For about three years now, the same conversation has been happening in every boardroom I sit in. Which model. Which provider. Which framework. Should we be on the latest. Should we wait. Are we behind. Are we ahead. The conversation has shape. The shape is technical procurement.
The conversation has been the wrong conversation.
Not because the technical choices do not matter — they do, marginally — but because the technical choices are not the moat. The technical choices are commodities. Anyone can buy access to the same model your company is buying. Anyone can switch frameworks in an afternoon. The thing the conversation is consistently failing to land on is the thing that is actually defensible.
Ideas are the moat. Specifically: a sharp answer to "what should this thing do for us, exactly, given who we are and what we are good at?"
Most teams cannot answer that question. They have meandering answers. They have committee-approved answers. They have aspirational answers that survived contact with three risk-assessment workshops and emerged as marshmallow. They do not have sharp answers. Without a sharp answer, the model is pointed at fog.
What a sharp answer looks like
A bank I worked with had a problem with their commercial-lending arrears process. The previous strategy was "use AI to improve risk modelling". Vague. Aspirational. A consulting-deck idea. Eight months in, nothing had shipped.
We replaced it with: "We can detect three early-warning patterns that human relationship managers miss in the first sixty days of a deteriorating loan, because we can read every email and meeting note the manager logged and apply pattern-matching that no individual manager has time for." That sentence took an afternoon to write. The pilot built around it shipped in six weeks.
Same model. Same data. Same engineers. The only thing that changed was the sharpness of the answer to "what should this do for us?"
I have been in the room when a thousand variations of this question got answered badly. Here is what badly looks like:
- "Use AI to do summarisation." Of what, for whom, replacing which current behaviour?
- "Build a chatbot for customer service." Which customer cohort, which support tier, what metric?
- "Apply AI to the operations team." Apply what to which process, displacing which manual step?
Each of those is a vendor pitch dressed as a strategy. Each is a request for the model to figure out the strategy on its team's behalf. That is not what models do. Models execute strategies. They do not author them.
The labour the model cannot do
The labour that should be happening in the boardroom is the labour of authoring a strategy with enough resolution that an engineer could sit down on Monday and know what to build. If your AI conversation is still abstract by Friday, the model is not your problem.
The constraint on strategic work has never been the speed of thinking. The constraint has been the courage to commit to a specific answer before you have all the data.
This is also why I am not particularly worried about AI replacing strategic work in the next decade. That is a human-discomfort problem. The model can help you think through options. The model cannot commit on your behalf. Or rather: the model can commit, but the commitment is unowned, which is to say not a commitment at all.
What this means for moat thinking
People keep asking me whether the rise of capable models will erode the advantages of incumbent companies. My read is: it will erode the operational advantages and amplify the ideas advantages. Companies that have been winning on superior process and execution will find their advantages compress. Companies that have been winning on superior judgement about where to point execution will find their advantages widen.
If you are an operator wondering which side you are on, the test is dull and revealing. Watch where your leadership team spends its most expensive hours. If those hours go to debating which tool to buy or which vendor to retain, you are an execution shop. If those hours go to committing to specific bets that the rest of the org argues against, you are an ideas shop.
The interesting tools are getting cheaper. The interesting ideas are not. Spend accordingly.
I publish one essay a month, roughly, on AI and operations. If something here struck a chord — or struck the wrong chord — I would like to hear about it. Write to mark@skycot.com. Long emails are welcome; I read all of them and reply to most.
If your team is wrestling with the patterns above and you would like to talk, the best way is the contact page. Advisory engagements are limited but I keep one or two openings.