Switching engineering partners is expensive, and most teams put it off six months too long. With AI projects, that delay is more costly: model choices age out, eval debt compounds, and the team that built the system loses interest in the team that needs to inherit it. Here are the signals worth acting on, and the playbook for switching cleanly.
Five signals it's time to switch
- You can't see an eval dashboard, and your vendor can't produce one quickly
- Cost per query has crept up and nobody can explain why
- Every change request becomes a 'big build', small iteration is gone
- Your vendor still recommends fine-tuning for problems frontier models now solve out of the box
- The team that built the product is no longer the team working on it
Why AI vendor lock-in is different
In a traditional codebase, lock-in is the code itself. In an AI product, lock-in is the eval dataset, the prompt history, and the institutional knowledge of which model versions behave well. A vendor that hasn't invested in those artefacts has, intentionally or not, built a product the client cannot move.
If your vendor can't hand you a runnable eval suite this week, you don't own your AI product. They do.
The transition playbook
We've inherited enough handovers to know what works. Phase the switch: a 4-week parallel discovery, an 8-week stabilisation phase where the new team owns evals while the old team owns deploys, then a clean cut. Skipping the parallel phase is the most common mistake, it's where the new team builds confidence in the system.
What you need to ask of the new partner
- Show me the eval harness you'd build for this product on day one
- Show me a prompt diff and explain how you'd review it in PR
- Tell me which of my fine-tunes you'd retire and which you'd keep, and why
- Walk me through your handover plan for the moment I leave you
- Show me a customer you handed over to a third party, and how that went
What a clean transition looks like
A clean transition is one where the eval pass-rate doesn't dip, the cost-per-query doesn't spike, and the team you replace remains professional throughout. We've been on both sides of this, and we know the second one is the harder test of a vendor's character.