Build-versus-buy debates about AI tend to go in circles because both sides are arguing from instinct. Engineers want to build. Executives have seen a vendor demo. Here are the questions that actually resolve it.
1. Is the feature the product, or is it plumbing?
If the AI feature is what customers will choose you for, you probably need to own it. If it's a capability every competitor will have in a year, buy it and spend your engineering on what makes you different.
2. Does the vendor's model see your data?
Read the data terms, not the marketing page. Where does your data go, is it retained, is it used for training, and can your security team live with the answer? A "no" here ends the conversation regardless of price.
3. Can you evaluate it?
If you can't run the vendor's feature against your own test set and measure the results, you're buying on faith. Ask for a trial that lets you do this. If they refuse, that's information.
4. What does it cost at your volume?
Per-seat pricing and per-request pricing produce wildly different numbers at scale. Model it at three times your expected volume, because AI features tend to get used more than planned once they work.
5. How locked in are you?
If the vendor's feature is deeply wired into your product and their pricing doubles, what's your move? Prefer vendors whose capability you could replace with your own build in a quarter if you had to.
6. Do you have the team to own a build?
Building means evals, monitoring, cost management, and model upgrades forever. If nobody's going to own that, buy. If someone will, building is cheaper than it looks.
Most teams should buy the plumbing and build the product. The checklist is for figuring out which is which.