FIELD NOTES — №04·10 July 2026·Silvester Vrbancic

Red flags when hiring an AI consultant.

Ask what a prospective AI consultant measured before they proposed anything, and what happens if the honest answer turns out to be "don't build this." Consultants who can't answer either question in plain terms are optimising for the sale, not for your business, and the rest of the conversation tends to confirm it.

§ 01 — What are the biggest red flags in an AI consulting sales conversation?

A polished demo with no baseline measurement. It looks like: an impressive live walkthrough of what the system will do, with no mention of the current process it's replacing or how much time or money that process costs today. The exposing question: "what's the current baseline, and how was it measured?" If there isn't a straight answer, there's no way to know afterward whether anything actually improved.

"Transformation program" pricing before the problem is understood. It looks like: a proposal for a multi-month, multi-stage engagement priced before anyone has watched the actual work happen. The exposing question: "what specific process does phase one measure, and what does it cost if I stop after that phase?"

Proprietary middleware and lock-in. It looks like: the system only works through the consultant's own platform, and leaving means rebuilding from scratch. The exposing question: "if I never speak to you again after this ships, what do I own, and can someone else maintain it?"

Retainers for systems that should run themselves. It looks like: an ongoing monthly fee for a tool that, once built and handed over, has no obvious reason to need continued paid attention. The exposing question: "what specifically does this retainer pay for, month to month?" A vague answer is the flag, not the retainer itself.

Jargon density as a smokescreen. It looks like: a conversation thick with terms like "agentic workflows" and "AI-native transformation" that, when you ask for a plain-English translation, turns out to describe something fairly ordinary. The exposing question: "can you explain that in one sentence, without the term?" If they can't, they may not fully understand it either.

Refusal to put predictions in writing. It looks like: confident verbal claims about time or cost savings that don't make it into the proposal document. The exposing question: "can you put that number in the contract as the target we're measuring against?"

No willingness to say AI is the wrong tool here. It looks like: every problem you describe gets an AI-shaped answer, regardless of fit. The exposing question: "what's an example of a project you've talked a client out of?" Someone who's never turned down work, or can't name one, hasn't been selective.

“Ask what they measured before they proposed anything. If the answer is a demo instead of a number, you're being sold a feeling, not a result.”— The one question that filters most of the field, § 01

§ 02 — Are there green flags too?

Yes, and they're mostly the mirror image of the above. A consultant who wants to watch the actual work before proposing anything. One who quotes a fixed fee for a defined, measurable piece of work rather than an open-ended engagement. One who hands over documentation and access you'd need to run the system without them. One who talks about a specific number from a past project instead of a general sense of experience. And one who's willing to tell you, in writing, that a given idea isn't worth building, even when saying so costs them the sale.

None of these are exotic. They're mostly just what an honest working relationship looks like, stated plainly enough to check.

§ 03 — When is an ongoing retainer actually legitimate?

Some AI work genuinely needs continued involvement. A system that depends on a third-party model provider may need updates when that provider changes its API or pricing. A tool monitoring live data for anomalies benefits from someone tuning its thresholds as conditions shift. A process that's still evolving month to month, rather than settled, may need ongoing adjustment rather than a one-off build.

The distinction isn't retainer versus no retainer. It's explained versus unexplained. A legitimate retainer has a stated reason tied to something that will keep changing. An illegitimate one is billed indefinitely for a system that, by any reasonable description, should now run on its own. If nobody can tell you specifically what the retainer is buying this month, that's the answer.

Compare any proposal you're looking at against the do/don't list here, and if something doesn't add up, send the brief and I'll give you an honest second opinion.

§ 04 — Questions people ask

Is a fixed price always better than hourly billing?

Not automatically, but a consultant who can't quote a fixed price for a defined piece of work usually hasn't measured the problem well enough to scope it, which is itself worth noticing.

Are all retainers a bad sign?

No. Ongoing monitoring, model updates, or genuinely evolving scope can justify one. The red flag is a retainer with no stated reason it needs to continue, for a system that should run on its own once built.

What if a consultant just seems confident and experienced, is that enough?

Confidence isn't evidence. Ask for a specific number from a past project, hours saved, cost recovered, error rate change, and how it was measured. A confident non-answer is still a non-answer.

Should I be worried if a consultant recommends against automating anything?

The opposite. A consultant willing to say 'this isn't worth it' in writing, on a project they'd otherwise be paid to build, is one of the strongest positive signals available.

If you've got a suspect inefficiency, send the brief. I'll tell you plainly whether it's worth fixing.

Send the brief →