Every piece of AI that is any use in a dental practice does one of three things. A vendor who won't tell you which one is selling you a word.
Most AI conversations in dentistry start in the wrong place — with the model. Which model, how large, how new. None of that tells you the thing that matters, which is what the software is allowed to do inside your practice and who does the work. The useful question is not "how smart is it?" It is "when it is wrong, what has already happened?"
There are only three honest answers, and they are worth learning as words: Advise, Assist and Act.
In Advise mode the software reads your practice and tells you what it would do. It does not post a claim, move an appointment, send a message or change a ledger. A person reads the recommendation and, if they agree, does the work themselves in whatever system they already use.
This is where every practice should start, and for many kinds of work it is where they should stay. An adviser that is wrong costs you a few minutes of reading. It cannot cost you a patient relationship or a payment. The test of a real Advise mode is simple: is there any path by which the software can write to your systems? If the answer is yes, even after a click, it is not Advise — it is Assist with the approval step hidden.
In Assist mode the software prepares the exact action — the claim with its attachment, the recall message with the patient's name and the open slot, the appeal letter with the clinical note — and a named person approves it. Then, and only then, it runs. The value is the preparation, which is most of the work. The control is the approval, which is the part that keeps you accountable.
Ask three things about any Assist mode. Who can approve, and is that a role or a person? Is approval per action or per batch — because "approve 340 recall messages" is not the same as approving one? And can the prepared action be edited before it goes, or is it take‑it‑or‑leave‑it? A well‑built Assist mode records the approval alongside the action, so that six months later you can see what was proposed, who said yes, and what changed.
In Act mode the software does the work inside limits you set, tells you what it did without being asked, and escalates the cases it should not handle. Notice that Act carries two obligations beyond doing the work. It must inform — silent autonomy is not a feature. And it must escalate, both when a rule says so (a cap, a floor, an anomaly) and when the situation is one it recognises it shouldn't handle even though nothing forbids it.
Three questions matter here more than any other. What counts as "proven"? Who decided that? And can the system promote itself? It should not be able to. Promotion should happen one kind of work at a time, on verified results (not on the fact that an action was executed — a claim that submits cleanly and is denied three weeks later is not a success), approved by a person with the authority to say so. And some things should never reach Act at all: voiding a posted payment, adjusting a ledger, changing the software's own authority. Those carry a permanent ceiling, however good the track record.
Two vendors can use the same underlying model and give you completely different risk. The difference is in which mode each task runs in, how promotion happens, and whether every action carries a record of what the system knew, when it knew it, which rule it applied and who approved it. That record is what makes the thing explainable — to you, to your CPA and, if it ever comes to it, to a court. A system that cannot produce it is asking you to trust arithmetic nobody can audit.
There is one more principle worth holding onto, because it separates careful vendors from careless ones: earned autonomy governs how much authority may be exercised; it never rewrites the risk of the action in front of it. A practice that has earned Act on posting routine claims has earned the right to act when circumstances are safe. It has not earned the right to have an unusual, high‑value, anomalous claim treated as routine. Authority should be hard to earn and easy to lose.
If you run more than one location, one rule keeps this sane: group policy can limit what the software does at a location. It cannot expand it. A location's owner sets the ceiling for their practice; the group can only lower it. Anything else puts a regional manager in the position of granting authority over a practice they don't sign for.
Take the three tasks you'd most like to hand off. For each, write the mode you'd accept today and the evidence you'd need to move it up one. Then ask the vendor to map their product onto that sheet. If they can't, that's your answer.
Vella starts every practice in Advise mode, for every kind of work. Nothing moves up without a verified track record and a person saying so — and some things never move up at all. How Vella OS works →
Back to the curriculum — pick the seat you're sitting in and read five, in order.
Canopy University →Ask her. She answers from the lessons here, and says so when the honest answer is that it isn't written yet.
Ask about this lesson →
Every lesson here is written for the person doing the work, whether or not they ever buy from us. Inside Vella OS the same questions get answered from your own record.
The scalable operating system that actually improves financial performance on its own.