What are the biggest red flags when hiring an AI vendor?

The reliable ones: a price quoted before anyone has seen your workflow, production experience proven only by demos, evals framed as an optional extra, hosting that must stay in the vendor’s accounts, AI recommended for every problem you describe, and a strategy phase sold before any contact with the actual work.

Why is a price before seeing the workflow a red flag?

Because scope in AI work is a function of the workflow’s real volume, exceptions, and integrations, none of which a vendor has seen yet. A number produced without looking is a guess dressed as a quote, and it predicts either padding or change-orders.

Is vendor-hosted AI always a red flag?

Hosting during the build is normal. Hosting that must remain with the vendor after handover is the flag: it converts your system into their recurring revenue and your switching cost. The test is whether the engagement can end with everything running in your accounts.

AI vendor red flags: what you will hear, and what it means

The short answer

The reliable AI vendor red flags are phrases that sound reassuring and predict trouble: a price before seeing the workflow, demos as production proof, evals as an add-on, mandatory vendor hosting, AI recommended for everything, strategy phases before contact with the work, no deprecation plan, vague answers about who does the work, and engagements with no defined ending. Each is the vendor describing their business model out loud.

Red flags in this market rarely look like warnings. They look like confidence, convenience, and speed. Here are the nine we would check for if we were buying, written as what you will actually hear in the call, with the translation attached.

The nine flags

–"Ballpark, this runs about X." A price before anyone has seen your workflow is a guess with a signature line. Translation: scope will be discovered later, at your expense, via padding or change-orders.
–"Let me show you a quick demo." As the answer to "show me production." Demos are the easiest artifact in AI to manufacture; a vendor whose proof is a demo has a pipeline of demos.
–"We can add evals if you want them." Evals are how a system proves it works; making them optional relocates quality assurance from their build to your production.
–"It runs on our platform, so you get updates automatically." Convenience framing for permanent hosting. Translation: your workflow becomes their recurring revenue, and leaving becomes a migration project you will price annually and never fund.
–"AI can definitely help with all of that." After you described six problems. Some of what you described is a cron job, and a vendor who cannot say so is selling their revenue model, not your outcome.
–"We start with a strategy phase." Billed monthly, conducted at a distance from the actual queue. Briefs written far from the work are how systems end up automating an imagined workflow, which is the leading documented failure mode.
–"The model? We always use the latest." As the answer to what happens when a version retires. Providers publish deprecation schedules; systems without a plan for them fail on someone else's calendar.
–"Our senior team oversees everything." Oversees is the load-bearing word. Ask who writes the code and sits in your workflow; if the names change after signature, so does the quality.
–"We see this as a long-term partnership." Before anything has shipped. Partnerships are earned by delivery; proposed up front, the phrase usually means an engagement designed without an ending. The antidote is a handover with a date on it, backed by ownership terms the law will not supply by default.

Red flags in this market look like confidence, convenience, and speed.

The pattern underneath

Every flag above is one incentive wearing nine outfits: revenue that continues whether or not your system works. Guessed prices convert to change-orders, demos convert to discovery phases, optional evals convert to support tickets, mandatory hosting converts to rent, and endless partnership converts to all of the above indefinitely. The inverse incentive exists and is checkable: fixed scope, production as the deliverable, ownership transferred at the end, which makes the vendor's payday depend on finishing.

Running the counter-moves

You do not need to argue with a flag; you need to test it. Each of the nine collapses under one of the fifteen questions: production proof against the demo, the eval question against the add-on framing, the handover walk-through against the hosting pitch, the refusal question against AI-for-everything. Ask them of every vendor symmetrically, including the one whose site you are reading, and let the pattern of answers decide. The full selection framework, engagement models included, is the field guide.