Don't marry the model

Every AI pitch leads with a brand name. But models age like phones, not buildings — and the one question that matters is what happens the day you want a different one.

Every AI pitch you'll hear this year leads with a name — this system runs on so-and-so's latest model, as though that settles the matter. Here is the awkward truth behind the brochure: whichever model it is, it will not be the best one for your job in eighteen months, and it may not even be the best one by the time the contract is signed. The market re-shuffles every few months. Prices drop, new entrants arrive, yesterday's leader has a bad quarter. So the question that actually decides your risk isn't "which model does it use?" It's "what happens on the day we want a different one?"

Models age like phones; the system around them is the building

It helps to separate two things the pitch runs together. The model is the reasoning engine — the part that reads a document and drafts a reply. The system is everything around it: the secure connections into your inbox and your records, the log of every action, the approval step where a person signs off, the rules about what waits for review. The system is the building — expensive, durable, worth doing properly. The model is more like the boiler: essential, replaceable, and improving somewhere else even as yours runs perfectly well. A system built so the boiler can be swapped is normal engineering. A building that has to be demolished to change the boiler is a design fault — and in the AI trade, that fault has a sales pitch.

How the trap gets built

Lock-in rarely arrives as a clause in bold. It gets built, quietly, when a supplier wires every part of the workflow to one provider's particular product — its exact way of being asked, its exclusive features, its formats. Nothing about that is malicious; it's simply quicker. But the bill arrives later: when the provider raises prices, changes terms, retires the version you depend on, or has an outage during your busiest week — and "let's switch" turns out to mean "let's rebuild". At that point you're not a customer anymore. You're a hostage with an invoice.

The alternative costs a little discipline up front. Build the workflow so it asks for a drafting model or a checking model — a role, not a brand — and which company answers is a setting, not an architecture. We say this with the confidence of the recently burned: mid-project, an account snag at our own AI provider stopped our newsroom cold. Because the pipeline was written against roles rather than one company's product, moving the whole thing to a different provider took one evening — a settings change and a night of testing, not a rebuild. That evening is the entire argument of this article.

The one question to ask any supplier

Put it exactly like this, and put it in writing: "If we wanted to change the AI provider next year, what would that involve?" The good answer is boring: a configuration change, a round of testing, perhaps a day or two of tuning. The bad answers announce themselves. A rebuild quote. A contract that bundles the software and the AI provider into one inseparable lump. Or the most revealing one — "why would you ever want to?" — which tells you the supplier has never had to, and that you'll be the test case. While you're at it, ask where your data and logs live and in what format; a system you can leave is also a system whose records you can take with you.

What swappability actually buys you

Four things, all of them money or safety. Negotiating power: a supplier who knows you can switch prices accordingly; so does an AI provider. Resilience: outages, policy changes and account problems at one provider become an inconvenient evening rather than a stopped business. The right tool per task: the routine sorting in a workflow can run on a cheap fast model while the judgement-heavy step uses a stronger one — our own pipeline does exactly this, and it's most of why an illustrated article costs us pennies. An honest renewal: when the contract comes round, you're comparing options, not calculating the cost of escape.

The honest caveats

Swappable isn't free, and anyone who says otherwise is selling something. The separating layer costs a little extra build time at the start — think a few per cent, not double. Models have personalities, so after a swap the prompts need retuning and the outputs need a fortnight of closer review; an evening got us running, but the fine-tuning trailed on for days. And some provider-specific features genuinely are better than the generic alternative — using them is fine, so long as somebody keeps a list of exactly what you're leaning on, because that list is your future switching bill. The point was never to avoid commitment. It's to make sure the commitment is to your workflow, your data and your logs — the parts that are actually yours. Marry the system. Date the model.