How to buy AI development: a founder's guide to evaluating vendors

Ask how they'll know it works

The single most revealing question you can ask an AI vendor: 'How will we measure whether this is working, and can I see the harness?' Teams that ship production AI answer instantly — they'll talk about a golden set built from your data, accuracy thresholds agreed before build, evals wired into deployment. Teams that don't will pivot to talking about the model.

Push one level deeper: ask what happens when the system is wrong. Every AI system is wrong sometimes. A serious vendor has a designed answer — confidence thresholds, human review on consequential outputs, an escalation path. A vendor who says the model is very accurate has not thought about the day it isn't.

The red flags, in rough order of severity

Most bad engagements telegraph themselves in the sales process. The pattern behind all of these is the same: a vendor optimizing for the signature, not the outcome.

• Demo-ware: a polished demo on their data, and resistance to running it on yours before signing
• No talk of evaluation: if measurement doesn't come up until you raise it, quality was never the plan
• Accuracy promises before seeing your data — nobody honest commits to a number blind
• Lock-in by architecture: proprietary platforms you can't leave, or a contract where you don't own the code, prompts, eval sets and fine-tuned weights
• A quote for a platform when you asked for a feature — six months and no user contact until the end
• No questions about your baseline: a vendor who doesn't ask what the workflow costs today can't prove they improved it
• AI-washing: a thin LLM wrapper pitched as proprietary technology, with hand-waving when you ask what's actually theirs

Interrogate the boring parts

The differences between AI vendors mostly live in the unglamorous layers: how data is handled, what happens when the model provider has an outage, how prompt changes are tested before they reach users, what the monthly inference bill will look like at 10x your pilot volume. Ask about all four and listen for specifics.

Data handling deserves particular attention. Where does your data go, which third parties see it, is it used for training, and what happens to it when the engagement ends? A vendor who answers with named subprocessors, retention rules and a deletion process has done this before. A vendor who says everything is encrypted, as if that were the question, has not.

Structure the first engagement to make failure cheap

Don't sign a six-month build with a vendor you haven't worked with. Structure a first phase of four to eight weeks around the riskiest question — usually 'can the model actually do the core task on our data, at acceptable quality?' — with a defined eval set, agreed success numbers, and a deliverable you own either way.

Pay for that phase honestly; a vendor eating discovery costs recovers them somewhere less visible. What you're buying is an answer and an option: if the numbers clear the bar, you continue with evidence instead of hope. If they don't, you've spent weeks, not quarters, and you own the eval set and the findings for whoever you work with next.

Own the assets that outlive the engagement

Whoever builds your AI system, four things should be contractually yours: the source code, the prompts and pipeline configuration, the evaluation sets, and any fine-tuned model weights. The eval set is the one teams forget, and it's arguably the most valuable — it encodes what 'good' means for your business, and it's what makes the next vendor, or your future in-house team, productive on day one.

The test of a healthy engagement is that you could leave. A vendor confident in their work will make you portable and bet on the relationship. A vendor engineering the exit to be painful is telling you how they expect the work to hold up.