How do you avoid getting locked into an AI vendor you can't leave?

Before you sign with an AI vendor, ask one question you won't find in their demo: what does it cost, in money and in months, to leave? Not because you're planning to leave — because a vendor who can't answer that question in a straight sentence hasn't thought about it either, and that's the same team that built whatever you're about to depend on.

That question has a name now — vendor lock-in — and this month it's showing up in a wave of independent pieces from very different corners: a Forbes column on AI contracts nobody reads before signing, a small-business-focused breakdown of what a bad first vendor actually costs, an enterprise research note on switching costs, a practical guide on how to exit one cleanly. Different audiences, same discovery: most companies signed the AI tool before they worked out how to unsign it.

What lock-in actually costs, in numbers people have started measuring

Small businesses waste $15K-$25K on average on their first AI vendor relationship that doesn't work out, according to a small-business-focused survey — not on the tool itself, but on the switching, retraining, and rebuilt integrations that follow a bad fit. In the same data: 43% said unclear pricing led to real cost overruns, 35% found the tool didn't do in production what it did in the demo, and 38% — over a third — couldn't export their own data when they tried to leave.

That last number is the one worth sitting with. A tool you can't get your data out of isn't a vendor relationship anymore; it's a one-way door. One example from that research: a marketing agency signed a two-year contract at a 40% discount, found after four months the tool couldn't handle their clients' industry requirements, and was stuck paying for 20 more months of something that didn't work — because the discount was priced into a term length nobody negotiated down.

Enterprise research on the same question makes a structural point: the true cost of lock-in is "the cost of undoing what you've already built," and it's a number that, per that research, doesn't appear in a single enterprise AI budget document anywhere. Not because it's small — because nobody measured it going in.

Five ways a vendor can lock you in — most of them silent

A useful framework from this week's coverage breaks lock-in into five distinct types, and it's worth checking a vendor against all five, not just the contract clause:

1. Model lock-in — your outputs depend on a specific model's behavior, and the vendor can swap models under you without a contract trigger. Your workflow was tuned to a model that no longer exists.
2. Data lock-in — your documents, corrections, and usage history live inside the platform with no documented export path.
3. Workflow lock-in — your team has built muscle memory around this tool's prompts, templates, and interface, and switching means retraining, not just re-plugging.
4. Integration lock-in — the tool is wired into your CRM, ticketing system, or ERP, and unwinding those connections is its own project.
5. Commercial lock-in — multi-year terms, prepaid credits, bundled discounts, and auto-renewal windows that quietly reset the clock.

Most vendor evaluations check price and features. Almost none check all five of these before signing — which is exactly why they surface as problems only once leaving is already expensive.

The contract questions that keep the door open

None of this requires distrust of every AI vendor — it requires asking the same due-diligence questions you'd ask before signing any long-term infrastructure commitment:

1. What's the named export format, and can we pull our own data without a support ticket? "Available on request" is not an export path.
2. How much notice do we get before a material model change, and can we pin a version if we need to? A vendor with nothing to hide will put a number of days on this in writing.
3. Is there a cap on renewal price increases? Auto-renewal without a stated ceiling is how a reasonable year-one price becomes an unreasonable year-three one.
4. What happens to the prompts, templates, and configurations we built — do we own them? If the vendor owns your customizations, you're rebuilding from zero on your way out, not migrating.
5. What does an exit actually look like — timeline, named contact, engineering hours? A vendor who can answer this in one sentence has done the exit planning already; one who can't hasn't thought about it, and you're the one who'll pay for that gap.

A practical operating habit worth borrowing from IT teams who've done this before: map every internal process that touches the vendor before you give notice, including tools individual teams adopted on their own — that's usually where the real disruption hides, not in the systems everyone already knew about. And start that mapping well before a renewal deadline, not the week of it.

How we build it, so leaving is never the expensive part

We build inside your own cloud account, on open-source components — N8N, PostgreSQL — with no per-seat licenses and no proprietary runtime standing between you and the system. Every build ships with a documented handover any competent engineer could pick up, not a black box only we can touch.

The documentation is the same policy, extended: the process map, the architecture notes, the runbook — every document we write belongs to you. Hand them to any vendor, including one that isn't us. That's not a clause we negotiated in; it's the default, because a system you can't hand off isn't one you actually own.

If you're about to sign with an AI vendor — us or anyone else — run their contract against the five questions above first. If you want a second set of eyes on what you're already running and what it would actually cost to change any of it, the free AI Audit starts with a process map and an honest read on what's automated today — no clock starts on you for asking. Prefer to talk it through? Book a call.