Don’t marry a model.

Every few months, the leaderboard changes hands. A sound AI strategy shouldn’t care. The gate is permanent; the model behind it is replaceable.

Somewhere right now, a leadership team is being asked to “pick an AI.” The board wants a name it recognises; the vendor wants a multi-year commitment; the team wants to stop having the meeting. So an organisation weds its processes, its prompts and its budget to one provider, and six months later a different model is better at the very tasks it bought the first one for.

This isn’t a failure of judgement. It’s the nature of the market. Frontier models leapfrog each other on a cadence measured in months, prices move by orders of magnitude, and strengths are uneven: one model reasons better over long documents, another writes cleaner drafts, another is cheaper by a factor of ten for routine work. Betting the organisation on any single point on that curve is like signing a decade-long contract for this year’s excavator model: the machine is fine; the commitment is the mistake.

Separate the permanent from the replaceable

The way out is an old engineering habit: separate what must last from what will be swapped. In a governed operating model, the lasting part is the gate: the policy layer every AI request passes through. Approved providers, for approved tasks, at a known cost, metered and logged. The gate encodes your rules, your permissions and your record. It doesn’t care whose model sits behind it.

Behind the gate, models become what they actually are: interchangeable machinery. Route long-document reasoning to the model that’s best at it this quarter; route routine drafting to the cheapest one that clears your quality bar; retire either the week something better arrives. Because every request already carries its evidence and lands in the record, swapping the machinery changes nothing about what the organisation can stand behind.

What this buys you

No lock-in. Provider negotiations happen from a position where leaving is an afternoon’s configuration, not a migration project. Early gains, safely. Each new model generation is adopted the week it clears evaluation, inside the same permissions, the same approval steps, the same audit trail, rather than after a year-long re-platforming. A stable answer for the board. “Which AI are we using?” becomes “whichever is currently best for each task, through a gate we control”, an answer that stays true no matter what the market does next.

This is why Square One is built model-agnostic. The platform’s commitments (permissions inherited from Microsoft 365, evidence attached to every output, approvals recorded, the ledger append-only) hold regardless of which provider answers a given request. The models will keep changing. Your operating model shouldn’t have to.