The demo was on Tuesday, and it went well. The product summarized a stack of documents in seconds, the account executive was funny, and the slide at the end had logos you recognized. On Wednesday morning the proposal arrived as a PDF, your boss forwarded it to you, and the whole message was one word: "Thoughts?"
That one word is how most AI purchases actually get decided. Somebody who liked the demo writes back "looks promising," and eighteen months later the same somebody is explaining the renewal.
Here is my view, and I hold it firmly. Most vendor evaluations are decided in the demo, and the criteria get written afterwards to explain a feeling people already had. I'd write the criteria first, before the next call, and let the vendor do the work of answering them. You do not need to understand how the model works to do this well. You need seven questions and the patience to wait for real answers.
If you want the tactics for the meeting itself, I've written about reading the pitch and about the one request that breaks a polished demo. This piece is the list you check everything against.
1. Business fit
The plain-English version: does this solve the specific piece of work you named, or a nearby problem the vendor is better at?
The question I'd ask: "Which of our workflows does this replace or speed up, and what does that workflow look like the week after we switch it on?"
A good answer names your workflow back to you and describes a Tuesday afternoon after launch. A weak answer talks about "transformation" and "use cases across the enterprise." When a vendor can't describe your ordinary week, they are selling you their product's best week instead.
2. Total cost
The plain-English version: what will this cost in year one, including the parts that are not on the pricing page?
The question I'd ask: "Can you put in writing what we'd pay at our current volume, and at double it?"
A good answer is a number with assumptions attached. A weak answer is "it depends on usage" followed by an offer to "model it together later." Usage-based pricing is fine. Pricing nobody will write down is a cost you will discover in the second quarter.
3. Data handling
The plain-English version: where does your data go, who can see it, and does the vendor use it to improve their product?
The question I'd ask: "Is our data used to train your models, and can you show me the contract language that says so?"
A good answer points to a paragraph in a document. A weak answer is a reassuring sentence from the salesperson. I'd also loop in whoever owns security at your company before you get attached, because they will ask this question eventually and it is better if they ask it now.
4. Output quality on your work
The plain-English version: is the output good enough on your own messy examples, rather than on the vendor's prepared ones?
The question I'd ask: "Can we send you five real examples this week and see the outputs before we go further?"
A good answer is "yes, send them." A weak answer is a case study from a company that looks nothing like yours. Your own examples will tell you more in an afternoon than the demo did in an hour.
5. Implementation reality
The plain-English version: who does the work to get this running, and how long does it take for a team like yours?
The question I'd ask: "For a team with no AI engineers, who does the setup, and what do you need from us in the first month?"
A good answer names people, hours and a date. A weak answer mentions a "professional services engagement" that was not in the proposal. That line is where the real price often lives.
6. Exit and lock-in
The plain-English version: if this doesn't work out, how hard is it to leave?
The question I'd ask: "If we cancel after a year, what do we get back, in what format, and what do we have to rebuild?"
A good answer is boring: an export, a standard format, a clean termination clause. A weak answer changes the subject to how happy their customers are. I'd ask this one while everyone is still friendly, because nobody enjoys asking it later.
7. Vendor stability
The plain-English version: will this company and this product still look the same by the time your contract renews?
The question I'd ask: "What's shipping in the next two quarters, and what happens to our price if you're acquired?"
A good answer separates what exists today from what is planned. A weak answer blends them together until you can't tell which features you are actually buying. I'd buy what works now and treat the roadmap as a nice surprise if it arrives.
The note to send back
Once you have the answers, reply to "Thoughts?" with seven lines. Give each criterion one sentence and one mark: green if the answer was clear and in writing, amber if it was clear but only spoken, red if you didn't get one.
Here's where we are on the proposal, against seven criteria:
1. Business fit: green. It replaces our intake triage and they described the week after launch in detail.
2. Total cost: amber. Year-one number is clear; the "double volume" number is promised by Friday.
3. Data handling: green. Contract says our data is not used for training; security has the paragraph.
4. Output quality: amber. Their examples look good; ours come back Thursday.
5. Implementation: red. Setup assumes an engineer we don't have, so I've asked who does it.
6. Exit: green. Standard export, 60-day termination.
7. Vendor stability: amber. Roadmap is vague past one quarter.
My recommendation: continue, conditional on 4 and 5 by next Friday.
Those example answers are placeholders, so write your own. The shape is the point. Your boss can forward it upward without editing, and nobody can accuse you of either blocking the tool or waving it through.
What this gets you
Most people answer "Thoughts?" with a feeling, because a feeling is all the demo gave them. You'll answer with seven lines that say what you checked, what you're still waiting for, and what you'd do next. That reads as judgment, and judgment is what gets you handed the next decision.
Write the seven questions into your notes before the next vendor call, send them over, and let the vendor spend the week answering. You have better things to do on a Wednesday than evaluate a mood.
P.S.
If you're comparing more than one vendor, or this one is a big contract, The AI Vendor Evaluation Kit gives you a sixteen-question bank for the call, a weighted scorecard, and a one-page recommendation leadership can act on. ($29. The seven questions above work on their own.)
This was a free one. There's a new one every week.
What's signal in the AI noise — and the move to make about it. No hype, no vendor pitch, no link dump.
Aiden Vector is an AI-assisted publication; this content is produced by AI under human editorial direction.