Ten Questions to Ask Any Dental AI Vendor

Why This List Exists

Practices have no framework for vendor evaluation conversations in the AI category. This is a genuine gap, and it is commercially convenient for vendors. A practice that walks into a demo without a prepared set of questions is participating in a sales conversation, not an evaluation. The demo is impressive. The pricing slide comes at the end. The questions that would surface whether this tool is appropriate for this practice rarely get asked.

A practice that walks into the same demo with these ten questions is in a fundamentally different position. The vendor knows they are talking to someone who has thought about this carefully. Weak answers are visible. Good answers are distinguishable from polished deflection. The practice is in a position to make an actual decision rather than responding to whether the demo was exciting.

These questions are not adversarial. They are the questions any practice should expect a vendor to answer with confidence if the vendor has built a sound product they understand well. A vendor who cannot answer them is telling you something important about the product, the organization, or both.

Before any demo: print this list. Assign someone to take notes on each answer. Ask every question, not just the ones you are already thinking about.

---

Q1: What Exactly Does It Do, and What Does It Explicitly Not Do?

Ask the vendor: “Walk us through specifically what this tool does in a dental practice. And then tell us what it explicitly does not handle.”

Why this matters: Vendor demos show the happy path. The demo workflow is selected because it works reliably and looks impressive. It is not selected to illustrate the edge cases, the failure modes, or the situations the tool was not designed for.

The question about what the tool does not handle is more revealing than the question about what it does. A vendor who has thought honestly about their product’s limitations can name them specifically. They can tell you: this tool does not handle secondary insurance processing, it does not work with X PMS, it is not designed for multi-provider scheduling scenarios. That specificity is a signal that the vendor understands their product.

A vendor who struggles to name limitations, or who answers this question by pivoting back to capabilities, is showing you that they have not thought carefully about failure conditions. That is a meaningful signal about the organization’s relationship with honest self-assessment.

Follow-up to be ready for: “Can you give us an example from a recent customer where the tool encountered a situation it was not designed for, and what happened?”

---

Q2: Does It Touch PHI? Will You Sign a BAA?

Ask the vendor: “Does this tool process, store, or transmit any patient health information? If so, will you sign a Business Associate Agreement?”

Why this matters: Any vendor that processes patient information on behalf of a HIPAA-covered entity is a business associate and is required by law to sign a BAA before any patient data is shared or processed. This is a compliance minimum, not a negotiating point.

A vendor who is unfamiliar with the BAA requirement, who says they will need to check with legal, or who expresses reluctance to sign is not prepared to work with a healthcare practice. Many AI tools were not built with healthcare compliance in mind and are being sold into healthcare settings regardless. The BAA question surfaces this immediately.

Verify with your healthcare counsel what the BAA should cover and what language is required for your specific situation.

Follow-up to be ready for: “Which version of your BAA is standard, and how long does the execution process typically take?” A vendor with an established healthcare practice will have a prepared BAA template and a defined process.

---

Q3: Where Is Data Processed and Stored? Is Our Data Used to Train Models?

Ask the vendor: “Where is our data processed and stored, specifically? And is any of our data used to train or improve your AI models?”

Why this matters: Data residency matters for state-specific privacy laws and may be relevant to payer contract requirements. A vendor with offshore infrastructure or processing may create compliance issues depending on the practice’s payer agreements and state law.

The question about training data is separately important. Some AI vendors use customer data to improve their models as a default, and the practice data they are using may include PHI. If patient data contributes to a shared model, that implicates PHI handling obligations and requires specific contractual language. A vendor who uses customer data for model training without disclosure is not a vendor to trust with patient information.

Follow-up to be ready for: “Can you point us to the specific contract language that addresses data residency and model training?” Verbal assurances are less useful than contract provisions.

---

Q4: What Is the Documented Error Rate, and How Were Errors Measured?

Ask the vendor: “What is this tool’s documented error rate? How was it measured, on what dataset, and by what process?”

Why this matters: Every AI system produces errors. The question is not whether errors occur but whether the vendor has measured them honestly and can describe what they mean.

“Our accuracy is 98%” is a marketing claim, not a documented error rate. The meaningful version includes: measured on what dataset (ideally a real-world dental practice dataset, not a controlled test set), over what time period, and using what definition of error. An error rate measured by the vendor on their own internal test set means something different from an error rate measured on a real-world implementation by an independent process.

If no documented error rate exists, ask why. The answer reveals whether the vendor has thought about quality measurement at all.

Follow-up to be ready for: “Has the error rate been measured independently? Can you share the methodology?” A vendor with a mature product should be able to answer this with specifics.

---

Q5: What Happens When It Is Wrong? Who Catches It, and How?

Ask the vendor: “Walk us through what happens when the tool produces incorrect output. Who catches it, and how does that work in practice?”

Why this matters: Every AI system will be wrong at some point. The design question is whether the system is built so that errors are catchable before they cause harm. For any workflow touching patient data, billing, or clinical records, there must be a defined human review point, and that review point must be staffed.

An AI tool that operates autonomously in a consequential workflow with no defined human checkpoint is a tool that will eventually cause a problem that nobody noticed until after the fact. Ask the vendor to walk you through a specific example: a message was drafted incorrectly, an extraction missed a key figure, an appeal was structured wrong. How does the practice discover that? How quickly? What is the correction process?

A vendor who says errors are rare is not answering the question. All vendors say errors are rare. The question is about what happens when they occur.

Follow-up to be ready for: “Which of your customers could we speak to about how they handle output errors?”

---

Q6: What Integrations Exist with Our Specific PMS, and Are They Read-Only or Write-Back?

Ask the vendor: “We use [specific PMS]. What is the exact integration with that system? Is it read-only, or can it write to our records? And what happens when our PMS vendor updates their software?”

Why this matters: “Integration” in a vendor conversation can describe a wide range of realities. On one end: a real-time, bidirectional data connection that reads from and writes back to the PMS with reliability and maintenance. On the other end: a manual export-import process that requires staff time and breaks whenever either system changes.

The question is not whether an integration exists but what it does. A read-only integration can query your data. A write-back integration can modify records in your PMS. Both are legitimate; they are very different things. Knowing which one you are getting matters for both workflow design and risk management.

The PMS update question is separately important. Practice management systems release updates. Those updates break integrations. Ask the vendor how they handle this, how quickly they detect and resolve breakage, and what the practice’s responsibility is during an outage.

Follow-up to be ready for: “Can you give us the contact information for another practice using this PMS that has been live with your integration for at least six months?”

---

Q7: What Is the Total Cost Including Configuration and Ongoing Maintenance?

Ask the vendor: “What is the total cost of this, not just the subscription, including configuration, integration setup, training, and ongoing maintenance? Can you give us a realistic estimate from a comparable practice?”

Why this matters: The subscription fee is the most visible number in a vendor proposal and often not the most significant one. Configuration time (either the practice’s time or the vendor’s billable time) is the most commonly underestimated line. Integration setup, staff training, and ongoing maintenance add further. For complex tools, the implementation cost can exceed the first year of subscription fees.

A vendor who cannot give you a realistic total-cost estimate from a comparable practice is either not tracking this data or not willing to share it. Both are worth understanding before you commit.

Follow-up to be ready for: “What does your implementation timeline look like from signed contract to fully operational? What does the practice team need to contribute during that period, in hours?”

---

Q8: What Does Implementation Actually Require from Our Team, in Hours?

Ask the vendor: “Give us a realistic estimate of the time our team will need to invest during implementation, not just at launch but in the first ninety days. Which roles are involved and for how many hours?”

Why this matters: Practices consistently underestimate implementation burden, and vendors consistently underestimate it in their sales conversations. A tool that requires two weeks of the office manager’s time to configure during an already full quarter is a different commitment than one that requires two hours. These are not the same tool even if the subscription price is the same.

Get a realistic hour estimate based on a comparable practice, not a best-case projection. Ask specifically which roles are involved: the office manager, a biller, a clinical coordinator? What does each person need to do? The answer tells you whether this fits the practice’s current capacity.

Follow-up to be ready for: “What is the most common reason implementations take longer than expected at comparable practices?”

---

Q9: What Is the Exit Path? Can We Get Our Data Out?

Ask the vendor: “If we cancel our subscription, what happens to our data? Can we export everything in a usable format, and how long does the vendor retain data after cancellation?”

Why this matters: Vendor relationships end. Contracts expire, products change, vendors pivot or exit the market. A practice that has accumulated a year or two of contact histories, call recordings, analytical outputs, or content in a vendor’s system needs to know what happens to that data before they are in the situation of needing to leave.

Data lock-in is a real and underappreciated risk. Some vendors make data export difficult by design because locked data creates customer retention. Others have no export capability at all because they did not build one. Find out before you are in the situation.

Follow-up to be ready for: “Can you point us to the specific contract language that covers data export and retention after cancellation?” Read it before signing.

---

Q10: Which Practices Like Ours Can We Speak to Directly?

Ask the vendor: “We would like to speak directly with three practices similar to ours: same size, same PMS, who have been using this tool for at least six months. Can you provide those contacts?”

Why this matters: Vendor-selected references are by definition favorable. A vendor will not connect you with a customer who is frustrated, who had a difficult implementation, or who is considering cancellation. That is understood and expected.

What is meaningful is whether the vendor can provide references that match your practice profile. References from a large multi-location group practice are not useful intelligence for a solo or small group practice. References from a practice using a different PMS tell you nothing about your integration experience. References from a practice that launched three months ago cannot tell you whether the tool holds up at six, twelve, or eighteen months.

If the vendor cannot provide matched references, ask why. If they can provide them but the matches are loose (different size, different PMS, short tenure), that tells you something about the depth of their customer base in your segment.

Follow-up to be ready for: “Are there any customers who tried this and cancelled? Would you be willing to provide a contact there?” This is an aggressive ask and the vendor may decline. Their response, and how they respond, is informative regardless.

---

What to Do with the Answers

Before the demo ends, ask the vendor to send written answers to any question they answered verbally but could not document in the meeting. A vendor who is confident in their answers will have no objection to putting them in writing. A vendor who deflects or says they will follow up and then does not is showing you where the weaknesses are.

Compare the written answers against the verbal answers. Inconsistencies matter. A commitment made verbally in a sales conversation that does not appear in the contract is not a commitment.

Get the BAA before any patient data is shared, the integration agreement before any PMS access is granted, and the total-cost estimate in writing before the subscription agreement is signed. The sequence matters.

---

The Final Filter

The most important AI adoption decisions a practice makes are often not the most marketed ones. A practice that implements solid internal document assistance and a well-governed after-hours inquiry capture tool moves meaningfully up the operational maturity ladder. Those tools are less impressive in a demo than a clinical imaging AI or a predictive scheduling system. They also carry a fraction of the risk and deliver value that is measurable within weeks.

Apply this ten-question framework to the unglamorous tools as rigorously as to the impressive ones. The questions are not less relevant for a message-drafting tool than for a clinical AI. The compliance questions, the error-rate questions, and the exit-path questions matter at every zone of the AI landscape.

A practice that cannot get satisfactory answers to these ten questions from a vendor should not commit to that vendor. Not because the vendor is necessarily dishonest, but because a vendor who cannot answer these questions has not thought carefully enough about the product to be trusted with the practice’s data, the practice’s workflows, or the practice’s patients.