AI vendor lock-in is created on six surfaces — data formats and export rights, fine-tuned model ownership, prompt and workflow portability, dependency on a single model provider, integration coupling, and contract terms — and almost all of it is negotiable before signature and expensive after. The checklist below works through each surface: the questions to ask, and what a good answer looks like. The premise is simple: every AI contract prices your exit, whether or not anyone mentions it, so the exit belongs in the evaluation alongside the demo.
Lock-in deserves particular attention in AI purchases because the market itself is unstable. Vendors pivot, reprice, and get acquired — Portkey, an India-origin AI gateway, was acquired by Palo Alto Networks in May 2026, part of a broader consolidation of AI infrastructure into large security vendors — and the model landscape shifts quarterly. A dependency that looks safe today can change ownership, price, or direction within one contract term. You cannot prevent that; you can make sure it is survivable.
Surface 1: data formats and export rights
The gravity well of any AI system is the data that accumulates in it: your documents, extracted fields, labels and corrections, conversation logs, evaluation results. The lock-in question is not whether the vendor stores it but whether you can get it out in a form another system can use.
Ask, and get the answers into the contract:
- Can we export everything — inputs, outputs, labels, configuration — on demand, without a services engagement?
- In what format? Good: open, documented formats — standard JSON, CSV, plain files with a published schema. Bad: proprietary exports, or "the data is available via support ticket."
- What does export cost, and how long does it take? Good: self-service, included, days. Bad: professional-services rates and a quarter's notice.
- What is deleted after termination, on what schedule, with what certification?
Corrections deserve special mention. Every time your team fixes an extraction or approves an output, it creates training-grade labelled data. If those corrections live only inside the vendor's platform, your accumulated review labour — often years of it — is part of what you forfeit by leaving.
Surface 2: fine-tuned model ownership
If the vendor tunes, trains, or configures a model on your data, ownership must be settled in writing before work starts. The questions:
- Who owns the tuned model and its weights? Good: you do, or you hold a perpetual licence with the right to export and run it elsewhere. Bad: the vendor owns it and your contract merely rents access.
- Can your data improve models served to other customers? Good: no by default, opt-in at most, in writing. Your correction history is competitive information; treat it that way.
- If ownership is impossible — as it often is when tuning sits on a foundation model you cannot host — what is the rebuild path? Good: you keep the training data, evaluation sets, and configuration needed to reproduce equivalent behaviour elsewhere. That package is the practical substitute for ownership.
This surface is where the build-vs-buy decision casts its longest shadow: a bought system tuned on five years of your documents is neither fully bought nor fully yours unless the contract says so.
Surface 3: prompt and workflow portability
Modern AI systems accumulate a second codebase that rarely appears in architecture diagrams: prompts, retrieval configurations, agent and workflow definitions, evaluation suites. Months of operational tuning live there. The lock-in test is whether that layer exists as artefacts you hold or only as settings inside the vendor's console.
Good answers: prompts and workflow definitions are exportable as files, storable in your own version control, and written against documented interfaces rather than undocumented platform behaviour. Your evaluation sets — the test cases that define "working" for your use cases — are yours and portable, which is what makes switching testable rather than terrifying. Bad answers: the workflow "is the platform," and leaving means rebuilding from memory.
Portability here is never perfect — prompts tuned for one model degrade on another and need re-tuning. The goal is not zero switching cost; it is that switching is a re-tuning exercise against your own evaluation suite, not a rewrite from nothing.
Surface 4: dependency on a single model provider
Per-token API dependence on one provider is the quietest surface, because it is created by application code rather than by contract. If your systems call one provider's API directly, that provider's pricing, deprecation schedule, rate limits, and outages are hard-wired into your operations — and model deprecations in this market arrive on the provider's calendar, not yours.
The structural mitigation is an abstraction layer between applications and providers, so a model change is configuration rather than a code change across every system — the approach covered in multi-model AI routing. Note the recursion, and apply this checklist to it: a routing gateway is itself a vendor, and a proprietary one becomes a new single point of dependency. The same six surfaces apply.
What good looks like: no application hard-codes a provider; at least one credible alternative model has actually been tested against your evaluation suite, not just assumed; and per-provider spend is visible enough that concentration is a known number, not a surprise.
Surface 5: integration coupling
Every webhook, API call, and embedded component that ties the vendor into your ERP, CRM, or document flow is switching cost. Some coupling is the point — an unintegrated system delivers nothing — but its shape is negotiable. Prefer integrations built on standard interfaces to proprietary connectors; keep the integration logic in your codebase rather than configured inside the vendor's platform where you cannot take it with you; and keep a current map of every touchpoint, because the migration bill is roughly proportional to the length of that list.
Good answer to demand in diligence: a written integration inventory from the vendor's side — everything they touch, everything they store, every credential they hold. It doubles as your security review.
Surface 6: contract terms — termination assistance and IP
The final surface is the paper itself. Four clauses do most of the work:
| Clause | What good looks like |
|---|---|
| Termination assistance | Vendor obliged to support migration for a defined period at defined (non-punitive) rates — cooperation is contractual, not goodwill |
| IP ownership | Deliverables, configurations, and anything built on your data: yours, or perpetually licensed; "vendor retains all IP" boilerplate challenged |
| Change of control | Acquisition of the vendor triggers your right to exit without penalty, with export and assistance obligations intact |
| Renewal and repricing | Caps on renewal increases; no pricing cliff once your data and workflows are embedded |
The renewal clause is where lock-in is monetised: the vendor who knows your exit costs can price up to them at renewal. Everything else in this checklist — export rights, owned evaluation sets, portable workflows, provider abstraction — is what keeps that number, and therefore your renewal price, low.
Running the checklist without killing the deal
This is diligence, not hostility, and it fits inside a normal procurement cycle: put the six surfaces into the RFP as written questions, score the answers alongside capability, and negotiate the gaps before signature — the moment your leverage peaks. Vendors who answer crisply are usually the ones with nothing to hide; a vendor who resists export rights or goes vague on model ownership before you have signed is showing you the relationship after you have. The same posture belongs in services engagements — the wider set of questions before hiring an AI consultancy covers that side — and in product categories where your documents themselves are the asset, where export rights should sit high in your evaluation of document processing vendors.
And keep the honest frame: the goal is not zero lock-in, which mostly means zero depth. It is lock-in taken knowingly, priced before signature, and paid only where it buys something real.