You own a three-dentist practice, and you're on a video call with a consultant. For twenty-five minutes you watch a polished demo: an AI voice books a cleaning, fields an insurance question and sends a confirmation text. It's genuinely impressive. When the call ends you realize the consultant never asked how many calls you get, where your front desk struggles, or which practice-management system holds the schedule.
A demo shows what a tool can do in ideal conditions. It tells you very little about whether the person showing it will understand your business, protect your patients' data or still pick up the phone when something breaks. These twelve questions do.
Questions about the plan
| Question | A good answer sounds like | A red flag sounds like |
|---|---|---|
| 1. What will you learn about our business before recommending anything? | “We'll follow real calls or jobs with your staff, time each step and check what your current software already does.” | “We already know what dental practices need.” |
| 2. Do you earn referral fees, commissions or reseller margins on any software you might recommend? | “No.” Or: “Yes, from these vendors, and we'll disclose it in writing beside any recommendation.” | Hesitation, or “that doesn't affect our advice” with no details |
| 3. Who owns the plan, the code and the configuration? | “You do. It's in the contract, and you can take it to anyone.” | “It runs on our proprietary system,” with no export terms |
Question 2 is the quickest filter. An adviser paid by a software vendor will struggle to recommend switching on a feature you already pay for, or doing nothing, even when that's the right answer. Disclosure doesn't make a consultant bad, but you should know who else is paying them before you weigh a recommendation.
Questions about your data
| Question | A good answer sounds like | A red flag sounds like |
|---|---|---|
| 4. Where will our data be stored and processed, and by whom? | Named vendors, US hosting where it matters, and a written list you can keep | “It's all secure in the cloud.” |
| 5. Will any vendor use our data to train its models? | “No. It's in each vendor's terms, and we'll show you where.” | “It's anonymized, so it doesn't matter.” |
| 6. Who will actually do the work, and which agreements will they and your vendors sign? | Named people, any subcontractors disclosed, a confidentiality agreement up front, and a Business Associate Agreement from everyone who touches patient data | Surprise at the question, “we're HIPAA compliant” with no agreement offered, or subcontractors you only hear about later |
For a dental or medical practice, question 6 isn't optional. HIPAA treats an outside party that handles patients' protected health information for your practice as a business associate, which needs a signed agreement with you before it starts, and that can include a consultant who works with patient records during setup. A consultant's own subcontractors need the same kind of agreement with the consultant, which is why the “who does the work” half of question 6 matters. That's a general description rather than legal advice, and state law can add requirements, so ask your own adviser what your practice needs.
Questions about mistakes
| Question | A good answer sounds like | A red flag sounds like |
|---|---|---|
| 7. What happens when the system gets something wrong? | Uncertain cases go to a person, every action is logged, and errors have a clear reporting path | “It's very accurate.” |
| 8. How will we test it before customers see it? | A side-by-side run on your real calls or documents, against a go-live bar agreed in advance | “We'll launch and tune it as we go.” |
| 9. Which decisions will always stay with our people? | A specific list: EV charger service sizing, anything code-related, and any caller reporting smoke or a burning smell, who hears your approved safety wording and is paged through to the on-call electrician | “It can handle pretty much anything.” |
Push hardest on question 9 if you run a trade. An electrical contractor's after-hours line will eventually get a caller describing a burning smell at the panel. A good setup reads only the safety wording you signed off on (stay clear of the panel, and get out and call 911 if smoke appears or the burning smell grows), then pages the on-call electrician right away. It never diagnoses.
A consultant who can describe that hand-off in detail has thought about your business. One who says the AI will “assess the situation” hasn't.
Questions about money and leaving
| Question | A good answer sounds like | A red flag sounds like |
|---|---|---|
| 10. Once it's live, what will we pay every month? | Usage, subscriptions and upkeep listed separately and priced at your volume | Only the setup fee is ever discussed |
| 11. How do we leave, and what do we take with us? | A notice period, data and recordings exported in a standard format, help with the handover | A long lock-in with no exit before you've seen results |
| 12. Who have you done similar work for, and can we call them? | References in a similar business you can call yourself, or an honest “you'd be our first in your field” | Logos on a slide and nobody to call |
When you make the reference calls for question 12, skip “were you happy?” and ask what you can't learn from a website: Is the system still in use today? What did it get wrong in the first month, and how quickly was that fixed? How did the monthly bill compare with the quote? Would you hire them again for something different?
Ten minutes on the phone with someone who has lived with the consultant's work tells you more than any case study.
Flowchart
Vetting a consultant in one pass
- StartFirst call booked with a consultant
- DecisionDid they ask about your process before naming a tool?NoStepAsk question 1 directly; if they still pitch a product, move onYes
- StepSend all twelve questions and ask for written answers
- DecisionAre ownership, data terms and exit terms answered in writing?NoStepTreat vague answers as a no and get a second proposalYes
- A personYou call two references yourself, without the consultant on the line
- StepAgree a small first phase with one defined deliverable
- DoneSign, knowing what you own, what it costs to run and how to leave
Before and after
A tool-first pitch vs. a process-first consultant
Tool-first pitch
- 01Opens with a demo of their voice assistant
- 02Asks for a signature to start a pilot
- 03Data and BAA questions answered “later”
- 04Running costs vague; contract runs a full year
ResultYou learn about their product, not your practice
Process-first consultant
- 01Opens by asking where the front desk loses time
- 02Asks for call counts and which scheduling system you use
- 03Raises the BAA before you do
- 04Proposes a short first phase with a written plan you own
ResultYou learn something useful even if you never hire them
Red flags that should end the conversation
- Guaranteed results. Nobody can promise a specific outcome before seeing your work. Regulators treat inflated AI claims like any other deceptive claim: the FTC's 2024 sweep, Operation AI Comply, went after companies promising easy AI-powered income and an AI “robot lawyer,” and the agency's then-chair said there is “no AI exemption from the laws on the books.”
- No talk of what stays human. A consultant who can't name the decisions that belong to your people hasn't thought about your risk.
- Pressure to sign today. Good advice survives a week of thinking it over.
- Your data as their asset. Any hint that your recordings or documents will improve their product for other clients.
- Claims you can't check. Partnerships or certifications that the named company or body won't confirm.
Score the answers simply. A clear, specific, written answer is a pass; anything else becomes a follow-up question. If you still have more than two or three follow-ups after a second conversation, keep looking.
To put these questions to us, our first AI consulting call runs 30 minutes, with straight answers and no sales push. If you're still deciding whether you need outside help at all, our guide to the AI opportunity audit shows what a good deliverable looks like.