Your AI strategy is chasing $1 coins

Most teams use AI for small tasks, while the biggest operating-model gains stay untouched.

Jul 27, 2026

Hi there,

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 lying on the footpath and focusing intently on the $1 coins instead.

AI change is an operating model rethink

I’ve written a lot about how AI is an opportunity to rethink your entire operating model. It’s not “just” a technology project. Yet, we are a few years into this technological shift and adoption is still too slow. Naysayers point to the economic data and claim nothing is happening.

If you actually use the tools, and don’t just repeat what is printed in newspapers, you know that a lot of the common “objections” to the use of AI aren’t really applicable any more.

Models have become more reliable, with fewer hallucinations and stronger logical and mathematical performance. If you are still seeing frequent errors, first check whether you are using the right model, context, tools and quality control process for the task.

This doesn’t mean that you can operate without safety guardrails or checking output quality. It means that many ordinary business tasks can already be performed better, faster, and cheaper by AI agents.

You just need to start delegating more tasks to AI agents instead of teams. And the sooner you get to that point, the better for remaining competitive.

What does that even mean?

One of the biggest change management challenges is getting your team open to the idea of change at all. Just because we want to roll out a tool or make a change, doesn’t mean it will be automatically adopted.

I’ve worked on many projects where the conversations around the people impact were more consequential than the technology. And what I see in AI transformation so far is a lot of focus on the technology, and not nearly enough on the people.

If you’re a technology optimist and an early adopter, it’s easy to project those values onto people who view technology as a work thing and have no real curiosity about learning a new application or way of working.

There is a whole change management literature, and when I reflect on my own change management experience in the corporate world, almost all of it lines up quite nicely.

When we think about a business we have different teams: finance, sales, operations, marketing. They all have capabilities to get tasks and bundles of tasks done, and coordinate inside and outside the firm to deliver value.

The organisation of the firm in the AI era shifts to pushing more tasks and bundles of tasks to AI agents instead of either internal or external human collaborators. Coordination via API endpoints instead of Slack messages and long-running email threads is the approach.

Every task has a cost-appropriate model

The $1 coins I referenced at the start of the article are the simple tasks most people have tried using AI for, or already use it for.

Summarising an email thread
Reviewing a long PDF with technical information
Building a daily automation to perform a simple task
Conducting basic competitive intelligence gathering

Even the cheapest models absolutely dominate many of these tasks now. If you are paying for API usage, it’s a waste of money to have Claude Opus 5, Kimi K3 or GPT-5.6 Sol perform a summary of an email thread.

You might get a really good summary, but it’s a pointless exercise. Gemini 3.6 Flash or GPT-5.6 Luna or MiniMax M3 can do just fine.

To get the real benefits from AI transformation though, more complex bundles of tasks need to be delegated to AI agents over time.

A lot of AI change initiatives fail because they roll out a licence for ChatGPT and do not provide guidance on how to design, test, and safely deploy much more complex workflows.

Many workers will not just “figure it out”. They need training and change management support to feel comfortable that mistakes they make at this stage of the early adopter era will be treated as part of their skill development and not used as fuel for building a regressive blame culture.

What Codex usage data found

This chart comes from a June 2026 paper, The Shift to Agentic AI: Evidence from Codex. It shows that between 1 November 2025 and 11 June 2026, median output-token generation among active OpenAI workers rose at least tenfold in every job function. The measure includes output from both ChatGPT and Codex.

This chart shows a stark difference in skill use. In the seven days ending 11 June 2026, 25.7% of active individual Codex users and 30.4% of active organisational users invoked at least one skill. Inside OpenAI, that figure was 96.2%. The paper treats this as evidence that intensive users are turning one-off interactions into reusable, systematised workflows.

The $100 bills to pick up are the most complex tasks that cross team boundaries, involve multiple systems or data sources, and incorporate elements of judgment and decision-making.

While there are commonalities in every business, part of leaning into AI change is experimentation and figuring out what works and doesn’t work in your business. If you review skill use across your team, how does it compare with the 30.4% organisational-user rate in the paper? And what would have to change to move closer to the frontier demonstrated inside OpenAI?

OpenAI is not a typical workplace benchmark. The paper describes it as an unusually favourable environment, with high familiarity, strong organisational support, internal training and extensive knowledge-sharing. But it provides a useful picture of what becomes possible when those adoption barriers are lowered to the greatest extent.

This is just one example where we are reminded that to realise the benefits from AI, you need to push the tools to their capability frontiers and keep pushing. This requires a culture that is open enough to make mistakes and learn from them.

A theme from a number of conversations I’ve had over the last 18 months is that because a lot of leaders are not technical, they have underestimated the level of technical hands-on training and coaching that is required to get the most out of AI tools.

The irony is that the AI tools themselves are the best teachers - yet hardly anyone is using them as teachers!

The prompt to run this playbook

Here is a prompt for you to try with your preferred tool. It will work best with something like ChatGPT/Codex on desktop or Claude Desktop.

Review the interactions and work history from the last 90 days that you can access. Begin by telling me what is and isn’t visible to you. Identify repeated workflows that should become skills but haven’t. Where an available plugin would help, recommend it and help me install and connect it with my approval. Identify gaps in how I use AI so I can upskill rapidly. Where I am doing manual work that you could help schedule or automate, identify it and interview me so we can fix it. Help me understand where I am both underusing and overusing your capabilities.

The next step is to understand what this exercise reveals. You most likely have a lot of duplication and token waste across your business if you don’t have common skills available for team members to copy what works from colleagues.

If your security policies or governance have blocked a lot of this functionality, again, you are leaving $100 bills on the footpath through waste and unnecessary manual effort.

None of this absolves you of accountability for output quality or responsibility for testing the agentic workflows and scheduled tasks you set up. But hopefully it provides a reset opportunity to take your AI use to the next level.

Drop a comment below if this prompt helps you get your work done better, faster, and cheaper - or if you already use something like it.

Getting AI To Work by Brennan McDonald

Discussion about this post

Ready for more?