic is simply different at scale.

The competitive picture over the next few years will look less uniform than the “startups always win” framing suggests. The incumbents that get serious will pull further ahead. The incumbents that stay cosmetic will get eaten. The spread between these two groups inside the same industry will be wider than anything I have seen in my career.

For a concrete example of where this gets interesting, think about Claude’s new design capabilities, which compete directly with tools like Figma and Canva. The strategic question for a large company is not “which tool do we buy.” It is a question about operating model. If AI tools make acceptable first-pass design work available across the organisation, what should the design function actually become? A production queue? A standards body? A brand governance layer? A centre of excellence for high-value creative judgement? The head of design role does not disappear, but what it governs shifts from producing every asset to setting standards, protecting brand coherence, and deciding where human craft is still worth the cost.

These are not small questions. They affect headcount plans, budget allocations, career paths and, ultimately, cost structures. The companies asking these questions seriously right now will have a very different cost base in three years from the ones still debating whether to licence the enterprise version of a particular SaaS tool.

What this means for you, this week

If you are a senior person in a corporate role, there are three moves you can make before Friday.

First, build the technical middle ground. Not a computer science degree, but genuine comfort with the terminal, debugging instincts, and a personal reference point for what these tools can do at the frontier. You cannot build this through corporate deployments alone. You have to experiment on your own devices, on your own time, on your own side projects. This is part of the unpaid permanent part-time job we all now have, alongside staying healthy, exercising and sleeping properly. It is not optional if you want to remain competitive.

Second, run your own cosmetic adoption test. Open the last three deliverables your team produced. Not the dashboards, not the training completion metrics, the actual work. Can you see where AI changed the output, or was it used as an input and then smoothed back into a work product essentially identical to what the team would have produced six months ago? If the work looks the same, the adoption is cosmetic. That is your first honest number.

Third, ask one diagnostic question in your next one-on-one with someone on your team. Not “how is AI going.” Instead: “what would you do with AI if nothing was stopping you.” Whatever they say tells you which of the five friction points is live. A task they do not know how to approach points to Capability. A fear of doing something they are not supposed to do is Clarity or Control. A shrug and “no one really expects much” is Consequences. A comment about what leadership is or is not doing with AI points to Credibility.

Those three moves cost almost nothing and will tell you more about where your organisation actually is than another quarterly adoption report.

I keep developing the full diagnostic toolkit here. Paid subscribers get the 5C Decision Tree, the Compliance Theatre Diagnostic, and the playbooks for each of the five friction points. If you want to move from diagnosing to changing, that is where to go next.

Subscribe now

Do you have a frame of reference for what is possible? Are you close enough to the frontier to make decisions that will still look sensible in twelve months? Most people cannot answer yes to these, and most organisations cannot either. That is the opportunity, and it is also the risk.

Not a bubble. A buildout.

The media’s default frame for AI is “bubble.” I do not think that fits what is actually happening.

The major AI labs are rate-limiting their users rather than cutting prices, which is what you do when demand is outrunning supply. Performance and latency are nowhere near what mature cloud services deliver. The hundreds of billions flowing into the underlying infrastructure each year do not go to an industry in terminal decline.

Bubbles do not look like this. Early-phase infrastructure buildouts do, where demand is real, supply is constrained, and the people paying closest attention are positioning themselves for what comes next.

The question is whether you are going to be one of them.

Best regards,

Brennan

Leave a comment

Start here if you are new to this series: Your AI Training Isn’t Working. Here’s Why.

Enterprise AI adoption stalls in five predictable places. The most expensive of them is not the one most leaders watch for. It is about meetings. Which ones AI reaches, and which ones it does not.

None of these are failures of competence. The default pathways in organisations push toward these exact choices. That is what makes them worth examining: they are structural failures, and structures have authors.

Most leaders are not making all five. But the one that is active is usually costing more than it appears, and the next one in the sequence is often close behind.

Subscribe now

Mistake One: Treating Adoption as a Technology Problem

Leaders invest in tools first and ask why the adoption is stalling later. They buy enterprise AI platforms. They build security architecture. They negotiate vendor agreements. And then they watch the adoption curve flatten after three months.

This happens because they have solved the wrong problem.

The friction in enterprise AI adoption is rarely a technology problem, at least not by the time it shows up. Once the platform has been bought, approved, and deployed, the binding constraint shifts from the model to belief, permission, trust, workflow, and incentives. The blockage has moved. Most leaders have not. People do not understand why they should change. Or they understand it and they do not trust it. Or they trust it and they do not have permission to use it. Or they have permission and they do not know how.

Better diagnostic questions start with the people. Who needs to believe this is valuable? Which conversation needs to happen before the technology matters? What would make a sceptical person in this room less sceptical? The work is there, in conversation rather than in procurement.

You end up with a fully deployed technology that no one actually uses, and the erosion of trust that comes from investing in tools no one wanted.

Mistake Two: Choosing the Wrong Champion

The second mistake follows quickly from the first. You need a champion to drive adoption. That is correct. But most leaders choose the wrong person.

They choose the enthusiast. The person who got excited about AI before anyone else. The person who already uses it in their inbox, who has seen the value, who wants to talk about it constantly. They think: this person will convince everyone else.

What actually happens is the opposite.

The enthusiast raises fear in the room, not interest. Because the enthusiast is not like the rest of the team. The enthusiast absorbs the learning curve more easily, bears less reputational risk in being early, and is not representative of the median employee. The rest of the team sees someone whose skills and confidence are already high and thinks: maybe I need to be someone else to use this.

The person you actually need is the trusted sceptic. The person everyone respects. The person who does their job well without AI and has asked hard questions about why this matters. When that person says, “I tried it. I was wrong about this,” other people listen. When the enthusiast says it, the room assumes the enthusiast is selling something.

The cost of choosing the enthusiast: you alienate the very people whose adoption would prove the technology works. You turn it into a special interest instead of a normal tool.

Share

Mistakes one and two stall adoption. Three, four, and five teach the organisation to hide that it is stalling.

Paid subscribers get mistakes three, four, and five, the compounding effect, the full Recovery Protocol map, and the attached protocol document with the three specific actions for each mistake. Subscribe.

Read more

Not everyone cares about technology. And that is fine.

I spend most of my day connected to a screen, actively trying to keep up with what is happening in AI, and I still miss things constantly. If that is true for me, you should not feel bad about it. You have a business to run, you have people to manage, you have a life to lead. Nobody should be expected to stay across all the technical detail, and honestly I think the expectation that you should is part of the problem.

What I have come to believe, after leaving corporate life and spending a frankly unreasonable amount of time and money going deep on this subject, is that AI adoption is a change management problem before it is a technology problem. If you want real transformation, you have to start with people. Not with models, not with tooling, not with benchmarks. People.

Once you accept that framing, the resistance you are seeing starts to make sense. The pushback, the scepticism, the blank stares in meetings, the quiet foot-dragging. Most of it comes from fear, and the fear is rational, because nobody in any position of authority is offering a clear and honest story about what this means for ordinary workers and households. How they benefit. How abundance actually reaches them. How their lives get better rather than more precarious. There is no believable transition story on the table. No positive vision being articulated by anyone. None.

That absence is the backdrop to everything I am about to describe. The race between labs, the geopolitical competition, the capital flooding in. All of it is happening at pace, and almost none of it is being translated into language that helps real people understand where they stand. So every piece of news below, I want to filter through a single question: what does this mean for how you lead your people through what is coming?

Leaders who engage now will outpace those who wait

I keep coming back to a simple observation: too many leaders are still treating AI as something to monitor rather than something to use. They read about it, they attend the conference sessions, they nod along in board meetings. But they are not using the tools themselves, and it shows.

Read more

Start here if you are new to this series: Your AI Training Isn’t Working. Here’s Why.

The Quiet Failure Mode Nobody Talks About

Last week’s essay explored why leaders avoid the hard question about their AI initiative: is it actually changing how work gets done? This week asks what happens when they keep avoiding it.

There is a kind of AI initiative that is harder to spot than the ones that fail visibly. It has not been cancelled. It has not been declared a success. It is still running. The conversations have not changed since month two. The pilot has not moved past pilot.

Nothing has visibly broken. Nothing has visibly worked either.

This is the quiet failure mode. Far more common than spectacular failure, and far more costly, because it consumes time and goodwill without showing that anything needs to change.

The pattern is recognisable. Six months of invisible stagnation, followed by a crisis that should have been a conversation three months earlier. The 5C Adoption Friction Model maps the forces that cause this. Stagnation is what happens when those forces sit ignored long enough that the initiative stops generating friction at all.

Share

The First Three Months

In month one, there was energy. You ran a launch meeting. You explained the why. You announced the first pilots. You took questions. You watched people’s faces. Some excited. Some sceptical. Most paying attention. The conversation was forward-looking.

By month three, the conversations were widening. New roles were asking questions. Integration challenges were surfacing. Real friction was appearing, the kind that comes from people actually trying to use the tools in their work. It was messy, but it was movement.

What Month Six Looks Like

If progress is real, month six looks untidy. New people have entered the conversation. The regular check-ins are asking different questions than they asked in month two. The budget has shifted, because the reality changed. People are finding edge cases nobody anticipated and using the system in ways the designers did not expect.

When stagnation has set in, the opposite is true. In an exploratory programme, too much stability is usually not a sign of control. It is a sign that nothing is moving.

Same people, same questions, same tools, same budget, same answers.

Nothing in the operating reality has forced adaptation. The budget, line for line, matches the original plan, which in exploratory work usually means no one learned anything worth adjusting for. The energy is gone. Not the dramatic kind of failure that forces a decision. The kind that fades quietly until nobody believes it will actually work.

The Three Data Points

These do not require a consultant. They require a calendar, a set of meeting notes, and ten minutes.

Data Point 1: How many new names are in the room? If the same small group is still doing the work in month six, adoption has stagnated. You cannot scale something that is not moving. Check your meeting attendance. Check your channel membership. Count the names that were not there in month one.

Data Point 2: What new friction has shown up? Friction is how you know people are actually using something. No new friction in month six that you did not have in month two means no one is using this enough to find new edge cases. Look at your support tickets, your feedback channels, your integration logs.

Data Point 3: What are your regular check-ins asking? A working initiative asks different questions in month six than in month two. If the conversation has not changed, someone is not learning. Pull last month’s agenda next to the one from month two. Compare.

Check your calendar. Check your notes. Write the numbers down. This is your diagnostic.

The Cost of Waiting

By month three, stagnation is recoverable. The team has not yet burned out. A pivot looks like learning.

By month six, the cost rises. People have invested time and attention. Some have made bets on this direction. A pivot now means admitting those early months did not lead where they should have. The people who believed in the original direction will need a reason to believe in the new one.

By month nine, the cost is very high. The initiative has history. Stakeholders have positions. Some people have been promoted because they “led the AI transformation.” The change is no longer a learning conversation. It is political.

By month twelve, you are in crisis mode. “If we had made this call in month three, it would have been a course correction. If we had made it in month six, it would have required a difficult conversation. In month twelve, it requires an apology.”

If you are between month six and month nine, the decision window is narrowing.

The Decision That Changes the Trajectory

If those three data points confirmed what you suspected, the next question is harder. Push harder? Change the approach? Or acknowledge that six months of effort has not produced the progress it was supposed to?

This is uncomfortable. The sunk cost is real. The stakeholder expectations are real. But the cost of delay is also real. Every week the decision waits, the initiative sinks deeper. Not into success. Into habit. Into expectation. Into the calendar.

You now have the diagnosis. The decision comes next: push forward, pivot, or reset.

Read more

About fourteen months ago, I quit a corporate job. I’d spent over 12 years in financial services working in technology and change, delivering projects across Australia and New Zealand. I had a young family. It was a calculated bet, not a leap of faith.

Share

This publication, Getting AI To Work, is about why AI initiatives stall inside real organisations, and what to do about it. That focus took me a while to find.

What pushed the decision was something specific. I had been experimenting with AI models in my own time, asking them questions about domains I knew well, areas where I had enough experience to judge the quality of the answer. The responses were so much better than anything a Google search could produce that it shocked me. I needed the space to dedicate real time to learning, experimenting, and building. That was not going to happen inside a corporate role.

So I left. And then I was sitting at a desk that was mine, in a house that was not quiet at all, writing into a newsletter that few at that point read.

The first few months were deliberate exploration. I wrote about vibe coding, about Gemini, about Studio Ghibli, about whatever caught my attention that week. If you go back through the archive, you can watch me think out loud. Some of those posts hold up. Some don’t. That is what exploration looks like when you are doing it in public.

Subscribe now

What I did not expect was how quickly my technical range changed. I’ve always been reasonably technical: some SQL, some Python, comfortable with the command line. But I was not a developer. And there was no pathway to experiment with this technology at work.

Since leaving, I have been building with AI coding tools almost every day. An AML risk assessment readiness tool for Australian businesses. An analytics platform that automates analysis of complex regulatory disclosures for clearinghouses. Several data and analytics tools to support my research.

None of these existed fourteen months ago, and I did not have the technical ability to build them. That direct, practical understanding of what AI tools can do informs everything I write about adoption. The rate of capability improvement still astonishes me.

By mid-2025, I had landed on a set of big frameworks. Operating Model Compression. The Liquid Organisation. The Triple Boundary Framework. I thought that if I could describe the structural forces clearly enough, the right people would find it useful. And many did.

But something was off.

What I got wrong

The frameworks described the world. They did not tell anyone what to do next.

I was writing for leaders who had already decided AI mattered. They did not need another explanation of why the world was changing. They needed a way to figure out why their specific initiative was stuck, and what to do about it on Monday morning.

The technology was moving so fast that the bottleneck had shifted. By mid-2025, the tools were good enough. The models were good enough. The constraint was no longer technical. It was organisational. Bureaucracy, governance theatre, change management failures, the politics of getting people to do something different when they are not sure it is safe to try. That was the real constraint. And my frameworks were not addressing it.

I did not have a single moment of revelation. It accumulated. The posts that got traction were the practical ones. The posts that got silence were the theoretical ones. The readers who reached out were not asking “What does the future look like?” They were asking “My team is not using the tools and I do not know why.”

That question became the 5C Adoption Friction Model, which became the foundation for everything I have written since December.

What this publication is now

If you are new here, particularly if you have come across from Global Custody Pro, here is where things stand.

The core of the publication right now is a diagnostic series on AI adoption friction. Each article names a specific friction point that blocks AI adoption and walks you through how to identify it, why it persists, and what to do about it. But I also write about whatever else I am thinking about, and that will not change.

The underlying framework is the 5C model: Clarity, Capability, Credibility, Control, Consequences. Five friction points. A specific sequence. Fix them out of order and you compound the problems. Fix them in order and each solution makes the next one easier.

For those coming from Global Custody Pro: this is different content. Global Custody Pro focused on AI transformation in custody and post-trade operations. This publication is industry-neutral.

The patterns I write about here show up in financial services, yes, but they also show up in healthcare, in professional services, in government, in any organisation where AI has been introduced but not adopted. The friction is human, not technical. The industry is incidental.

Every article stands on its own, and the foundational article that introduces the full framework is here. Start there if you want the complete picture, or start with whichever title sounds like your problem.

What paid subscribers receive

Each article includes a diagnostic tool for paid subscribers: conversation scripts, audit frameworks, action plans, checklists. They are practical tools designed to be used in a real conversation with a real team, not filed away. The library grows with each article, and paid subscribers have access to the full archive.

Subscribe now

Start here

AI Adoption Is Stalling. More Training Won’t Fix It. introduces the full 5C model. Every article since sits on top of it.

How to Unblock Your AI Roadmap in One Conversation gives you a single conversation to figure out what is blocking your team.

Operating Model Compression: A 2025 AI Year in Review is the bridge between the earlier framework writing and the diagnostic series.

Why Smart Leaders Are Still Hesitant About AI names the patterns behind executive hesitation that most commentary misses.

If you follow the Substack Notes or my other social media channels, you already know I think out loud there regularly. I am going to start doing more of that in the newsletter itself. Alongside the main articles, I will be publishing a shorter weekly notes post: what I am reading, what I am building, observations that do not fit neatly into a diagnostic framework but are worth sharing. The kind of writing that got me started in the first place.

If you have been reading along, thank you. If you are new, welcome.

I would like to hear where your own team is getting stuck right now: Clarity, Capability, Credibility, Control, or Consequences?

Hit reply to this email or send me a DM.

Normal service resumes on Wednesday.

Best regards,

Brennan