The Real Cost of Vendor Lock-In in AI

Every serious discussion of lock-in ends up as an estimate of the migration bill. That number prices an event most organisations will never trigger, while the cost they actually pay accrues silently in the requests their own people quietly stop making.

In a quarterly planning session at a large distributor, an operations manager opens a document to write up an idea she has been carrying for a month: a way to have incoming supplier correspondence read, reconciled against open purchase orders, and escalated only when the discrepancy is material. She gets two paragraphs in, rereads them, and deletes the file. Nothing about the idea is unreasonable, and she has not been told no. She simply knows, from three years of watching how these conversations go, roughly what will happen — the capability is not in the platform's current shape, the integration would need an interface nobody exposes, and the last two proposals of this kind dissolved into a roadmap discussion and then into nothing. The idea does not become a request. It is never scoped, never rejected, never logged, never counted, and by the time the planning cycle closes there is no artefact anywhere in the organisation recording that it existed.

That deleted document is the real cost of vendor lock-in, and it appears in no analysis of it. Ask a technology leader what lock-in costs them and the answer will be an exit estimate — data extraction, reimplementation, retraining, a parallel run, the professional services to get it all across. Those are legitimate figures and people compute them carefully. But they price a transaction most organisations will not execute, in a year that keeps receding. Meanwhile the organisation that stays — which is nearly all of them, nearly all the time — is paying something else entirely, every quarter, in a currency nobody has an account for.

The exit estimate prices the one thing you will probably never do

There is a ritual quality to the way migration costs get modelled. Someone builds the case, the number comes back large, and the conversation resolves in relief: leaving would be painful, therefore staying is correct, therefore the question is settled for another cycle. Occasionally the number comes back smaller than feared, and the relief takes a different form — we could leave if we needed to, so we are not really locked in. Either result terminates the inquiry, which is precisely what makes the exercise so comfortable and so useless. Both answers treat lock-in as a boundary condition that only matters at the moment of departure, when it is in fact a force acting continuously on everything the organisation decides to attempt.

The estimate also has a subtler defect that almost nobody catches. It prices the migration of what you currently have — the integrations you built, the data you accumulated, the processes you encoded — and by doing so it silently accepts your current footprint as the natural extent of the work. But the footprint is itself an artefact of the platform. It is the subset of your ambitions that survived contact with what the platform made easy. A faithful accounting of lock-in would have to compare your actual estate against the one you would have built had every capability been equally available, and no finance function on earth has a method for costing a comparison against a thing that never existed. So the estimate quietly narrows itself to the part that can be counted, and the part that can be counted is the small part.

Organisations learn, and what they learn is not to ask

An enterprise is a learning system, and like any learning system it responds to reinforcement rather than to policy. When a request is met, people make more requests of that kind. When a request meets friction — a long qualification process, a partner engagement, an architectural review that ends in a suggestion to use the supported pattern instead — people do not usually complain. They adapt. The next specification is written a little closer to what the platform natively does, and the one after that closer still, until the vocabulary of what the business asks for has been trained, over several cycles, on the contours of a single product. Nobody issues a memo announcing that ambition has been rescoped. The rescoping happens in the drafting, in the pre-meeting, in the moment an analyst decides which version of the idea is worth the political capital.

What makes this so difficult to see from inside is that it presents as maturity. Teams that have learned not to ask describe themselves as disciplined, realistic, focused on adoption rather than novelty, sensibly buying rather than building. Some of them genuinely are — constraint chosen deliberately is one of the more valuable things an engineering culture can develop. The difference between chosen constraint and absorbed constraint is entirely a matter of whether anyone still evaluates the excluded option on its merits, and that difference leaves no trace in any artefact. Two organisations with identical roadmaps can be in completely different conditions, one having considered and declined a set of possibilities, the other having stopped being able to see them.

The practical consequence is that lock-in is not principally contractual or technical. It is the accumulated inventory of things the people inside a company have concluded are not worth proposing. That inventory grows with every quarter of successful, unremarkable operation, because success is exactly what teaches everyone where the easy paths run. It is invisible during good times and it is invisible during bad times, and it is the mechanism by which a company can be extremely satisfied with a platform for six years and still emerge from those years having attempted a fraction of what it might have.

Nothing that was never proposed appears in a business case

Corporate accounting is built to track transfers: a cost has a counterparty, an invoice, an owner, a line. A foregone option has none of these. It generates no spend, triggers no variance, breaches no control, and shows up in no post-implementation review, because there was no implementation. This is why the industry can produce confident numbers about failure and almost nothing about absence. When Gartner predicts that more than forty percent of agentic AI projects will be cancelled by the end of 2027, citing escalating costs and unclear business value, it is measuring initiatives that at least got proposed, funded, staffed and killed. Every one of those left a record. The proposal that died in a planning session, or in the ten minutes before one, is beneath the resolution of every instrument we have.

The incentive gradient runs the same direction. A migration that goes badly is vivid, attributable, and career-relevant; the person who championed it owns the outcome. An ambition that quietly contracts is attributable to nobody, produces a calm roadmap and a healthy renewal, and is generally read as good governance. Given a choice between a legible risk and an illegible one, institutions choose the illegible one every time, which is not a failure of intelligence but a structural feature of how organisations perceive themselves. It also explains why the lock-in conversation never improves. The instrument that would detect the real cost would have to measure a counterfactual, and no procurement process has ever been asked to do that.

None of this is an argument against platforms, and it would be dishonest to pretend it exempts the one I work on. StudioX accrues exactly the same gravity as anything else in the category: once an organisation's knowledge, its missions, its human-in-the-loop gates and its model access all run through one system, that system becomes the default answer to every new question, and the default answer is where ambition goes to get shaped. Certain choices genuinely lower the wall — an open protocol surface like the Model Context Protocol so capabilities are not hostage to one vendor's connector catalogue, an LLM gateway so model choice remains a decision rather than an inheritance, enterprise knowledge that leaves in a usable form. Those things reduce the exit cost, which is the visible cost, and they do very little about the invisible one. What makes any platform valuable is that its defaults work, and defaults that work are precisely the mechanism by which an organisation's sense of the possible gets quietly bounded. The recurring theme across the reporting on how autonomous enterprises actually get built is that the binding constraint is rarely the technology on offer; it is what the organisation has learned to believe is on offer.

So the useful test is not what it would cost to leave. It is to sit down with the people closest to the work and ask them to write the list they would write if they assumed every answer would be yes — not the roadmap, not the backlog, but the ideas that never made it as far as a ticket. The gap between that list and what your organisation actually asked for last year is the true price of your lock-in, denominated in the only unit that matters. Run it annually and it becomes a health metric with real teeth, because a platform relationship in good condition is one where the ambitious version of the request still gets written down, argued about, and sometimes refused on its merits. When the ambitious version stops getting written at all, you are no longer choosing a platform. You are being described by one.