In January 2025, the European Union's digital operational resilience regulation for the financial sector, known as DORA, came into force. Among other requirements, Regulation (EU) 2022/2554 now obliges banks and insurers to keep a documented exit strategy for every critical technology vendor — a plan for migrating without stopping the operation, written before it's needed.
In the United Kingdom, the Bank of England and the Financial Conduct Authority followed the same path. Since January 1, 2025, vendors deemed critical to the UK financial system must show how they manage their own chain of dependencies and how an eventual exit would be handled without threatening the sector. The regulator itself recorded that issues tied to third parties were the leading cause of operational incidents reported between 2022 and 2023.
The logic behind these rules isn't exclusive to banks. Any company that builds its AI operation on top of a single vendor runs the same risk regulators are addressing: discovering, at the moment it needs to switch, that leaving is far more expensive and slower than joining was. And whoever decides where AI runs inside the company rarely thinks about this before signing the first contract.
The Bill That Only Shows Up at the Exit
A Zapier survey of 542 US executives, cited by The Register in April 2026, shows the gap between expectation and reality: 90% believed they could switch AI vendors within four weeks, and 41% expected it to take just a few business days. In practice, only 42% of companies that attempted to migrate from one AI platform to another say the move went smoothly.
The survey's own explanation is blunt: once AI is already woven into internal processes, connected to other systems, and tuned to specific workflows, it carries dependencies, edge cases, and small adjustments nobody documented. Each of those adjustments stays invisible until the day the vendor changes — and turns into rebuilding work.
And vendors change more often than companies expect. According to the same report, Anthropic changed its Claude enterprise billing in April 2026, moving away from fixed pricing to usage-based charges — a shift that, according to experts the outlet spoke with, could double or triple costs for heavy users. OpenAI, for its part, raised the price per processed unit for its flagship model from $1.25 to $5.75 between one version and the next.
None of these changes are announced with enough notice for a company to switch vendors at the same pace. And when someone on the leadership team asks why the company hasn't switched, the answer tends to sound a lot like the one European banks gave before DORA: because nobody drew up the exit plan before it was needed.
Two Answers That Fall Short
The most common answer is commercial: negotiate a longer contract or a cap on price increases. It helps the budget, but it doesn't fix the underlying problem — a vendor can discontinue a model, change its behavior, or scale back what it offers without touching a single line of the pricing contract. According to the same report, Anthropic retired more than one Claude version with little notice over the past year; OpenAI did the same with an image-generation tool, swapping out the model behind it without even naming the previous version.
The second common answer is technical: signing with more than one vendor at once, as insurance. That reduces the risk of a total outage, but it doesn't fix the switching problem — because what actually gets locked to a vendor is rarely the technical connection. It's the workflow someone spent months tuning: the exact way to phrase a request, the exceptions learned through trial and error, the sequence of steps that, in practice, learned to work with that specific model. None of that migrates on its own to the next one.
In other words, both common remedies treat the symptom — price and availability — and leave the cause untouched: the knowledge and rules of how the company works became part of the vendor, instead of staying part of the company.
Once AI is already woven into internal processes and tuned to specific workflows, it carries dependencies and small adjustments nobody documented.
What Has to Be in Place
Company knowledge lives outside the model, with a source and an owner. Policies, processes, and playbooks stay stored in a place the company controls, with a source and a current version — not baked into one vendor's fine-tuning. Switching models doesn't erase what the company knows.
Work rules are documented in the platform, not in the head of whoever uses the model today. Who approves what, in which order, with which exception: that's a company rule, written down and enforced by the environment AI runs in — not a habit that only exists because someone memorized the right way to ask a specific model.
Each task is routed to the most suitable AI model, not to one fixed model. When the platform decides, task by task, which model handles it, swapping one vendor for another — or running both at once, for cost or availability reasons — becomes a configuration change, not a rebuild of the whole operation.
The contract spells out how to leave before the company needs to leave. A transition period, an export format for the data and for what was learned, minimum notice before a price change or a discontinuation: clauses like these cost a negotiation today and avoid an expensive rescue later.
The audit trail and history stay with the company, not only with the vendor. A record of what was approved, by whom, and why, kept in the company's own environment. If the model vendor changes, the history of what the company built doesn't disappear with it.
None of these mechanisms depend on guessing which vendor will raise prices first. This is how Skyller was designed: company knowledge is kept as an asset of its own, with a source and an owner, and each task is routed to the most suitable AI model — switching vendors becomes an adjustment, not a rebuild.
Getting Leverage Back
The first gain is the most direct one: negotiating leverage. A company that can switch models without restarting its operation can also say no to a price increase — and the vendor knows it from the first renewal meeting.
The second gain shows up when a vendor discontinues a model or the service goes down for a few hours. Instead of the operation stalling until someone manually rebuilds everything, the task gets routed to another model already configured for that same type of work, without whoever requested it even noticing the difference.
The third is financial, and the easiest to show on a spreadsheet: not every task needs the most expensive model. A simple support question and a complex contract analysis don't require the same processing effort — and with the right routing, the company pays the price of the task, not a single vendor's flat rate for everything. Skyller, for example, handles that task-by-task routing continuously, without anyone having to pick the model by hand for every request.
The fourth gain is less visible, but it's what's left after years: the knowledge built along the way — the playbooks that work, the exceptions learned, the history of decisions — keeps belonging to the company, and doesn't evaporate when a contract changes.
Five Questions Before Signing the Next Contract
- If this vendor disappeared tomorrow, what would stop working? List the processes that currently depend on a single AI model with no configured alternative.
- Does what the company learned belong to the company, or to the vendor? Playbooks, exceptions, and fine-tuning that only exist inside one specific model aren't a company asset — they're a hostage of it.
- Does the contract state how much transition time the company gets if the price changes or the model is discontinued? If the answer is "we don't know," that's the first clause to renegotiate.
- Is every task going to the right AI model, or does everything run through the same one out of habit? A simple task on an expensive model is wasted money; a complex task on a simple model is a risk.
- Who, today, could rebuild the company's AI operation without one specific person? If the answer is a single name, the knowledge isn't with the company — it's with a person and a vendor.
