Answer

How do you tell whether an AI vendor is legitimate?

Justas Butkus is a fractional AI officer based in Vilnius, Lithuania, working with mid-market companies and scale-ups across the UK, EU and US at two to four days a month. He builds and operates the production AI systems he advises on, and takes no commission, referral fee or revenue share from any vendor.

The test buyers themselves use

Think back to the last sales call where the vendor started sharing their screen in the first five minutes. You had not finished explaining what was going wrong in your week, and there was already a product on the slide, with a name, and a logo, and a very confident arrow. Something in you went quiet and said: they have not asked me anything yet.

That instinct has a name, and it is not my framework. It is how an experienced buyer described their own decision process, unprompted, in a public forum:

Both halves carry weight. A vendor who will not scope before a contract is protecting themselves from discovering the work is small, and a vendor who names a product first has a solution looking for your problem.

Ten questions, in order

Ask these in a first conversation, and in this order, because the early ones are hardest to fake and the later ones only make sense once the early ones are answered. You do not need to ask them like an interrogator. A curious person asking slowly will hear the same things.

  1. What is breaking in our day, as you understand it?If they cannot restate your problem before pitching, they were not listening. This should be question one and it disqualifies quickly.
  2. Will you produce a short scoping document before we commit to a build?The buyers' own test. A no here is close to disqualifying on its own.
  3. What have you personally put into production, and can we look at it?Not a prototype, not a demo environment. Something running that you can inspect.
  4. What would you tell us not to do?Anyone who cannot name work they would decline is selling capacity rather than judgement.
  5. What happens when the system produces a wrong answer?People who have operated a real system answer instantly and specifically. People who have not talk about accuracy percentages.
  6. Who owns the accounts, the infrastructure and the prompts?If the answer is them, you are renting, and the switching cost is the product.
  7. What does this cost to keep running, not to build?Most proposals omit this entirely. It is often the larger number over three years.
  8. Do you take commission, referral fees or revenue share from any vendor you might recommend?Ask it directly. Ask every advisor you are considering.
  9. What happens if we stop working with you?Documentation, handover and whether a competent internal engineer could pick it up.
  10. What would make you say this project should not go ahead?A vendor with no such answer will build whatever you fund, including the wrong thing.

Red flags and green flags

Side by side, the two kinds of conversation sound very different from the first minute. The left column is what you hear when the vendor is selling a tool, the right when they are working on your problem.

Signals in an AI vendor conversation
Red flagGreen flag
Names a product before asking about your processRestates your problem before proposing anything
Will not scope without a signed contractProduces a short written scope up front
Confident return figures before seeing your dataDeclines to forecast returns until data is examined
Case studies with no numbers, or logos with no caseSomething running that you can inspect yourself
Accounts and infrastructure in their nameEverything in your name from day one
Build cost quoted, running cost unmentionedThree-year running cost stated unprompted
Every idea you raise is a great ideaTells you which of your ideas is not worth doing
Recommends a product they have a relationship withDiscloses relationships without being asked

The seventh row is my favourite, because it costs a vendor nothing to say and almost none of them do. If every idea you bring is a great idea, you are talking to a salesperson, however friendly.

What "a wrapper with a logo" actually means

Buyers use that phrase for the most common disappointment in this market, and it is worth understanding precisely, because it is not always fraud and sometimes it is perfectly fine.

A thin wrapper is a general-purpose assistant with a system prompt naming your company, presented as a bespoke system. Sometimes that is genuinely all a problem requires, and paying a modest amount for someone to configure it well is reasonable. It becomes a problem when it is priced and described as custom engineering, and the difference shows up as soon as you ask three plain questions.

You, about change

What happens to the system when our process changes?

A real system has somewhere that knowledge lives and is updated. A wrapper requires rewriting the prompt.

You, about evaluation

How do you know it still works next month?

Real systems have an evaluation harness. Wrappers have someone trying it occasionally.

You, about your data

Does it actually use our specific data?

Or does it only sound like your company? The honest answer tells you which one you are buying.

The question to ask every advisor, including me

Anyone advising on AI who also sells AI products has a conflict. It is not disqualifying, and builders often give better advice than pure advisors, but it has to be handled out loud rather than left unmentioned.

Ask: what do you not procure from yourself, what do you disclose, and at what point do you step out of a decision? If the answer is that there is no conflict, that is either an evasion or it has not been thought about.

So the next time a vendor opens with a slide, what is the first question you will ask, and did they have an answer before you finished asking it?

Short answer

Ask them to produce a short scoping document, before any build commitment, describing what they would automate and what it would save. Buyers themselves name this as the test that works. The second signal is whether they name a product before they have asked what is actually breaking in your day.

If you have a proposal in front of you

The ten questions above are the whole method. A second opinion on a specific proposal is a short conversation.

Frequently asked questions

How do I know if an AI vendor is legitimate?

Ask them to produce a short scoping document describing what they would automate and what it would save, before any build commitment. Buyers themselves name this as the test that works. Also watch whether they name a specific product before asking what is breaking in your operation.

What are the biggest red flags with AI vendors?

Naming a product before understanding your process, refusing to scope without a signed contract, confident return figures before seeing your data, holding accounts and infrastructure in their name, and quoting a build cost without a running cost.

What is an AI wrapper?

A general-purpose assistant configured with a system prompt naming your company, presented as bespoke engineering. Sometimes that is genuinely sufficient and worth a modest fee. It is a problem when priced and described as custom development.

Should I ask an AI vendor about commissions?

Yes, directly, and ask every advisor you are considering. Whether they take commission, referral fees or revenue share from products they might recommend changes what their recommendation is worth.

What if the vendor will not scope before a contract?

Treat it as close to disqualifying. Refusing to scope protects the vendor from discovering the work is smaller than the proposal, which is precisely the information you need before committing.

How do I check what they have actually built?

Ask for something running that you can inspect or use yourself, rather than a demo environment or a client under confidentiality. Anything you cannot verify independently is a claim, not evidence.