Clarity,
applied.

Most AI pilots do not fail. They succeed. They produce a working demo, an enthusiastic email from a sponsor, and a slide that finds its way into the next board pack. And then nothing changes.

We call this pilot theatre. It is the most common pattern we see in firms that have spent meaningful money on AI in the last two years. Twelve pilots, three of them technically impressive, none of them in production, none of them anyone can name a P&L line for.

There are three reasons it happens, and two ways out.

One: the pilot was scoped to be a pilot. It was sized to fit a quarter, not a workflow. By the time the demo lands, no one has the budget to take it the rest of the way, and the team that built it has already moved on.

Two: the workflow it lives inside has no owner. AI sits between functions. If no single person is accountable for the workflow end-to-end, no one takes the pilot through the messy work of becoming a system.

Three: the firm has not changed any of the rituals around the workflow. The morning meeting still runs the same way. The weekly review still asks the same questions. The pilot is real. The operating model is unchanged.

The way out is unglamorous. Pick two workflows. Make one person accountable for each. Rebuild the meeting that runs them. Then — and only then — turn the pilot on.

We have written this for clients enough times that we are now writing it once, here.

An operating model is, in practice, a calendar. It is the set of meetings in which the business decides things, and the artefacts those meetings produce.

There are five meetings every operating model needs. Most businesses we work with have three of them, and the other two are happening informally — in corridors, on Slack, after the official meeting has ended.

The weekly run. What happened, what's about to. Forty minutes. No slides.

The monthly review. What's working, what's not, what we're changing. One page in, one page out.

The quarterly bet. Where the next 90 days of effort go. Decided, not debated.

The annual frame. The narrative, the priorities, the price the business is paying for them.

The board. The smallest meeting, often the loudest. It works when the other four do.

If two of these are missing or rolled into another, the business is being run somewhere else. It is worth finding where.

We use language models every day. We also write a lot of things by hand. The line between the two is more useful than people make it.

Let a model write: meeting notes, summaries of long documents, first drafts of recurring artefacts (status updates, briefs that follow a template), translation, formatting.

Write yourself: the strategy memo, the resignation letter, the email to a client whose project has just gone wrong, the document that names a thing for the first time. The work where the writing is the thinking.

The principle is simple. If the value is in the words being correct, a model is fine. If the value is in the act of writing — in being forced to think the thought all the way through — a model robs you of it.

The harder cases are the ones in the middle. The board paper. The proposal. The all-hands. We tend to draft these by hand and use a model to pressure-test them, not the other way round.

We get asked, regularly, when we plan to grow. The honest answer is: slowly, and not by much.

There is a size at which a consultancy stops doing the work and starts selling the work. The leverage model demands it. Once the partner-to-associate ratio crosses a certain threshold, the partner is no longer in the room — they are managing the people who are.

We have decided that the work is the product. So we are staying the size at which the partner is in the room, every meeting, every memo. That is a ceiling on how many clients we can take. It is also the thing clients are paying for.

If you want a hundred-person firm, there are good ones. We are not trying to be one of them.