This week Anthropic filed its S-1, the SEC document required ahead of an IPO.

Watching the reaction across the internet, and especially in capital markets spaces, has been fun.

Anthropic have built a fantastic business, and too many people are lost in the headlines and not actually taking the time to read the document and think about what all of this means for your own business.

Take for example the “closing the loop” concept.

“closing the loop” refers to a development cycle in which successive versions of an AI model assist in building the next, with the goal of compounding capability gains over time.

For business owners, Anthropic is essentially implementing continuous improvement with AI models, as a mental model for how to use AI to run your business. You want AI to be optimising every constraint and removing every blocker at pace.

Anthropic runs on Slack. Claude is in Slack and has access to all that context. Agents autonomously do things. Code written by AI is evolving to AI research experiments driven by AI supervised by the PhD holding experts.

Subscribe now

This is operating model compression, like I’ve been writing about in this newsletter for ages, happening at warp speed.

If Anthropic has been able to drive this to revenue per employee of US$6.6 million per the S-1 filing, imagine what even a modest implementation of the same concept can do for your business.

Yet adoption is still too slow and not aggressive enough. People are scared about risks, worried about governance, and not understanding that you need the best AI models to defend yourself from cyber risks now.

Here is what Anthropic disclosed about how spend is evolving inside enterprise customers:

In addition, as of September 30, 2026, we had more than 1,900 customers with $1 million or more of revenue on a TTM basis and approximately 250 customers with $10 million or more of revenue on a TTM basis. Despite this growth, our penetration into the largest enterprises remains early. Over 50% of the Fortune 500 and over 30% of the Global 2000 were customers with $100,000 or more of revenue on a TTM basis as of September 30, 2026.

We believe we are in the early stages of our expansion within these customers, with our revenue from these Fortune 500 customers representing, on average, less than 0.1% of their total expenses for the three months ended September 30, 2026. We compare this measure against our largest Fortune 500 customers, who are further along in their adoption of AI.

For the three months ended September 30, 2026, revenue from our top 10 Fortune 500 customers represented, on average, less than 1% of their total expenses—still a small fraction of their overall spend, but over eight times the average of our Fortune 500 customers with $100,000 or more of revenue on a TTM basis.

In the S-1 filing, Anthropic disclose that spend intensity is still tiny even at their biggest clients. To my mind, it highlights how curiosity and experimentation are so important.

If you are behind the curve on AI adoption, it’s a people and change problem. The technology is already good enough to help AI frontier labs accelerate. Why are you still thinking that your business is special?

Leave a comment

Increasing AI spend intensity is one of the easiest plays a business owner can run today. Instead of worrying about cost control of modest plans or making horrifically short-sighted decisions like restricting use of the best models to “save money”, time block a free spending experimentation phase.

How do I run this playbook?

Read more

Before ChatGPT came on the scene and we started getting better AI models, an audit finding or consultant report was infrequent and expensive. I handled a number of these engagements in my corporate career.

Subscribe now

This isn’t an easy or fun process, and takes weeks or months to get an outcome that is commercially useful. Often you get your own ideas remixed and replayed back to you in exchange for a 6 or 7 figure invoice!

Quite often when you had these sorts of reports, you had to go off and do a lot of mapping work to turn a consultant finding into something workable inside that firm’s environment.

With an AI audit, every finding you don’t follow up on is going to keep coming back every time you run the next AI audit. If you do something like set up an AI agent to act as a monitor for a particular process, then until you fix it, it’s just going to keep alerting you.

This shows that a lot of people who use AI to find problems still aren’t thinking about using AI to help fix them.

This used to sound far-fetched to me, having worked in the corporate world for a long time, but this is what I think is the right mental model. We can’t plan and execute our AI transformations without understanding that AI can do far more than most people in the business world either appreciate yet or want to admit.

In the next section, I’ll talk about how to think about accountability for these findings and how the AI era shifts how we handle this process of exercising judgment around what we do and don’t choose to fix.

Read more

When you make the tough calls in an AI transformation, the process debt and technical debt inside your operating model are the next challenge to tackle.

When you have AI agents help you audit everything and find the process improvements and new system integrations, you now have a list of problems to solve.

Many leaders expect a quick return on investment, and while you might find a vendor you can immediately cancel or a setting you can change to rapidly cut cost on a service, there are no guarantees.

The early weeks of a serious AI transformation reveal low level things that are acting as blockers for realising benefits.

Share

For example, different systems might define key terms their own way, leading to different teams having a different idea of what “customer” means, or process steps that haven’t made sense for years.

To move through the process and start to actually get the return on spend, you now have a list of things to fix. From conversations I’ve had and work I’ve done over the years, I’m convinced that this is where many leaders are deciding that “AI doesn’t work for us” because that’s when they are shown the cost of getting the foundations right.

This doesn’t mean that there aren’t small use cases you can quickly ship and realise benefits from. It does mean that a lot more effort will be required because you are effectively embarking on a digital transformation of old with perhaps 10x the number of decisions required. You can’t delegate this one.

Leave a comment

The technical reason is because AI agents need access to your context and data. If you don’t share something and it doesn’t have access to your system and files, there will be gaps. A desire to control risk by reducing what AI agents can access is likely limiting the scope and impact of your AI initiatives before you’ve even started.

Every little shortcut will have the missing test coverage and business logic gaps revealed. There’ll be a series of questions asking why System A defines something as X when everywhere else defines it as Y. AI agents are now doing the work of requirements gathering, validation, and overlaying architecture without your detailed nudging.

This is why greenfield approaches - starting from scratch - are working well for many firms. They want the outcomes and business logic, but appreciate that because data quality and documentation quality aren’t high enough in their existing stack, they probably need to rebuild some things from scratch and progressively migrate things to their new operating model over time.

All of this adds up to why better managed firms are realising better returns from AI spend - first you need to have the right foundation and setup to actually be in a position to let AI agents help you.

The more operating model debt you have to pay off, the longer this initial dip period will take, but I think this is worth accepting, because the firms that stop their AI initiatives too soon are probably those at the highest strategic risk of disruption.

There are definitely quick wins out there. I’ve helped identify many. If a project doesn’t look like it will quickly deliver a return, many leaders will want it dropped. A lack of corporate patience and grit won’t help your firm survive the AI era.

Yes, budgets and constraints of stakeholder expectations are real, but you need to see the rebuild through if that’s what it takes. The firms that went into the last few years with strong foundations have been able to build on this.

There’s a clear reason why many regulated firms are seeing good results with their AI experiments - they already had quite prescriptive and structured ways of working, which is catnip for AI agents.

How do you get through this dip period though and not lose the stakeholders you need to keep happy?

The rest of this article is available for paid subscribers. If this sort of thinking about AI is useful, consider upgrading now.

Read more

A lot of stalled AI projects can be traced back to a hard decision that wasn’t taken.

When you use AI to help audit your operating model, a frontier model today can easily surface decision points that led to the current situation.

Our operating model is the sum of all the decisions taken and not taken since the business was incorporated, big or small.

When I’m working through a complex project with the help of AI agents, as things progress, more and more tough decisions are surfaced.

The direction setting of AI agents when they’re helping us transform a business and the front-loading of decisions where we can is a way to bring the pain forward and avoid rework after billions of tokens are generated.

Subscribe now

Decision debt is the accumulated pile of decisions, big and small, that weren’t made either intentionally or unintentionally.

It’s different to tech debt, which is the accumulated impact of less-than-optimal quick wins and operating at pace that can leave technology platforms hard to refactor or improve.

Some firms are really good on governance and code quality. Others are under so much pressure they just solve problems as they arise with less focus on precision.

Over time, all of this adds up to a stack of decisions - some one-way gate, some reversible - that require thinking and reflection time at leadership level and hard trade-offs to be both understood and made.

Share

When I look at stalled AI projects and even declining businesses, a preference for avoiding hard decisions, as opposed to being realistic about the options faced, is responsible for many businesses losing competitive position.

What this looks like in your business is issue registers not being closed out at a suitable pace and items where an accountable owner is hard to confirm. You really don’t want or need this in normal times, but in the AI era, kicking tough decisions down the road will be terminal for your business.

Read more

In today’s article:

• The first prompt advantage

• How things actually work

• Data quality matters

• Ownership and accountability

• Determined machine problem solvers

• Avoid failure modes

• Why it’s different this time

• Low-cost automated audits

• What the AI sunlight finds

A lot of people think that an AI transformation is just going to help them do things better, faster and cheaper, and that that’s that. My view is a little bit different.

I think AI is first and foremost a sunlight disinfectant that enables us to understand what’s really going on under the hood at a low technical level of detail in a much more efficient way than we might have had to do the investigation just one year ago.

This change is in what we see across the operating model. We’re still working in much the same way, but with the help of AI agents, we can rapidly find the bottlenecks and use agentic workflows to solve them much more quickly than we could in the past.

If you change how a business is using AI to an adversarial and audit-type style, the first thing you start to see is all of the little bits of extra fix and repair and maintenance work you need to do. You could almost think of it as an engine tuning or performance improvement exercise.

Every business has $100 bills sitting on the footpath that they haven’t picked up yet. I don’t think there’s much benefit in trying to put cosmetic KPIs around AI usage that don’t link up to what you actually want to achieve strategically.

When you start applying AI well, you start to get detailed, low-level insights into what is going on in your operating model that may not have previously been reported to you unless you had an incident and had to do a deep dive on a particular function or process.

Subscribe now

The first prompt advantage

You start with a blank prompt window in the tool of your choice. An AI agent won’t start automatically fixing your business. You still need to steer it in the right direction and set things up so that it can actually operate with the level of agency frontier models have today.

This is why I think that leaders getting on the tools and working extensively with AI is important. If they haven’t experienced both the success modes and failure modes themselves, the way they guide and govern AI transformation is going to be undercooked.

All of the standard things you need to think about when doing a transformation apply to an AI transformation. To make the right decisions and nudge the AI models in the right direction, having knowledge about how things actually work under the hood in a business or the curiosity and tacit knowledge to figure out where the answers are or who has them is more important than ever as a skill.

This comes naturally to a lot of business leaders when they are in direct control and directly accountable for a piece of work. They can operate. They have the context. They know the right questions to ask.

I think many people are missing that if you put in a goal that is modest, you will get a modest outcome. The AI models can do much more for us if we think expansively about what is possible.

I think one of the reasons why a lot of people are very sceptical of AI is that it’s very easy to get stuck in a doom loop of prompting. You can set up a very fancy set of workflows and dashboards and look like you’ve done an awful lot of work, but the business metrics that you care about don’t move.

Share

How things actually work

Processes, and how they’re designed, how long they’ve been operating, and the problems that people already know there are, are another key area here for exploration and interrogation.

The processes that everyone knows how they work often don’t actually map to exactly how they’re configured in systems and exactly how things flow from customer to firm to customer again, or customer to supplier to you.

You need to understand and appreciate that one of the downsides of being in a continuous-improvement, AI-first era is that every little process improvement win you achieve can probably be made better with the next generation of models.

In my corporate experience, I did a lot of business analysis work, sitting with people, watching how they work, asking them questions and tracing entire processes end to end and documenting them. When you do that process many times over many years, you understand the gap between how we think our business works and how it actually works at the front line. It can be two very different things.

It actually links back to the decision-making challenge around how we execute AI transformation, because what I see as one of the biggest value-adds of AI agents is that sunlight exercise: the requirements gathering, the configuration validation, mapping the plans and documentation (or lack of documentation) to what actually exists in our operating model today. This exercise is really important.

There’s this gap between tacit knowledge, which lives in people’s heads, and explicit knowledge, which is written down in documentation. A lot of that gap is the IP or competitive differentiation that enables firms to deliver value to their customers.

One thing which AI helps us do is close that gap at a much faster pace than having a team of business analysts come in and do interviews with everyone. They can still do that, and in fact, if I was tasked with setting up a requirements-gathering exercise today, I would be leveraging AI tools by doing things like recording all the transcripts and feeding them straight to my preferred AI agent to synthesise and tease out all of the things to make sure nothing’s been missed.

Why? Because context is everything in the AI era, and anything that isn’t accessible to an AI agent is something that you can’t use to optimise your operating model further.

Leave a comment

Data quality matters

Having worked in a data analytics team and having a lot of personal experience with data cleaning, data scrubbing, data processing and figuring out how to actually get golden-source data into a beautiful dashboard for stakeholders, I care about data quality.

I know that throughout the corporate world, and the business world in particular, data governance has been a buzzword as opposed to how people are actually operating their businesses.

This is one thing which I’ve written about a lot on Getting AI To Work, but data quality matters so much because if an AI agent is looking at the wrong piece of context, or has an inaccurate idea about what different data points mean or the value they have in optimising different processes or delivering specific bits of value to customers or complying with specific regulatory or legislative requirements, you can have really unfortunate consequences from not taking it seriously.

What are AI agents doing in this aspect of thinking about AI transformation? Well, AI agents can help you do those data quality deep dives and build out data remediation plans at pace.

You need to have a lot of testing, quality control and business expertise sitting across these processes, but to my mind this is yet another example of how, when you’re doing these transformations, a lot of the tasks which may have needed to have been done sequentially can now be done in parallel.

At the same time as doing all your deep diving on your operating model, you can kick off data quality interrogation agents to go and look at things and understand: how are things flowing? Where do you not have those business definitions? Where are the blockers in your data ecosystem that could lead to inadvertent AI agent mistakes?

Ownership and accountability

A lot of organisation charts have evolved over time and constantly get restructured or reorganised based on how leaders want to manage things and achieve other objectives through the stock-standard corporate approach of disestablishing roles, moving things offshore, or changing who reports to who in order to achieve the outcome they personally want inside the organisation’s goals.

Now, you can see from the attempts of the frontier AI labs to not really talk about the liability for actions of AI agents that, in corporate environments, making sure that someone is accountable for every single thing that an AI agent does in your business is quite an important thing.

There’s also a problem where that gap between what AI agents can do well, can’t do well, and need human help or help outside of the firm needs all of the right governance controls sitting around it, so that you know and can understand to the best that you can inside the volume of work that is now going to be done in these AI-first operating models, where things are happening at a high and detailed level alike.

Because when things go wrong, and they will go wrong, rapid remediation is now an expectation from external stakeholders. Sitting around for weeks and months doing a response to a regulatory incident isn’t going to cut it. People now expect the speed of AI disaster to match the speed of AI-enabled response and recovery.

So this ownership and accountability problem is a serious issue because, in my opinion, you need both the right business knowledge and technical knowledge. It’s no longer good enough to be delegating AI-related things to technical people.

One way to think about it might be: the most successful business leaders over the coming years are going to be the most AI-native ones, who have leaned into this technology to apply their domain knowledge and expertise to get the leverage and have the ownership over the overarching quality of every component of that operating model they are responsible for.

Determined machine problem solvers

Why do AI agents serve as auditors so easily? Well, if you give a human worker a task, they are going to route around the gaps in knowledge through doing the work. They’re going to go and ask people they know. They’re going to ask people they know who else they should talk to. They’re going to look at a system and know from their own experience the documentation and some basic research: what’s working, what’s not working.

The difference is with AI agents, they will not route around anything. They will either get stuck or complete what they’re asked to do. If they are aggressive and goal-maximising, they will find things which people are shocked and surprised about.

One thing I’ve found useful in my work so far in this AI journey is seeing the low-level things that get surfaced when AI agents operating at speed to achieve a goal run headfirst into the blockers. I think about times in my corporate career when low-level issues were sometimes only surfaced after enormous amounts of effort.

The fact that we can now just ask an agent to audit an entire database and prepare a detailed report is useful. We still need the experience and skills to judge and assess the output.

The flip side of having all of this is that the logs and audit trail are now evidence. So if someone in your business asks a question and finds something that requires regulatory disclosure or a chat with the auditors, more companies are going to find themselves having to do that.

From a wider societal perspective, this is probably better for increasing quality and compliance across all manner of industries because AI agents will start to surface things that even the best human experts haven’t considered over many years and regulators would love to hear about.

Share

Avoid failure modes

So what I’ve seen myself is that when I’m working through these sorts of tasks, the first run doesn’t really run into model limitations. It’s almost always context and access. So the key part of getting this investigative and exploratory side of an AI transformation right is around getting things set up properly and making sure you give things access.

I believe that a lot of pilot projects and proof-of-concept projects are completely set up to fail because they’re given such a narrow-scoped set of things to focus on that the wider value of AI isn’t being leveraged.

There are so many things which are really interesting to me, especially things like system misconfigurations, process steps that people have forgotten about or aren’t clearly documented, and handoffs that only work because there’s tacit knowledge about what needs to be done when. People haven’t stepped back and thought about end-to-end processes and how they should actually be working.

So there are a lot of issues in existing operating models that we might have dealt with previously through process improvement, continuous improvement, digital transformation, or however we might want to label these sorts of exercises. But all of these things are still present when we are thinking about doing an AI transformation. It doesn’t just stop.

So is this all just a rerun of “garbage in, garbage out”? I don’t think so. It is true that bad inputs have always produced bad outputs, but there are some things about AI and its capabilities that make it a little bit different this time.

Why it’s different this time

Read more

Today I’m going to talk about an incident where an influencer partnership Xero had in the UK led to a backlash from accountants. I think this isn’t actually a clumsy marketing mistake, but it’s exposing why software vendors have a lot of opportunity in the AI era, and intermediaries like accountants and other professionals face enormous risks.

This influencer claimed she could save £800 a month for management reports that previously an accountant performed. This was via a read-only Claude MCP connection to her Xero account. This use case is just the tip of the iceberg in terms of what sorts of things business owners who pay accountants lots of money will choose to do themselves.

I think that a lot of the initial reaction shows a lot of accountants haven’t thought through the strategic risks they actually face from AI disruption. Just because the current tool is read-only doesn’t mean that other AI tools that exist today can’t use browser use and computer use and access to APIs to do many more things that they thought they might be safe from disruption on.

Read more

This week’s newsletter is a little bit shorter than usual. You see, I’m getting increasingly disappointed at slow AI adoption.

There are clear reasons why this is the case. Business owners a…

Read more

Everyone wants to talk about how they’re using AI, yet much of the evidence we see from companies applying AI in its public form is AI-generated LinkedIn posts.

In today’s post, I want to share how I think about AI process improvement and what I think a lot of firms are missing when they select AI use cases without first taking a step back to reimagine how those use cases fit within an AI-first operating model.

An operating model is just a concept of how a business is put together. It’s about how people, processes and platforms roll up into capabilities that enable a company to ship products and services to its customers.

This AI-first operating model perspective focuses every single part of operating model change on reimagining how everything is done as if you were starting from a blank whiteboard with the power of not just the best AI tools as they exist today, but what they will be like in six months’ time and beyond.

An example of the old way of approaching operating model change is how companies traditionally approached outsourcing. Let’s say they have a call centre with 500 employees. They know they have to pay for all of the systems and processes and controls and audits around operating this call centre.

The traditional way of doing this is to get a consulting firm in and get them to map out how the call centre currently works. Then get them to do a market scan to see if there are outsourced call centre providers who can replicate that capability, meet all of the different quality and cost targets that we want to achieve from this operating model change, and then give us back a nice pretty report and execution plan.

For us to go to the board and say, we want to outsource our call centre, here’s the plan, and “brand name consulting firm” helped us put this together. Can we get this spend of $10 million approved because we will save $15 million a year ongoing if we outsource this?

A more cynical take might be that the old way of thinking about operating model change involves a lot more predetermination and constraint in terms of reference and assignments of work to arrive at the preordained outcome of, oh wow, we have to outsource or offshore this thing that we’ve wanted to do for ages, and here is all the evidence to back that up, and here is the concrete plan for us to execute it, and look here, the consulting firm that helped us do this just happens to have a preferred outsourced call centre “strategic partner”!

The AI-first operating model change is a little bit different. When I talk about starting at a blank whiteboard, I’m serious. You want to be imagining how you would rebuild the capability you’re trying to rebuild. In this example, a call centre, and whether that is even the right boundary of capability you want to be rebuilding.

What are the jobs to be done? What is the right blend of human and AI agent? What are our customer expectations around the sorts of things they can just do self-service and the sorts of things where they do need to speak to either a human customer services person or an AI agent? What are all of the regulatory challenges we face around voice interaction with customers?

Read more

Over the last couple of days, I’ve been researching something you probably think about when you hear corporate leaders talk about what they’re doing with AI, but may not have examined in detail: what’s going on under the hood.

When a CEO claims on an earnings call that 10,000 people are being laid off because of investments in AI, that’s just the surface-level proclamation. When you look under the hood, what is often happening is that the 10,000 number is made up of people who were going to leave anyway because of retirement or churn, the movement of headcount from a big corporate to an outsourced vendor, or even an FTE-equivalent number made up of hours saved from processes that started getting automated years before the ChatGPT moment.

Every case is going to turn on its own facts, but over the last few years, before AI model capability could even justify some of the claims being made, it was clear that trying to label every cost-cutting measure as AI-linked would be rewarded by capital markets.

My take that I’m going to unpack in this article is that there are three things going on here:

1. AI capability is very real and already delivering results

2. The evidence is weaker for AI already causing large-scale job losses

3. There is an AI capability-to-deployment gap, not fully explained by culture

That is to say that deploying AI yourself and deploying AI in an enterprise environment are two very different things. There are all manner of considerations that a business needs to worry about: provisioning and permissions, how AI tools integrate and connect to other systems, and what sort of evaluation and testing processes are in place.

When you read an article that claims AI does work or doesn’t work, this is so simplistic that it’s essentially worthless if you’re a business owner or leader thinking about how you actually get AI working in your business.

Every operating model has so many moving parts. That is why, when I talk about an AI-first operating model, I talk about the need to rethink everything in it. This is all happening in an environment where frontier AI capability has moved much faster than the ability of a business to take that upgraded capability and turn it into reliable workflows and use cases that are actually delivering value.

Only the companies that care the most and have put in the most effort are getting these rewards. A lot of companies were already behind on the technology journey. They are suddenly finding themselves blocked or restricted in the number and quality of use cases they can put into production.

AI essentially acts as sunlight and brings all of the poor decisions and underinvestment of the past to the surface. It is this environment in which things like AI washing or attributing cost savings to an AI project when it was actually a plain old automation project can start to be the response to incentives you see from some corporate leaders.

What are the variants of AI washing?

There are several different subtypes of AI washing to consider. These are substitution washing, reallocation washing, bundle washing, denominator washing, forecast washing and vendor washing.

Read more

You’re thinking about AI costs the wrong way.

A lot of commentary is written about AI costs and how much firms are spending on tokens.

I’ve even indulged in some of this myself.

In today’s post I’m going to share why large spending on AI can make complete commercial sense.

Can you spend too much on AI?

Cost per task matters when you are building an AI-first operating model.

If a solo operator spends $100,000 a year on AI, is that too low, about right, or too much?

If a 1,000-person firm spends $100 million a year on AI, is that too low, about right, or too much?

Whether AI spend makes sense depends on its wider context.

• How big is the business?

• Who is running this initiative?

• Are the right guardrails and governance in place?

• What are they really using AI for?

• How big is the opportunity that they’re targeting?

• How long will this spend level last - is it build or run?

• Is this removing things from the operating model permanently?

• Is this a major cost avoidance or reduction in regulatory risk?

The right unit of analysis isn’t a token

I think the “token maxing” trend was downstream of a wider anti-AI pushback trend in the workplace where many workers, even when given access to the best technology of the era, don’t want to put any effort into using it and learning something new.

Instead of thinking about tokens consumed, think about the business benefits. Time savings, error rates and rework are the easiest to plan, measure and realise.

For each use case, articulate the benefits and how they will be realised. An AI project is just another project.

One common failure mode I’ve seen is finance-minded folks thinking they can control cost by restricting models. This can be catastrophic!

For a detailed adversarial review of a plan, Fable 5 can justify a higher per-task cost than Gemini 3.6 Flash.

If you’re not operating at that level of detail, you shouldn’t be imposing cost controls and usage guardrails at that level.

This is where many firms go wrong: severe guardrails and cost controls imposed up front tie teams’ hands before they can experiment towards better cost efficiency.

Big returns are plausible for some firms

You don’t need major percentage improvements if an initiative is wide enough.

Not backfilling roles through natural attrition, or redeploying people between functions, represents a real benefit to a business.

Deferring hiring for six months in roles previously filled at scale can generate savings beyond generous AI budgets.

Removing a vendor category, SaaS subscription or enterprise product feature at renewal can produce the same result.

At larger firms, a 3% reduction on a $100 million spend can justify significant investment, depending on hurdle rate and expected payback.

Enormous numbers can be plausible when you are improving a $10 billion payroll or $3 billion vendor spend.

There is a jagged frontier of adoption

The jagged frontier means some AI use cases are proven enough that piloting them is a waste of time: agentic engineering, long document review, or infographics.

Everyone competent knows that you can use these use cases and execute them at a decent level of quality. This is limited only by your own skill level and the models and harnesses you’re using.

For more complex use cases, a human in the loop with the skill and taste to see what needs rework is still required.

In many industries, human-in-the-loop quality review is non-negotiable, especially when work interfaces with regulatory requirements. In legal practice, rules vary by jurisdiction, but lawyers remain responsible for verifying AI-assisted submissions and following any court-specific directions on disclosure or use.

AI transformation is trial and error, like every new technology. Clear scope matters, but AI is surfacing blockers elsewhere in the business faster.

If systems cannot talk via API, or data rules are unresolved, agentic coding tools quickly expose the legacy technology and poor decisions underneath.

The call many leaders make is that this is now “too hard” and “AI didn’t work for us” - when all it did was force a look in the mirror.

Read more

When I quit my corporate job, I didn’t have much idea what I wanted to do.

I knew that AI was the most consequential technology shift since the Industrial Revolution, and that it could become more important than the internet or smartphones.

I needed to find a way to share that thinking with like-minded people who understood that it was now a people problem, not a technology problem.

The technology can do the task. The people are still pushing back.

• They hate AI

• They don’t want it in the workplace

• They don’t want AI-generated content

• They don’t want AI-generated work product

The exact people who would get the most leverage out of AI are people with experience, people with taste, and people with domain expertise.

So what is going on here?

Read more

You need to rethink your entire operating model, and because no one is actually doing this properly yet in the AI era, there’s a window where first movers can build big advantages over their…

Read more

In today’s newsletter:

• Benchmarks are getting saturated

• AI-era operating model redesign

• Shipping use cases to production

• Implementing this playbook (mega-prompt for paid subscribers)

Share

Benchmarks are getting saturated

Every few months, we get much better AI models. A benchmark is a set of tasks of varying complexity that gives us a basis for comparing Claude Fable 5 with Kimi K3.

Benchmarks are getting saturated - in June, Artificial Analysis removed IFBench from its Intelligence Index because it no longer distinguished frontier models well enough.

Artificial Analysis recently launched Doomscroll, a randomly shuffled feed of charts showing how AI models compare across many measures.

Many well-specified tasks can already be done faster and more cheaply by AI agents.

You don’t need to spend your money running a pilot project to prove that an AI model can extract text from an invoice and turn it into something useful.

Read more

I’m writing this article because increasingly, in conversations I have and company filings I review, I’m building a picture where I feel like too many folks are walking past the $100 bills l…

Read more

Implementation can be delegated, but we still need leaders to care about the judgement they exercise. The point of all this is achieving real business outcomes.

One thing common to technology projects the world over is the tendency for leaders to ask technical people questions that should already have an answer from the business itself.

Subscribe now

This sometimes manifests in technology projects being designed, approved and implemented as ends in themselves. There is no clear link to the overarching business strategy or what the future operating model needs to be. Sometimes there isn’t even a clear people strategy. No thought is given to change management.

There is a difference between a trusted advisor making a recommendation to you, making a decision for you, and executing the decision for you. When we engage external help, it is because we need to go from a state that we are not happy with to one where we have achieved our goal.

Share

There are three lanes you might want to think about - the leadership lane, the specialist lane, and the team lane.

PS: I’m also going live on Substack with at 1800 EDT 23 Jul / 0800 24 Jul AEST to talk about human in the loop in the AI era.

If you have your own AI change problem you’re stuck with, I have limited availability over the next 90 days to work with a small number of clients to help them unblock their AI initiative. DM me on Substack or reply to this email if that’s something you’re interested in.

In the next section, paid subscribers can access a decision map to help stop decisions bouncing between governance forums and overburdening business owners and leaders with decisions relating to AI initiatives.

Read more

This project can still have no real business owner. In fact, there might even be an executive name in the “Business Owner” box - yet this person was “voluntold” and will be polite about the project but won’t put their best effort into making it a success.

The gap between a sponsor going through the governance motions and a business owner who cares deeply about achieving the end goal because their reputation is on the line couldn’t be starker.

Transformation projects already carried high stakes before AI began dominating boardroom agendas around the world, so it’s no surprise that CEO-level ownership matters so much.

When I worked in change roles in the corporate world, I often came across well-documented projects with the finest hand-curated governance arrangements. Yet, when I looked under the hood and did the testing myself, I often found a documentation-reality gap.

Well-funded change programmes with “best practice” governance can produce amazing governance documentation while accountability and ownership dilute with every decision and meeting.

There is nothing worse for culture and buy-in to a change agenda than a delivery team waiting for business decisions while no one senior understands why AI projects are surfacing complex issues that block progress. Many AI projects are also surfacing hard-to-reverse, one-way-door decisions that leaders cannot afford to defer.

Silo structures and corporate politics are always fun to navigate, but on AI projects, they can be even more disruptive because of the additional hysteria floating around in relation to job loss and other topics like outsourcing and offshoring.

This requires careful planning and a lot of intensive stakeholder management through regular check-ins and over-communication. All of the different “project roles” can’t afford to be passive when so much is riding on AI projects at many firms.

What are the key warning signs?

• Delayed decisions that keep being rolled over from meeting to meeting

• Steering committee meetings that deviate from required decisions or actions

• No clear business owner with the authority to make decisions

• The gap between formal communication and informal vibes widening quickly

• Constant escalation of problems due to lack of clarity and delegated authority

A theme I’ve written about a lot on Getting AI To Work is that AI projects surface the cans that have been kicked down the road. If you took a shortcut on a data project three years ago and those gaps now undermine your AI project outputs, well, you need to fix that blocker to achieve the desired standard.

A technical team can’t decide a firm’s risk appetite, figure out the entire operating model in isolation, or manage the people consequences of AI change without the right level of executive support.

In the next section, paid subscribers will get further analysis and a short ownership test they can apply to a live AI initiative in their business. If you subscribe now, you’ll get the full ownership test and access to the archive of practical AI transformation and change playbooks.

Read more

As the second half of 2026 proceeds, the distribution of outcomes across firms with active AI initiatives is wider than optimists might have predicted.

The governance is in place, the budgets have been approved, and the licences rolled out - yet the operating model is the same as it was last year.

I’ve spent many an hour over the course of my corporate career in a workshop or meeting that danced around the real choice a business faces.

Sometimes you can build a stakeholder coalition and get the right outcome. Other times, politics and commercial realities push the tougher calls down the agenda.

Share

One example is a use case that crosses departmental boundaries. Stakeholders need to agree that the way finance and sales run the process must change to realise the benefits. This is a 30-second decision from a leader, but a 10-week drama at many firms.

Deciding who is accountable for AI output, and what risk profile is acceptable for work that reaches customers or regulators, is a big part of the complex decision-making that can accelerate benefits realisation or cause a project to stall.

The core idea is this: decision debt is the accumulation of tough calls you haven’t made. It increases cost, confusion and resistance in much the same way as tech debt - or keeping an underperforming supplier in place.

The consequence of this is that when leaders do not make tough decisions, people reporting into them and the people reporting into those people start to become more cautious and risk-averse.

People worry more about their job security and the political consequences of making a judgement call. They start to wait for permission and preserve the old processes. That is the opposite of what an AI transformation, or any change initiative, needs.

Paid subscribers will be able to read on to see the five questions I would use to expose the decision debt inside an AI change initiative and how I think about this challenge for AI projects of all sizes.

PS: If you haven’t already subscribed to my YouTube channel, do that now.

Read more

• The rise of AI cost consciousness

• What many leaders are missing

• How to implement this playbook

PS: Check out my latest video over on YouTube:

The rise of AI cost consciousness

The money you spend on AI subscriptions and API bills is visible. The money you spend on time wasted, duplication, and inefficiency often hides in your management reports.

The rise of AI cost consciousness is what I’m calling the intense focus in recent months on how much AI is costing some businesses. If you’re using APIs or coding agents at enterprise scale, every token can add to the bill.

And the pricing for the best models isn’t cheap. When you read about senior Uber executives complaining about blowing their Claude Code budget for 2026 in just a few months, you can see how that happens.

This is because of both the capability and the marketing. GPT-5.6 Sol can just do more for you at a higher level of quality. If you’re on a subscription plan, using it up to the limits or even stacking multiple $200/month plans is something I increasingly see among power users.

When you have worked in the corporate world or run a business, you know that it is quite normal for departments to have a budget and be managed to that budget. For some commentators on social media, that sounds outrageous, and increasingly I think that they’re not wrong.

Read more