Perspectives.

Potential is not a strategy.

Walk into most mid-sized companies today and you will find AI potential stored everywhere: a chatbot on the website that nobody uses, a licence for a copilot that sixty managers were given and twelve opened, a data-science contractor nine months into a model that has never touched production, and a board asking for "an AI plan" by year-end. None of this is a technology problem. All of it is an ownership problem.

Stored potential converts into operating advantage when one person is accountable for it — not for the pilots, but for the P&L line the pilots were supposed to move. That person does three unglamorous things in the first month. They rank: which uses of AI change revenue, cost or speed in this business, in order, with a number against each. They stop: the vendor that is not used, the model that will not ship, the pilot that was never a priority. And they decide where the humans sit: which work moves to software, which checkpoints a person must sign, and what record is kept of every decision made along the way.

The test of whether a company has an AI strategy, rather than AI potential, is embarrassingly simple. Is there a line in next year's budget that depends on it? Is there a decision log a director could read? Is there a named person who loses sleep when it slips? If the answer to any of these is no, the company has a collection of experiments, and experiments do not compound.

The owner also has to have built the thing. Not read about it — built it, with the scars that come from data that does not reconcile, models that are confident and wrong, and users who will not change how they work for a tool that does not respect how they work. That experience is what separates a plan that survives contact with the first quarter from a plan that survives only the board meeting.

Potential is a balance-sheet item. Strategy is an income-statement item. The job is moving it from one to the other.

Choose your model on your own tasks, not on benchmarksFORTHCOMING

A vendor-neutral account of running the same six tasks through the leading AI labs — research and numbers, a long-document read, a coding task, an unattended agent, a client-ready memo, and a live-information question — scored on correctness, speed, cost, how much babysitting each needed, and whether a client would trust the output. Which lab for which job, dated, because the answer changes monthly.

Before AI touches a capital-account statementFORTHCOMING

What a family office or wealth manager should insist on before any AI reads, reconciles or reports on client holdings: provenance on every figure, exceptions routed to people rather than hidden, no training on client data, a human release gate, and controls sized to the sensitivity of the records.