A patient texts a dental office at 9 p.m.:
Patient: Do you take Delta Dental? And can I move Thursday's cleaning?
Those are two different kinds of question. The first is answered from the office's own insurance page:
AI: Yes, we're in network with Delta Dental PPO.
Get that wrong and a patient shows up expecting coverage they don't have. The second is a booking, and no AI should be making that up:
AI: I see your cleaning with Dr. Lee on Thursday at 2:00. I can offer Friday at 10:00 or Monday at 3:30. Which works for you? Patient: Friday. AI: To confirm: I'll move your cleaning with Dr. Lee from Thursday at 2:00 to Friday at 10:00. Shall I go ahead? Patient: Yes. AI: Done. You're booked for Friday at 10:00.
Every time offered came from the office's real calendar, and nothing changed until the patient said yes.
That one text holds the number that rarely makes it onto an AI customer service spec sheet: accuracy, how often the answer is right. Speed, channels and deflection are easy to measure, so they're what most spec sheets lead with. How often the answer is right is harder to measure and harder to publish. Here is ours, and how we reach it.
Two kinds of accuracy
The booking is a transaction, and a transaction can't be "mostly right". In our system the AI doesn't generate it at all. It recognizes that the patient wants to reschedule and hands the conversation to JSFE, a deterministic workflow engine we built. JSFE finds the real record in the business's own scheduling system, offers only valid options, validates each answer, reads back from the system's own data, and commits only on an explicit yes. Otherwise it cancels and offers a person. That is why those transactional steps, payments, bookings, account lookups, carry a 100 percent accuracy guarantee: the worst case is a request it didn't recognize, most likely caught on the next turn, never a wrong action. We explain the mechanics in How JSFE Brings Trusted Workflow Control to Agentic AI.
The insurance question is different. It is a conversational answer, generated from the office's own information, and that is where the hard accuracy problem lives.
Our number
Our conversational answers measure 97 percent or higher, in production at Omie in 2024 and at Curacao in 2025, against an industry average that tends to sit closer to 90.
Seven points sounds small until you do the math. At 90 percent, one in ten answers is wrong. At 97 percent, it's closer to one in thirty. So the 90 percent tool makes roughly three times as many mistakes, and across the thousands of questions a busy business gets, each one is a customer told the wrong thing: the patient who shows up expecting coverage, the complaint waiting to happen.
How it's reached
It looks things up instead of remembering. One way to make an AI answer about a business is to bake the information into the model through retraining, which goes stale the moment the business changes. The other, the one we use, is called retrieval: the AI looks up the business's current information every time it answers and builds the response on what it found. For the insurance question, that means answering from the insurance page as it reads today, not from something memorized months ago.
It has to find the source first. AI goes wrong by generating answers that sound right and aren't grounded in anything. An AI that must find the relevant piece of real content before it answers is structurally much harder to lead into confident nonsense.
It doesn't rely on one method or one model. We combine several retrieval methods, so the system is better at finding the right page in the first place, and we can draw on the best-suited model from across competing AI vendors for a given task. That is where much of the gap over the industry average comes from. The same choice helps availability: if one provider has an outage, the system falls back to another, so the answers keep coming and stay grounded.
The limits, stated plainly
Generated answers aren't 100 percent, ours included. 97 percent means about three in a hundred conversational answers still miss, and most of those misses trace back to content the search didn't find, which is why they can be found and fixed.
Accuracy also depends on what the system is given. Retrieval can only find answers that exist, so thin, contradictory or out-of-date information drags the number down no matter how good the system is. If the insurance page still lists a plan the office dropped last year, the answer will be wrong in exactly the same way.
Misses you can find and fix
Because answers are anchored to the business's content, a wrong one can almost always be traced to a source page that was unclear, missing or contradicted by another. Fix the page and the answer is fixed. Staff can mark any answer correct or incorrect in the dashboard, the system flags answers it judged incomplete, and the list filters down to just the misses, which gives a map of where the information has gaps.
Questions to ask any AI vendor
- What is your accuracy on conversational answers, measured in production, and where?
- Which steps are generated by a model, and which are handled by something deterministic?
- When it takes an action, does it read back from the system's real data and wait for a yes?
- When it's wrong, can you see why, and fix the source?
- What happens when one AI provider goes down?
Back to that 9 p.m. text. The patient got a correct answer about coverage and a cleaning moved to a slot that really exists, without anyone at the office picking up a phone. That is why we're willing to put our number in writing.
