A founder I talked with this summer had done everything right, or at least everything the market told him to do. He picked a vendor early, committed properly instead of dabbling, and spent about eighteen months teaching the thing his business. His team corrected it constantly. They fed it their pricing logic, their objection handling, the way they talk to a general contractor versus a facilities director. By spring it was genuinely good, and it was good because it held eighteen months of their corrections.
Then a better model shipped, priced roughly a third of what he was paying. He went to move.
What he could export was a JSON file of conversation transcripts. Everything that made the system worth having existed only as tuning inside a product he was trying to leave. He is still on the old vendor. Not because it is better, but because leaving means starting over.
That is the real lock-in of this era, and it does not look like the lock-in anybody prepared for.
Nobody Is Standing Still, Including You
The instinct is to treat this as a bad-vendor-choice story. It is not. Switching is the normal condition of this market, and the data is unambiguous.
Menlo Ventures found that Anthropic's share of enterprise LLM API usage reached 40% in 2025, up from 12% two years earlier, while OpenAI fell to 27% from 50% in 2023. Those are enormous moves in twenty four months, and they represent real companies really migrating. Multi-model is now the default posture rather than a hedge: a16z's survey of enterprise CIOs found 37% now run five or more models in production, up from 29% the year before, and Menlo's earlier report found organizations typically deploy three or more foundation models, routing by use case.
The economics guarantee the churn continues. Stanford HAI's AI Index tracked inference cost for a given capability level falling from $20.00 per million tokens to $0.07, a drop of more than 280 times in about eighteen months. When the price of what you are buying falls that far that fast, staying put is not loyalty. It is a tax.
Subscribe
Get notified by email when we publish a new post. No spam, unsubscribe anytime.
Most companies know they are exposed. In a survey of 542 US enterprise decision-makers commissioned by Zapier, 81% reported at least some concern about dependency on specific AI vendors, and 74% said losing their primary vendor would disrupt operations or that they were completely reliant on it. The gap between confidence and experience is the telling part: 89% believed they could switch within a month, but of the 66% who had actually attempted a migration, only 42% called it smooth.
Lock-In Moved from the Data to the Interpretation
The reason those migrations go badly is that most people are guarding the wrong asset.
We spent twenty years learning to demand data portability, and we mostly won that fight. You can get your CRM records out. You can export your invoices, your tickets, your contacts. Everyone asks about it in procurement now.
Then AI arrived and the valuable thing quietly moved. What matters now is not the records. It is the layer on top of them: which customers are exceptions and why, what your team corrected the model on last March, which past decision turned out badly, the reasoning that sits behind the entries rather than in them. That layer is what makes a system useful for your specific business instead of generically capable, and it is almost never in the export.
McKinsey has made the same point about where advantage comes from, noting that when everyone adopts the same tools, differentiation has to come from proprietary data, and that roughly 90% of an enterprise's data is unstructured. The structured records were always portable. The unstructured interpretation, the part that took you two years to accumulate, is the part sitting inside somebody else's product.
This is the same argument I made about competitive advantage in Context Is the Whole Game, turned around and pointed at a risk instead of an opportunity. If context is the moat, then context living inside a vendor means the moat is not yours.
The Model Is the Rental, the Context Is the Property
There is a useful way to think about the split, and it survives contact with procurement.
The model is a utility. It will get better, it will get cheaper, and you will change it, probably more than once in the next three years. Treat that as a given rather than a failure of planning. Nobody feels bad about switching electricity suppliers.
Your context is property. The corrections, the exceptions, the reasoning, the record of what worked, all of it is an asset your business built and should own outright, in a form you can read without the vendor's help. So the right question in an evaluation is not "how good is this model." Models converge, and by the time the pilot ends the answer has changed anyway. The question is: if we leave in two years, what do we walk out with?
Most vendors have never been asked that directly. The answer you want is an exportable representation of everything the system learned about your business, in a format that means something outside their product. The answer you will often get is a transcript archive, which is like being handed a recording of every meeting instead of the decisions that came out of them.
What Portable Context Actually Requires
Three things, in my experience, and none of them are exotic.
Corrections have to be stored as records, not as tuning. When someone tells the system it got something wrong, that correction should exist as a durable, readable entry attached to the thing it corrected. If it lives only as an adjustment inside a model's behavior, it is not yours and it does not survive the migration. I wrote about why a system has to keep learning from its own corrections in RAG Is Only Half the Story. What that post treated as a quality question is also an ownership question.
Provenance has to travel with the answer. Every conclusion should carry where it came from. Without that, an exported context layer is a pile of assertions you cannot check, which is worthless the moment you load it somewhere else and cannot tell what is still true.
The context has to be model-agnostic by construction. If your structured knowledge only makes sense when read by one particular system, you have built a dependency and called it an integration. Written down properly, your context should be legible to a competent human, which is a decent proxy for being legible to whatever model you use next.
Do that and model choice becomes a procurement decision rather than a strategic bet, which is what it should be. It is a large part of what I meant in The AI-Native Company by intelligence being structural. Structure that only works with one vendor is not structure. It is a rental agreement.
The Takeaway
The AI market is going to keep moving. Share shifted by tens of points in two years, most enterprises now run several models at once, and the cost of a given capability fell more than two hundred fold in about eighteen months. You will change vendors. Plan on it.
What should not change is everything your company has learned about itself along the way. If eighteen months of corrections and exceptions and hard-won judgment cannot leave with you, then you did not build an intelligence layer. You rented one, and the rent is about to go up.
Ask your vendor what you walk out with. The answer tells you whether you are building an asset or feeding someone else's.
Tracy Thayne* is the founder of Expona, an AI-powered operational intelligence platform for B2B marketing. Read the Expona founder story or subscribe to the blog (below) for weekly insights on context, AI, and the operating model of the next decade.*
