A reference call on an AI enablement purchase has one job: find out whether anyone worked differently after the vendor left. Almost every reference script online was written for a software implementation, so it asks about budget, timeline, and integration partners, and a training vendor passes all three while teaching nothing. If you authorized ChatGPT, Claude, Gemini, or Microsoft Copilot seats at a 50-to-500-person company and have now been asked what the seats bought, the two reference calls on your calendar are the cheapest diligence available to you. Below are the nine questions worth the time, the answers that should end the conversation, and how to reach a client the vendor did not hand you. The wider case for buying capability rather than more licenses sits in our argument against license-first rollouts.
Why the Reference List You Were Handed Is the Wrong Input
The names you received are the vendor’s three happiest buyers, and the vendor knows which story each one tells. The error is treating that list as a sample rather than a highlight reel, then asking it generic procurement questions it can pass easily.
Look at the standard script. The widely copied DelCor version asks how you are using the technology, whether the vendor held budget and timeline, how they cooperated with integration partners, and whether you would select them again. A vendor who flew in, ran a polished two-day session, and left nothing transferable answers every one of those cleanly.
The other common script fits no better. The “questions to ask an AI vendor” lists cover data provenance, subprocessors, breach windows, and model versioning: right for software that touches your data, beside the point for two facilitated days with forty people. One idea carries over. A vendor who defines success only in model metrics has not defined business value. The training equivalent is a vendor whose proof is an attendance count.
What makes this purchase different is that the product is a change in human behavior, and behavior is what nobody measures. ATD’s Effective Evaluation: Measuring Learning Programs for Success found that 54% of organizations evaluate some programs at the behavior level, and those organizations do it for only 34% of their programs. So your questions have to work on a reference who ran no study, by asking for recollections instead of findings.
The 9 Questions to Ask an AI Enablement Vendor’s References
Nine, in the order to ask them. The first three carry the weight; if the call gets cut short, ask those.
1. Who stood at the front of the room, and are they still at the firm?
Training is delivered by a person, and a glowing reference is usually about that person. Get the name, check it against your proposal, and ask the vendor whether that facilitator is assigned to your dates and whether they are an employee or a subcontractor. Software does not resign; facilitators do, and generic scripts have no equivalent of this question.
2. What were your people doing differently 90 days later?
Ask for a task and a verb, not a verdict. The answer you want: the proposal team drafts first versions in the tool now and edits rather than writes. The answer that tells you nothing: it went well, people felt energized. BCG’s November 2025 analysis of stalled AI spend describes a progression from foundational learning to applied learning to embedded learning that establishes lasting habits, and reports that only about 5% of companies generate value from AI at scale. This question is the cheapest way to learn which of those three stages a vendor reaches.
3. What did you measure, and could you re-run it without them?
The second part matters more. A reference whose vendor produced a before-and-after reading has told you the vendor measures. One who could re-run that measurement alone next quarter has told you the vendor hands over the instrument, not just the chart. Numbers that live only in a vendor dashboard describe a dependency. How a finance reviewer reads those figures sits in our guide to underwriting the spend.
4. How much of the material was built on your own documents?
Enablement that sticks has people practicing on their own work (their close pack, their proposal template, their ticket queue) rather than on a vendor’s sample dataset. Ask what the exercises used. Working on their own files means the vendor prepares properly. A generic demo means you are buying a webinar with travel costs.
5. What share of the session was teaching, and what share was selling the next phase?
Blunt, and worth asking plainly, because it is a recurring complaint in public reviews of AI upskilling providers: the session that turns into a pitch for the bigger program. References answer this one with feeling. Anything beyond a closing ten minutes is a workshop built as a lead magnet.
6. What did they get wrong, and what happened next?
The one generic question that transfers perfectly. Every engagement has a bad hour: a hostile room, a tool that changed since the run sheet was written, an exercise that fell flat. Listen for whether the vendor noticed and adjusted live, or whether it only surfaced in the survey. A reference who can name nothing is either not being candid or bought a single day.
7. Who owned it inside your company afterward, and did that work?
Programs decay without a named internal owner, and no vendor supplies one. Ask who it landed on, whether they had time, and whether anyone followed up in month three. The honest answer at this company size is usually that it became someone’s second job and quietly stopped, which tells you what your own scope has to force into being. That thread is picked up in the full arc from rollout to measured capability.
8. What did you end up doing yourself that you expected them to do?
This finds the quiet exclusions: chasing attendance, redacting documents, writing the internal announcement, compiling the prompt library the vendor showed but never assembled. None of it is unreasonable for a client to own, and all of it should have been in the scope. A pattern of surprises here is the shape of a cheap quote, covered in the real cost of a low bid.
9. Would you buy the same scope again, or a different one?
The generic version, would you buy again, gets a polite yes. The sharpened version gets the design flaw. References will say they would run half as many people twice rather than everyone once, or insist on role-specific sessions, or buy the follow-up instead of the launch. That is free scope advice from someone who paid for it.
The Answer Key
What to do with each answer.
| Question | Reassuring answer | Answer that should stop the deal |
|---|---|---|
| Who stood at the front | A named person still at the firm and named in your proposal | Nobody remembers, or the name does not appear in your scope |
| Doing differently at 90 days | A specific task, with a verb and a role | Enthusiasm with no task attached |
| What was measured | A before-and-after reading the client can re-run alone | Attendance, satisfaction scores, or a dashboard the vendor holds |
| Own documents | Exercises ran on the client’s real files | A generic demo dataset |
| Teaching versus selling | A closing ten minutes at most | A large block of the session |
| What went wrong | A real incident the vendor caught and fixed live | Nothing went wrong |
| Internal ownership | A named owner with time who ran a follow-up | It landed on nobody and stopped |
| Unexpected client work | Prep tasks that were in the scope anyway | A list of surprises |
| Same scope again | A refinement they would make | A different vendor |
Two or three disqualifying answers in one call is a result. One is a conversation to have with the vendor, and a good vendor will engage with it.
How to Reach a Reference the Vendor Did Not Choose
The curated list is the floor of your diligence, not the ceiling. Two off-list routes need no help from the vendor.
Find a client who did not renew. Case studies and conference pages name companies and often the sponsor. A logo in material from two years ago and nowhere since points to the most informative call available to you. Ask rather than accuse: we are considering the same vendor, what would you want to know if you were us. Most people answer.
Talk to an attendee, not the buyer. Your reference is the person who signed; the person who sat in the room for two days has the more useful view, and the vendor will usually pass on one name from the cohort. A sponsor reports whether the project ran smoothly. An attendee reports whether they learned anything.
What an AI Enablement Reference Call Cannot Tell You
Knowing the boundary keeps you from over-reading a good call.
It cannot tell you the price to pay. References rarely disclose fees, the figures they do give are stale, and scopes are not comparable across companies. Hold your own market anchors instead: hands-on workshop programs land roughly in the $2,000 to $15,000 band, and the custom automation work that sometimes follows runs far higher, $25,000 to $150,000.
It cannot tell you whether the vendor will be good with your people. A reference’s culture, baseline skill level, and manager support are not yours, and manager support most often decides whether anything survives the week.
It cannot replace reading the contract. Ownership of the material your staff produces, the license term on the vendor’s frameworks, and the deliverables list are documents rather than anecdotes, and what should arrive at the end of an engagement is a check run on paper.
Checking References on a Two-Person Vendor
Much of this market is boutiques and solo practitioners, and a three-client vendor has no reference program to run. Adjust rather than disqualify.
Ask for two clients and accept one, provided you also get something you can evaluate directly: a real run sheet with the timings on it, a redacted prompt library, a 30-minute working session with the facilitator. Direct evidence beats references here, because the person you are evaluating is the person who will teach. The sequence that keeps a small-vendor evaluation honest is in our rules for choosing a partner.
Do not concede on the facilitator. With a large vendor you are checking that the named person stays assigned; with a small one, that they exist and are good, which a live session settles faster than any call.
Twenty-Five Minutes on a Thursday
A run sheet for the call you had no time to prepare for.
- Two minutes: confirm their company size, the scope they bought, and the date. More than two years ago and both the tooling and the material have changed.
- Twelve minutes: questions one, two, and three. Do not leave the 90-day question until you have a task and a verb.
- Six minutes: questions four through eight, fast. Each is a yes-or-no with one follow-up.
- Five minutes: question nine, then the oldest closer in reference checking (“is there anything you wish I had asked”), then ask for one attendee’s name.
Log the answers in that vendor’s column while the call is fresh. The pattern across two calls is the finding, not any single quote.
Frequently Asked Questions
What should I ask an AI enablement vendor’s references?
Nine: who facilitated and whether they are still at the firm, what people did differently 90 days later, what was measured and whether the client can re-run it alone, how much of the material used their own documents, how much of the session was selling the next phase, what went wrong, who owned the program afterward, what work the client did themselves, and whether they would buy the same scope again. The first three decide most calls.
How many references should I ask for?
Two from the vendor’s list is enough; a third curated reference adds little. Spend the saved effort on one off-list contact, either a client who did not renew or an attendee rather than the buyer. That call is worth more than the whole curated set.
Can I ask to speak to a client who stopped working with them?
You can ask, and a confident vendor sometimes agrees. If they decline, note it and find the person yourself through case studies, conference listings, or your own network. A lapsed client is rarely a hostile witness; they usually describe a scope mismatch, which is exactly what you are trying not to repeat.
What if the reference cannot tell me whether the training worked?
Expect it, and switch from findings to recollections. Ask what one person does differently now, what their team stopped doing by hand, and whether anything in the weekly routine changed. Someone who ran no study can still describe behavior, and no change worth describing is itself a clear answer.
Is a logo wall the same as a reference?
No. A logo means an invoice was paid, and in corporate AI enablement it often means one pilot session for one department years ago. Treat the wall as a list of names to call rather than evidence, and ask when each engagement ran.
What questions are unfair to ask a reference?
Fees and contract terms, which put the reference in an awkward spot and produce unreliable numbers anyway. Named people who performed poorly. Anything asking them to speculate about your company. Everything else above is fair, and a reference annoyed by a direct question about measurement is rare.
Should I check references before or after the proposal?
After the proposal, before the final negotiation. You need the scope in hand so your questions can test it, and room to negotiate on whatever the calls turn up. Reference-checking a vendor you have already verbally committed to is theater.
How do I check references on a vendor with only three clients?
Accept one reference and ask for direct evidence: a real run sheet with timings, a redacted prompt library, and a short working session with the facilitator who would teach your team. With a small vendor the facilitator is the product, so evaluate the facilitator.
What is the single most useful question on the list?
What were your people doing differently 90 days later. It is the only question a well-run session that taught nothing cannot answer, and the one most vendors have never been asked.
Where to Start
Put the nine questions across the top of a sheet, book the two calls, then spend twenty minutes finding one name the vendor did not give you. That call reorders a shortlist more often than the proposals do.
For a second read on the answers you collect, book a free AI-readiness call. We will say which responses are genuinely disqualifying, which are normal for this market, and whether your team needs a paid program at all right now.
Arthur Wandzel
