Here's a number that explains why your inbox is full of AI pitches: 43% of small businesses now use AI, according to an October survey of 691 owners. When nearly half the market is buying, everybody starts selling. Some of those sellers are real implementers who've wired up actual businesses. Some are slide decks with a logo.
Telling them apart isn't about their website or their client logos. It's about the questions you ask in the first meeting and what their answers sound like. Nine questions, in the order that matters.
1. Who does the work?
Not the firm. The person. Get a name, a background, and how many other projects that person is carrying right now. Plenty of firms send a senior partner to the sales meeting and a junior to your project. That's not automatically bad, but you should know it's happening before you sign, not after the third missed deadline.
Good answer: a named person, their relevant experience, and an honest statement of their current load. Dodge: "we have a deep bench" with no names attached.
2. What do I own when you leave?
This is the question that separates an implementer from a landlord. When the engagement ends, do you own the workflows, the accounts, the logins, and the documentation? Or does everything live in the consultant's accounts, on the consultant's subscriptions, behind the consultant's passwords?
Good answer: everything runs in accounts you own, from day one, with documentation your team can follow. Dodge: vague talk about "our platform" or "our proprietary system" with no clear handoff.
3. What happens when it breaks?
Every automation breaks eventually. An API changes, a vendor updates a pricing page, a webhook goes quiet. The question isn't whether it breaks. It's who notices, how fast, and who's on the hook to fix it.
Good answer: monitoring that alerts someone (them, during the engagement, and you after), a named response time, and a plain-English explanation of the most common failure modes for your specific setup. The second-year costs nobody quotes live right here. Dodge: "it's very reliable" with no monitoring described.
4. Show me something running.
Not a slide. Not a mockup. A real system, doing real work, that you can watch or click through. It doesn't have to be in your industry. It has to be alive.
Good answer: a screen share of a live workflow, with honest commentary about what broke during the build. Dodge: a case study PDF with no live artifact behind it. One practitioner put it well this month: big claims with nothing you can inspect are the same red flag as a vague answer about a production incident.
5. What are the kill criteria?
Before you start, agree on what "this isn't working" looks like and what happens then. A pilot with no kill criteria is a project with no end. You want a consultant who'll tell you at week three that the workflow you picked was the wrong one, not one who'll bill through month six hoping it turns around.
Good answer: specific metrics, a check-in date, and a pre-agreed decision. Dodge: "we'll evaluate as we go" with nothing written down.
6. How is my data handled?
Your customer list, your pricing, your inbox. Where does it go when the AI touches it? Who else's models train on it? What happens to it when the engagement ends? The federal AI Risk Management Framework from NIST exists precisely because these questions matter, and any consultant worth hiring can answer them without reaching for a lawyer.
Good answer: data stays in systems you control, vendors named with their data-use terms, deletion on exit. Dodge: "we take security very seriously" with no specifics.
7. What does month two look like?
Month one is the build. Month two is real life. Who trains your team? Who answers when something looks wrong? What does ongoing support cost, and what happens if you want to end it?
Good answer: a training plan with names and dates, a support arrangement with a price, and an exit path that doesn't punish you. Dodge: the proposal ends at go-live and everything after is "to be discussed."
8. Who trains my team?
If your people can't run it, you didn't buy a system. You bought a dependency. The engagement should leave your team more capable than it found them, even if that's just one office manager who knows where the bodies are buried.
Good answer: named training sessions, written runbooks, and a handoff call where your team drives. Dodge: "it's very intuitive" as a training plan. If they can't describe how they check their own AI-assisted work before it reaches you, assume they don't.
9. What's the total first-year cost?
Not the project price. The total: build plus subscriptions plus per-message or per-execution usage plus maintenance plus the support arrangement from question seven. Anyone who quotes you a single number before understanding your workflows is guessing, and the guess is always low.
Good answer: a line-item first-year total with the assumptions stated, so you can check the math. Dodge: a low build price with the ongoing costs discovered later. We've written the full breakdown of what the bands actually look like if you want to sanity-check a quote.
The red flags that end the conversation
A few signals should stop the process no matter how good the pitch sounds. One tool recommended for every problem before they've understood yours. No mention of cost controls or failure handling, which are the hard parts of production AI. Reluctance to transfer knowledge, because an advisor who wants you permanently dependent is optimizing for their revenue, not your outcome. And the inability to describe work they'd turn down. Ask what would make them decline your project. A firm that can't name anything it would refuse has no boundaries, and a proposal with no boundaries commits them to nothing.
The one-page scorecard
Put two vendors side by side on these nine questions and the decision usually makes itself. The right consultant feels less like a vendor closing a deal and more like a candid partner who'd rather lose the engagement than sell you the wrong one. If you're comparing models too, the consultant vs. agency vs. DIY breakdown will help you make sure you're even shopping in the right lane, and the build vs. buy scorecard will tell you whether you need custom work at all.
Book a free 30-minute diagnostic at krastor.com/diagnostic. Ask us all nine questions. We keep score on ourselves the same way.
Frequently asked questions
What should I ask an AI consultant before hiring?
Who does the work, what do you own when they leave, what happens when it breaks, can they show you something running, what are the kill criteria, how is your data handled, what does month two look like, who trains your team, and what is the total first-year cost. Nine questions, and the vague answers tell you everything.
How much should an AI consultant cost a small business?
Anyone who quotes you a number before understanding your workflows is guessing. What you should demand instead is a phase-one price in writing, a total first-year cost including maintenance and usage fees, and a clear statement of what happens financially if you walk away.
What are red flags when hiring AI help?
All confidence and no evidence, one tool recommended for every problem, no mention of cost controls or failure handling, reluctance to hand over knowledge, and an inability to describe work they would turn down. Any one of these is a reason to keep looking.
Should I hire a consultant, an agency, or do it myself?
One broken, expensive workflow calls for a consultant. An empty pipeline calls for an agency. A simple, low-risk task is fine to DIY. The three models sell different things, so match the model to the problem before you compare prices.
Continue reading
- What AI Automation Actually Costs a Small Business (Real Numbers, No Pitch)
- AI Consultant vs. Agency vs. DIY: What Each One Is Actually For
- Build vs. Buy: The AI Automation Decision for a 20-Person Company
- Book your free 30-minute diagnostic
Common questions
- What should I ask an AI consultant before hiring?
- Who does the work, what do you own when they leave, what happens when it breaks, can they show you something running, what are the kill criteria, how is your data handled, what does month two look like, who trains your team, and what is the total first-year cost. Nine questions, and the vague answers tell you everything.
- How much should an AI consultant cost a small business?
- Anyone who quotes you a number before understanding your workflows is guessing. What you should demand instead is a phase-one price in writing, a total first-year cost including maintenance and usage fees, and a clear statement of what happens financially if you walk away.
- What are red flags when hiring AI help?
- All confidence and no evidence, one tool recommended for every problem, no mention of cost controls or failure handling, reluctance to hand over knowledge, and an inability to describe work they would turn down. Any one of these is a reason to keep looking.
- Should I hire a consultant, an agency, or do it myself?
- One broken, expensive workflow calls for a consultant. An empty pipeline calls for an agency. A simple, low-risk task is fine to DIY. The three models sell different things, so match the model to the problem before you compare prices.