A 30-staff financial services firm I spoke with last quarter had spent fourteen months building an internal compliance assistant on Claude. The prompts were tuned for Claude’s behaviour. The retrieval layer was embedded with the recommended embedding model. The fine-tuned helper that summarised FCA-language documents was trained on Anthropic’s API. Then OpenAI shipped a feature the team genuinely needed, and the operations director asked the obvious question: what would it take to switch?
Her engineering lead came back with a number. Prompt rewrites, three to four weeks of senior time. Re-embedding the knowledge base into a different vector space, two weeks of compute plus re-evaluation. Rebuilding the fine-tune, twelve to eighteen thousand pounds. Rebuilding the evaluation harness, because the two models scored differently on the firm’s own benchmarks. Total: roughly 67,000 pounds and four months. The firm decided not to switch. The procurement decision they had made fourteen months earlier had committed them to an architecture, which is a bigger commitment than picking a vendor.
That is what vendor lock-in actually looks like in 2026.
What is vendor lock-in?
Vendor lock-in is anything that makes switching providers expensive or impractical. With AI it operates differently to traditional cloud lock-in. Cloud lock-in is mostly about visible infrastructure: compute, storage, networking, contracts. AI lock-in is mostly about decisions you have already made: how prompts are written, what data you have embedded, which agent framework you have chosen, which model your fine-tunes target. None of those show up on a bill until you try to leave.
Some lock-in is unavoidable in any meaningful AI deployment. The useful question is not whether you have lock-in. The useful question is which kind, how much, and whether you priced it before you signed. CloudZero’s 2025 analysis put it cleanly: AI lock-in is API-driven rather than infrastructure-driven, embedded inside product features rather than sitting as a standalone system, and that is what makes it harder to detect early.
Why it matters for your business
It matters because the cost of leaving compounds quietly across five vectors that owners rarely price up front. First, prompt engineering: prompts tuned for one model do not behave the same on another, and a 2025 academic study reported prompts dropping 68 percentage points in performance when moved between model families without re-tuning. Second, fine-tuning: fine-tunes are model-specific by design and do not transfer.
Third, RAG knowledge bases: the embedding space created by one vendor’s model is incompatible with another’s, so re-embedding a sizeable corpus is not optional if you switch. Fourth, agentic frameworks: each provider’s tool-calling format differs (OpenAI function calling, Anthropic tool_use, Google function declarations), and migration touches every agent. Fifth, contractual: multi-year commits, exit fees, and the cloud-credit nudge where Azure or AWS credits flow only to their hosted models.
The cloud-credit pattern is particularly subtle. SMEs with Microsoft Azure credits naturally gravitate to Azure OpenAI Service. Once embedded there, switching means losing credit value. Holland & Knight and Morgan Lewis have both flagged the same risk on the contractual side: exit mechanics in AI agreements are routinely thinner than in traditional outsourcing, leaving the customer stranded if pricing or roadmap changes.
Where you will meet it
You will meet vendor lock-in in three places that look different to the team and the same to the balance sheet. The first is the procurement conversation, where the lock-in is largely invisible. The vendor demos a working tool, the team gets excited, and nobody asks the migration questions because the system is not yet live. The architectural commitment is made here. It is also the cheapest moment to price the lock-in honestly.
The second is the eighteen-month review. By that point the prompt library has been tuned, the RAG corpus is embedded, agents are wired to one provider’s tool-calling format, and the team has built genuine expertise in one model’s quirks. A cheaper or better alternative appears, the team estimates the migration, and the answer is uncomfortable. Cost-savings of a few hundred pounds a month do not pay back forty thousand pounds of engineering effort. The decision is made for you.
The third is the renewal. The contract is up, the vendor proposes a price increase or a new minimum commit, and your bargaining position is gone. Without a credible alternative inside the firm’s stack, the negotiation is short. This is where the contractual layer (multi-year auto-renewals, 90-day notice windows, remaining-balance termination fees) does its real work. RMOK Legal has documented several mid-market firms locked into a third year of pricing they would never sign today, simply because the renewal clock ticked past their notice window.
When to care about it, when to ignore it
Care about it whenever your AI spend crosses 25,000 pounds a year, or when the deployment is touching a regulated process or a customer-facing product. At that scale the switching cost is real money and the architectural choices have compounded enough to matter. Architect for partial portability from day one: an abstraction layer between your code and the vendor API, an open-weight option as a credible alternative, a re-embeddable RAG design, and a tool-calling abstraction.
Ignore it when the deployment is small, the spend is under a few hundred pounds a month, and the system can be rebuilt in a fortnight if the vendor disappears. A pilot, a personal-productivity tool, a small internal helper. Trying to architect those for portability slows the team down and costs more than the lock-in you are avoiding. The honest answer for a 5,000-pound-a-year deployment is to accept the lock-in, ship the value, and revisit at year three.
The middle ground is where many firms sit, and it is where the procurement questions earn their keep. Before you sign anything material, ask the vendor what migration tools and documentation they provide, what their model-deprecation policy is, whether they support open-weight alternatives on their infrastructure, and what the exit terms actually look like. Vendors who answer cleanly are signalling commercial maturity. Vendors who deflect are telling you that lock-in is part of how the product works.
Related concepts
Total cost of ownership, or TCO, is where vendor lock-in shows up on the spreadsheet. Switching cost is a TCO line, even if your finance system has no place to record it. A clean TCO model for an AI deployment includes a switching-cost estimate alongside running cost, integration cost, and ongoing tuning. Without that line, the procurement conversation is structurally optimistic.
Hybrid pricing interacts with lock-in through multi-year commits and minimum spend tiers. The vendor offers a discount in exchange for predictability, and the predictability is yours to give. A three-year commit at 25 percent discount looks attractive in month one and looks expensive in month fourteen when a better model ships. Negotiate the exit terms before you accept the discount.
Open-weight models (Llama, Mistral, DeepSeek, Qwen) are the structural hedge. Because the weights are public, you can host the same model on AWS Bedrock, on Together, on Replicate, or on your own infrastructure. That removes lock-in at the model layer. It does not remove prompt-engineering or RAG lock-in, and it shifts operational burden onto your team. For many SMEs the right pattern is a tiered approach: a proprietary API for fast-moving small workloads, an open-weight fallback for high-volume or data-sensitive ones.
The Model Context Protocol (MCP) is the emerging standard for connecting tools and data to language models. As an open protocol it is designed to reduce lock-in. In practice each vendor’s implementation has its own quirks, so optimising hard for one provider’s MCP behaviour can still create stickiness. Treat MCP as a hedge worth having, not a guarantee of portability.
The honest test of any AI procurement decision is the migration question. If you cannot answer it cleanly today, you have not yet priced the lock-in. The work is to price it before you sign, not after.
