Insights Business| SaaS| Technology Where AI Exit Cost Accrues and What to Demand in Model Provider Contracts
Business
|
SaaS
|
Technology
•
Nov 18, 2026

Where AI Exit Cost Accrues and What to Demand in Model Provider Contracts

AUTHOR

James A. Wondrasek James A. Wondrasek
Where AI Exit Cost Accrues and What to Demand in Contracts

Model deprecation notices tend to arrive with about two weeks of warning. When one lands, the endpoint changes, and everything built on top of the model has to be re-validated before production can resume. Exit cost accrues in that re-validation.

Exit cost is the aggregate of re-prompting, re-evaluating, re-tuning and re-certifying work that leaving a provider forces. The exit cost of your AI stack sets the context, and the notice windows companion piece documents the actual deprecation notice periods you’re negotiating against.

Why does model deprecation matter, and where does the real exit cost accumulate?

Deprecation matters because it is a forced migration with a deadline someone else set. The model layer is the fastest-moving, most commoditised part of the stack, and treating it as permanent builds the exit cost in. The endpoint swap is a small part of the work.

Switching cost is the one-off migration bill, made up of the labour, parallel running and data conversion you pay once to get off a provider. Exit cost is the ongoing total that compounds the longer you stay. When a deprecation notice lands, you have to pay that total to migrate. Short-notice walk-aways are routine, and the re-validation that follows concentrates in the agent layer, made heavier by the data that has already accumulated around it.

Where does AI vendor lock-in actually live?

Vendor lock-in concentrates in the agent layer. It comes from the components your team built around the model, including prompt schemas and prompt libraries, tool definitions and MCP servers, guardrails, agent state and memory, evals, and multi-step agent workflows. Stacklok’s survey found the agent layer stickier than the model, with MCP integrations hardest and prompt libraries last.

Fine-tuning is the asset you cannot move. The weights do not transfer, so a fine-tune is irreproducible without re-training and re-evaluating, which makes it the strongest lock-in. The harness is the layer around the model. Your team builds vendor-specific habits and integrations into it, so moving to another provider means retraining people and rebuilding those integrations, and that rebuilding is the harness lock-in.

Embeddings are model-specific, so changing the embedding model means re-embedding the whole corpus, and permissions, audit history and integrations have already accumulated around the data. Running more than one model doesn’t move the built layer, and 92% of teams already run more than one.

Why do leaders rank lock-in last when they actually buy AI infrastructure?

Portability is a property you have to engineer, and it’s not granted at signing. Stacklok’s survey of 520 leaders found 84% say avoiding lock-in matters, yet switching ability ranked sixth of eight factors and vendor independence ranked last. The consequences show up: 81% have delayed, scaled back or restructured a migration, 57% estimate four months or more, and 3% could not migrate at all.

Leaders’ stated priorities diverge from their buying behaviour, and this pattern predates AI. A company migrating off Salesforce can accidentally build its own Salesforce clone. Every “Splunk is too expensive” migration ends with a permanently understaffed team and frustrated users with less capability than before. Kubernetes made compute portable, Terraform made a single configuration run on any cloud, and ONNX gave models a common interchange format. Apache Iceberg separated storage from compute so data could move between warehouses, and Airflow kept orchestration running across platforms. AI has no complete analogue yet. The procurement template for avoiding lock-in comes from modular contracting and performance-based acquisition, a practice born in the United States and Canada that breaks a project into interoperable pieces, ties payment to outcomes, and keeps more than one qualified supplier in play. Infrastructure-as-code and containerisation round it out, making deployments reproducible across providers.

How do you assess which parts of your AI stack are hardest to move before a renewal?

Price each asset class by the re-certification effort it would take to move it, separating the endpoint from prompts, evals, fine-tunes, gateway semantics, embeddings and the vector store. Prompts are the cheap class to move. Evals and fine-tunes are expensive to move, and MCP integrations and multi-step workflows are hard to move because they encode business process. A mid-market forced switch prices the classes at 15 to 30 person-days to rebuild prompts and pipelines, a week or two of compute to re-vectorise the knowledge base, and 5 to 10 person-days to retrain the team.

A gateway lets you swap the endpoint while the built layer stays put. It decouples application logic from provider APIs and enables failover, but it only helps if something behind it survives.

How do you get an honest estimate of how long a migration off your current model provider would actually take?

Start from the honest baseline that Stacklok’s data provides: 57% of leaders estimate four months or more, 30% one to three months, and 3% say they could not migrate at all. Only 10% estimate less than a month, so a sub-month estimate is the outlier.

Start by pricing the agent layer. Re-certifying is the slowest part, because human sign-off is slower than compute, and re-embedding is a full-corpus job. During a real migration, three things surface that one migration account spells out: quality was never measured, the prompt is coupled to the old model, and model behaviour was never versioned.

Walk it through a tabletop exercise, shadow-run the fallback, and keep a current dependency map. Run the Data Export Completeness check at proof-of-concept, the point where you are still trialling the provider and switching costs nothing. The check tests whether every artefact your team created exports in a format another platform can ingest, so the estimate reflects what leaves.

What should you demand in a model provider contract to protect your exit runway?

Demand a notice-period guarantee, model-lifecycle disclosure, fine-tune export rights, scoped transition assistance, and data portability you can test. Parity of model availability across channels and regions and model-version support SLAs complete the picture.

Each contract demand corresponds to a specific architectural risk. The notice period mirrors the short-notice deprecation risk: OpenAI’s policy promises up to six months while its June 2024 country block delivered roughly two weeks, and Anthropic’s 60-day policy sat alongside its cutoff of Windsurf, a third-party tool whose direct access to Claude models was reduced with what its CEO called “little notice”. Those are the windows to negotiate against, because the published policy rarely matches what happens.

Model-lifecycle disclosure corresponds to the risk of a model that stays available while its output degrades. A model in that state is functionally impaired even when it meets a conventional uptime SLA, and service levels should cover model version lifecycles. Fine-tune export rights mirror the weights that don’t transfer, so the contract specifies substitute deliverables like the tuning dataset, prompts, evals and telemetry.

Once you give notice, you lose your negotiating position on price, so transition assistance has to be a scoped, priced service fixed at signing. “The parties will cooperate in good faith” has no scope, no deliverable and no price. Data portability is tested through the Data Export Completeness check, and parity and the support SLAs are anchored to the actual notice windows.

Exit cost accrues as re-validation work in the agent layer, and the say-do gap is why it keeps going unmeasured until a deprecation notice makes it real. Pricing the stack by re-certification effort before renewal shows you the hardest assets while you still have the power to negotiate the clauses that protect them.

Contract language converts a notice-period promise into an enforceable runway, and that runway has to be tied to the documented notice periods in the notice-period piece and the broader exit-cost analysis. Whether gateways and open-weight fallbacks remove that cost or relocate it is taken up next.

Frequently Asked Questions

Does running multiple models already protect me from vendor lock-in?

No. Running several models makes the endpoint easy to swap, but it does nothing for the agent layer you have built around each one. Your prompts, evals, guardrails and multi-step workflows still have to be re-validated when a model disappears. Stacklok’s data shows 92% of teams already run more than one model, yet lock-in persists because the stickiness lives in the built layer, not the model itself.

What is a fine-tune, and why can’t I move it to another model?

A fine-tune is a model you have adapted on your own data so it behaves to your specification. The trained weights do not transfer to another model, which makes a fine-tune irreproducible without re-training and re-evaluating from scratch. It is the highest-mobility-cost asset in the stack, which is why your contract should specify substitute deliverables like the tuning dataset, prompt libraries and evals when the model is deprecated.

What happens to my embeddings and vector store if I switch embedding models?

Embeddings are model-specific, so changing the embedding model means re-embedding your entire corpus before your vector store is useful again. That is a full-corpus re-computation, not a configuration change. Data gravity compounds the problem: permissions, audit history and integrations have already accumulated around the existing data, so the re-embedding is only the visible half of the migration.

How is data gravity different from vendor lock-in?

Vendor lock-in is the commercial constraint a provider creates. Data gravity is the physical and process weight of your data as permissions, audit history, embeddings and integrations accumulate around it. Lock-in is what you negotiate against in a contract; data gravity is what you engineer against in your architecture. Both raise exit cost, but they respond to different tools.

Why do providers deprecate models with so little notice?

Providers deprecate models for cost, safety and infrastructure reasons, and those decisions are made on the provider’s commercial timetable, not yours. The precedents are documented: OpenAI’s June 2024 country blocks landed with roughly two weeks’ warning, and Anthropic has cut off models with what it called little notice. Short-notice walk-aways are routine, which is why a notice-period clause has to be fixed at signing.

How much notice will I actually get before a provider deprecates a model?

Assume two weeks unless your contract says otherwise. OpenAI’s June 2024 country blocks gave roughly two weeks’ notice, and Anthropic’s Windsurf cutoff arrived with little notice at all. Those are the real windows to negotiate against, not the marketing assurances. The actual per-provider notice periods are catalogued in the companion piece on deprecation notice windows, so you can anchor your clause to a documented number.

When should I start testing whether my data will actually export?

At proof-of-concept, not renewal. Run a Data Export Completeness check early so you know what actually leaves the platform in an ingestible format before you have committed. Testing at renewal is too late, because the estimate is no longer grounded in what the export really contains. The moment of maximum leverage is before signature, and that is when the export should be verified.

What does re-certification actually involve in an AI migration?

Re-certification is re-validating that the migrated system still meets your compliance, safety and quality obligations under a new model. Guardrails have to be re-tested, output quality re-checked, and any regulated or customer-facing behaviour re-approved. It is the slowest part of the re-validation work because it depends on human sign-off rather than compute, and it is exactly what gets under-priced when a deprecation notice lands.

What if my provider refuses to include these exit clauses in the contract?

Treat the refusal as data, not a dead end. If a provider will not commit to a notice period, lifecycle disclosure or fine-tune export rights, the price they are really quoting is the full migration cost they expect you to pay later. Walk the ask through the procurement pattern of modular contracting and performance-based acquisition, and if the clause stays off the table, re-price the deal accordingly before you sign.

Is a notice-period promise in a contract actually enforceable?

Yes, if it is written as a specific obligation rather than a marketing assurance. A notice-period guarantee tied to a documented number of days, with a remedy for breach, converts a runway promise into an enforceable runway. Vague comfort language is not enforceable and is worth nothing. The clause is only as strong as the number and the consequence attached to it, so both have to be fixed at signing.

Can self-hosting an open-weight model eliminate my exit cost?

Not by itself. Self-hosting removes the provider’s control over the endpoint, but your agent layer still has to be re-validated when the model changes, and open-weight models have their own lifecycle risks. The cost is relocated, not removed. Whether an open-weight fallback genuinely removes the cost or only moves it is explored in the companion piece on open-weight models and gateways.

AUTHOR

James A. Wondrasek James A. Wondrasek

SHARE ARTICLE

Share
Copy Link

Related Articles

Need a reliable team to help achieve your software goals?

Drop us a line! We'd love to discuss your project.

Offices Dots
Offices

BUSINESS HOURS

Monday - Friday
9 AM - 9 PM (Sydney Time)
9 AM - 5 PM (Yogyakarta Time)

Monday - Friday
9 AM - 9 PM (Sydney Time)
9 AM - 5 PM (Yogyakarta Time)

Sydney

SYDNEY

55 Pyrmont Bridge Road
Pyrmont, NSW, 2009
Australia

55 Pyrmont Bridge Road, Pyrmont, NSW, 2009, Australia

+61 2-8123-0997

Yogyakarta

YOGYAKARTA

Unit A & B
Jl. Prof. Herman Yohanes No.1125, Terban, Gondokusuman, Yogyakarta,
Daerah Istimewa Yogyakarta 55223
Indonesia

Unit A & B Jl. Prof. Herman Yohanes No.1125, Yogyakarta, Daerah Istimewa Yogyakarta 55223, Indonesia

+62 274-4539660
Bandung

BANDUNG

JL. Banda No. 30
Bandung 40115
Indonesia

JL. Banda No. 30, Bandung 40115, Indonesia

+62 858-6514-9577

Subscribe to our newsletter