On 24 March 2026, OpenAI announced it was shutting down Sora. The app went dark on 26 April; the API followed on 24 September — despite a $1 billion deal with Disney signed only months earlier (Wikipedia, CostGoat). Every business that had built a video pipeline on Sora had six months to rebuild it somewhere else.
That is what lock-in looks like in AI: not a slow price creep over five years, but a migration project you did not plan, triggered by someone else’s strategy meeting. This article explains why AI lock-in is different from classic SaaS lock-in, which layers of an AI system you need to own, the contract clauses that make ownership real, and a simple exit drill that tells you — before you need it — whether you could actually leave.
Why AI lock-in is different
Classic SaaS lock-in is mostly about data gravity and switching costs. AI adds four accelerants.
Models are retired quickly. Google’s Veo preview models in the Gemini API are being retired from 22 October 2026 at the earliest, replaced by newer versions (CloudZero). Luma moved all new users into a different product, Luma Agents, from September 2026 (Wikipedia). When Udio settled with Universal Music, users got 48 hours to download their tracks before the service became a “walled garden” (Wikipedia).
Pricing models change mid-stream. Bland AI switched to plan-based pricing on 5 December 2025. Synthflow now offers new customers enterprise contracts only, starting at around $30,000 a year (DevToolLab). HubSpot moved its Breeze agents to outcome-based pricing in April 2026 (Resolve247). HeyGen closed its Team plan in January 2026. None of these are scandals — but each one forces customers to recalculate.
Access can disappear. OpenRouter stopped serving an entire country in June 2026 (ProxyAPI). Vendors restrict regions, change acceptable-use policies and suspend accounts with little notice.
Vendors fail. Builder.ai raised hundreds of millions, including from Microsoft, and became insolvent in May 2025 after it emerged that much of the “AI” work was done by people and revenue was overstated (Wikipedia). Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027 (Gartner) — some of those cancellations will be vendors’, not customers’.
European buyers sense this. In surveys reported from Deloitte, 77% of companies weigh the country of origin of their AI vendor and 66% worry about dependence on foreign AI (DEV Community).
The seven layers you should own
An AI system is not one thing. It is a stack, and lock-in can hide in any layer.
| Layer | What to own | Lock-in symptom | How to keep it portable |
|---|---|---|---|
| 1. Code and workflows | Source code, automation flows, infrastructure config | “The agency’s account”, flows only in a vendor UI | Repository in your organisation; export flows (e.g. n8n JSON) to git |
| 2. Prompts and agent configs | System prompts, tool definitions, routing rules | Prompts editable only in the vendor dashboard | Versioned in your repo, deployed to the platform |
| 3. Evals | Test questions, expected answers, scoring rules | “We’ll know quality when customers complain” | Eval set in your repo, run on every change |
| 4. Data and indexes | Source documents, knowledge base, conversation logs | Only copy of knowledge lives in a vendor’s vector store | Source of truth outside; index rebuildable from it; regular exports |
| 5. Accounts and keys | Model API accounts, cloud, billing | Keys and billing under the agency’s name | Accounts in your company’s name; agency gets scoped access |
| 6. Channels and numbers | Phone numbers, WhatsApp Business account, domains | Number rented through the bot vendor | Numbers in your telephony account; bring your own SIP trunk |
| 7. Generated assets | Images, video, voice, 3D files and their licences | Free-tier outputs with vendor ownership | Paid plans with ownership terms; licence log per asset |
A few details from the market make the table concrete. Retell and Vapi let you connect your own SIP trunk at no platform charge, and Twilio’s BYOC trunking costs $0.004 per minute (Vapi, Twilio) — so your phone numbers don’t have to belong to your voice-AI vendor. On generated assets, licence terms differ sharply: on Recraft’s free plan the images belong to Recraft and commercial use is not allowed, while paid plans grant full ownership that survives cancellation (Recraft); Meshy’s free tier publishes models under CC BY 4.0, paid plans keep assets private and owned by the customer (Meshy).
Architecture patterns that keep you free
Put a gateway between your code and the models. Instead of calling one vendor’s SDK everywhere, route all model calls through a single layer that handles provider selection, fallbacks, budgets and logging. Switching models becomes a configuration change. We cover the design in LLM gateway: model routing, cost control and observability.
Use open standards for tools. The Model Context Protocol (MCP) has become the standard way to connect agents to systems. It was donated to the Agentic AI Foundation under the Linux Foundation on 9 December 2025, and there are more than 10,000 public MCP servers (MCP Blog, Digital Applied). A CRM connector built as an MCP server works with any MCP-capable model or agent framework. More in MCP explained.
Keep an open-weight fallback. For sensitive data or as an emergency route, an open-weight model hosted in the EU — Mistral is the common European choice (DEV Community) — removes dependence on any single API. Licences matter here too: FLUX.2 [klein 4B] is released under Apache 2.0, while the [dev] weights need a commercial licence for self-hosting (invideo). See private and on-prem LLMs.
Treat evals as portability insurance. The real reason teams fear switching models is not code — it is not knowing whether the new model is as good. An eval set of a few hundred real questions with expected answers turns that fear into a measurement: run it against the candidate model, compare scores, cost and latency, decide.
Self-host the glue where it makes sense. The community edition of n8n is free to self-host (n8n), which keeps workflows, credentials and execution data on your own server.
What portability costs — and what it saves
Portability is not free. A gateway, an eval set and versioned prompts add upfront work to a first project, and someone has to run the quarterly drill. Against that, weigh the alternative: an unplanned migration under a deadline set by a vendor, with no way to prove the new setup is as good as the old one. There is a cost argument too. Gartner expects spending on AI platforms and models to grow 63% in 2026, which makes budgets harder to predict (Gartner), and in KPMG’s Q3 2026 pulse 74% of large US companies had built cost reviews into AI approvals while 43% run explicit token budgets (KPMG). The ability to move a workload to a cheaper model when quality allows is one of the few levers you control.
Contract clauses that make ownership real
Ownership that is not in the contract does not exist. When an agency or freelancer builds AI systems for you, check these points:
| Clause | What to insist on |
|---|---|
| IP assignment | Code, prompts, configurations, evals and documentation assigned to you from day one, not after final payment months later |
| Accounts | Model, cloud and telephony accounts in your company’s name; contractor works with scoped access |
| Data | Your data is used only to deliver the service; no training on it; export in open formats on request and on exit |
| Deletion | Written confirmation of deletion at contract end, including backups and sub-processors |
| Change notice | Advance notice of model, vendor or price-model changes that affect you |
| Handover | Architecture overview, runbooks, credentials list, eval results — the “bus factor” package |
| Token costs | Either you pay providers directly, or re-billing with a transparent fixed markup |
| Exit assistance | Defined hours and rates for migration support |
Strong studios already work this way: some fixed-price MVP studios hand over code from the first day and put scope, price and warranty in writing before payment (HouseofMVPs). On tokens, the common market options are inclusion with a buffer, re-billing with a 15–25% markup, or direct payment by the client (Digital Applied) — the important thing is that you can see the real cost.
Copyright: you may not own what you think you own
Ownership of AI output is a legal question, not only a technical one. In the US, purely AI-generated content is not protected by copyright: on 2 March 2026 the Supreme Court declined to hear Thaler v. Perlmutter, leaving the human authorship requirement intact (Finnegan). The US Copyright Office states that prompts alone are not enough, while human selection, arrangement and modification can be protected (Copyright Office). In practice: use tools whose terms grant you rights, document human contribution for important assets, and secure your position through contracts rather than assuming copyright.
Buy vs build is not the same as own vs rent
None of this argues for building everything yourself. In 2025, 76% of enterprise AI use cases were bought rather than built (Menlo Ventures), and MIT NANDA found that buying from specialised vendors succeeds about 67% of the time versus roughly a third for internal builds (Fortune). Buying is usually right. The question is which layers you rent and which you keep. Rent commodity capabilities — models, transcription, telephony transport. Own the layers that encode your business: prompts, evals, data, workflows and customer channels. Our build vs buy framework goes deeper.
Five signs you are already locked in
- Nobody on your team can say which model version answers your customers today.
- The latest prompts exist only in a vendor dashboard or in a contractor’s head.
- Your support phone number or WhatsApp Business account is registered to the bot vendor.
- You could not produce last month’s conversations in a usable file within a day.
- The monthly AI invoice comes from your agency as one line, without per-provider costs.
If two or more apply, start with the cheapest fixes: move accounts into your name, export prompts and workflows into a repository, and build a first eval set from real conversations.
The exit drill: test your freedom quarterly
Once a quarter, spend half a day answering one question: could we switch the main model or platform within two weeks?
- Export. Pull prompts, workflows, the knowledge base and the last 30 days of conversation data. Did everything come out in usable formats?
- Rebuild. Rebuild the retrieval index from source documents in a test environment.
- Swap. Point the gateway at an alternative model.
- Measure. Run the eval set. Compare quality, latency and cost per conversation.
- Record. Write down the gaps and fix the cheapest ones first.
Most teams discover something on the first drill — a workflow that only exists in a UI, an API key on a former contractor’s card, a phone number registered to the bot vendor. Better to find it in a drill than during a shutdown notice.
Typical mistakes
- Prompts living only in a vendor dashboard. Your best business logic, with no version history and no export.
- Agency-owned accounts. When the relationship ends, so does your access — and your call history.
- No eval set. Every model change becomes a gamble, so nobody dares to change anything.
- Vector store as the only copy of knowledge. Indexes are caches, not archives.
- Free tiers in production. Licence terms on free plans often deny ownership or commercial use.
- Optimising only for the lowest token price. A slightly cheaper model behind a proprietary orchestration layer can be the most expensive choice once you need to leave.
Conclusion
Lock-in in AI arrives fast and from directions you do not control: model retirements, price-model changes, regional access cuts, vendor failures. The defence is not to avoid vendors — it is to own the layers that matter: code, prompts, evals, data, accounts, channels and assets, backed by contract clauses and a quarterly exit drill. Portability costs a little discipline up front and saves a migration crisis later.
This is the principle behind our name. At Lotus Libre, everything we build — AI systems and integrations and custom software — lives in your accounts and repositories, with evals and a handover package included, so you can switch models or partners whenever it makes sense. Want to know how portable your current AI setup is? Book a free 15-minute audit.
This article is general information, not legal advice. Licence terms and IP law vary by vendor and country; have important contracts reviewed by a qualified lawyer.