AI in the Operatory: A Practical Guide to Getting Started Responsibly 

By Dr. Shreyas Oza, Diplomate, American Board of Endodontics 

I have spent an embarrassing amount of time playing around with large language models over the past couple of years. Some of it has been genuinely useful — even this article has been “polished” by AI! Some of it has been entertaining. And some of it has involved watching an AI confidently give me an answer that was completely wrong. 

Somewhere along the way, I stopped thinking about AI as something I had to sit down at a computer to “use.” I started wondering whether it could actually be useful during a normal clinical day. Not someday. Not after some futuristic integration with the EHR. Could I use it tomorrow, with a patient sitting in my chair? 

The answer is a resounding yes. 

There are obviously much larger conversations happening around AI in health care — privacy, liability, accuracy, patient autonomy, clinical responsibility and how professional organizations should approach all of this. Those conversations are important, and they are still evolving. 

This article is not an attempt to answer them or establish what the profession’s position on AI should be.  

Instead, I want to answer a much smaller question: If an endodontist wants to understand what a large language model, or LLM, can actually do in clinical practice, where should they start? 

My suggestion is surprisingly simple. 

Start with your phone. 

Don’t Start With AI Diagnosing Patients 

When people hear “AI in dentistry,” I think our minds naturally jump to diagnostic applications: interpreting radiographs, analyzing CBCTs, detecting lesions and recommending treatment. 

That is interesting, but it isn’t where I would start. 

I think of an LLM more like a second workspace for my brain. I collect the history. I test the teeth. I look at the imaging. I make the diagnosis. Then I can give the LLM the information I collected and ask it to organize it, challenge it, explain it or document it. 

That difference is important. 

I don’t particularly want AI telling me what to do. I am much more interested in asking it: 

“What am I forgetting to think about?” 

Here’s What This Could Actually Look Like 

Imagine an emergency patient walks in. You do exactly what you already do: chief complaint, history, PA, cold, percussion, palpation, probing, bite test — whatever the situation requires. 

Nothing about the examination changes. 

When you’re finished, take out your phone, open your LLM of choice, hit the microphone and dictate: 

“42-year-old male. No. 30. Three days of spontaneous pain. Wakes him at night. Ibuprofen helps. No swelling. Cold on 29 normal, 30 severe lingering approximately 25 seconds, 31 normal. Percussion 30 positive. Palpation negative. Probing normal. No mobility. PA shows a deep restoration approaching the pulp and normal apical structures. Organize this consultation, suggest the pulpal and apical diagnostic categories, and tell me what I may have missed.” 

That takes maybe 20 seconds. 

No patient name. No date of birth. No chart number. 

For someone learning how to incorporate LLMs into clinical practice, I think intentionally using de-identified information is the easiest place to start. Once protected health information, or PHI, or continuous recording enters the equation, there are additional privacy, contractual, security, legal and practice-policy considerations. 

For now, keep it simple. 

Give It Instructions Once 

You also don’t want to explain the assignment every time you see a patient. I would create a dedicated conversation or appropriately configured workspace and give the LLM standing instructions along these lines: 

“When I give you an endodontic consultation, organize it into relevant history, diagnostic testing, proposed pulpal and apical diagnoses, missing or inconsistent findings, things I may have overlooked, reasonable treatment options, important points to discuss with the patient, and a concise draft clinical note. Never invent missing information. Clearly identify uncertainty. I will independently determine all diagnoses and treatment decisions.” 

Now you can simply dictate your findings. 

The goal isn’t to become a world-class prompt engineer. The goal is to spend as little time prompting as possible. 

My Favorite Question: “What Doesn’t Fit?” 

This is where I think things get interesting. 

Imagine a previously treated No. 19. Percussion is strongly positive. Palpation is negative. Probing is normal everywhere except for an isolated 9-mm distal-lingual defect. There is a J-shaped radiolucency associated with the distal root. 

You could ask: 

“What is the diagnosis?” 

Which, in my opinion, would be the wrong question. 

Instead, ask: 

“What doesn’t fit here, and what should I make sure I have considered before recommending treatment?” 

The model may point out that the isolated probing defect and radiographic pattern should raise concern for a vertical root fracture. 

Now, hopefully, no endodontist reading The Advocate just learned for the first time that an isolated deep probing defect could be associated with a root fracture. 

That’s not the point. 

The point is that we practice in the real world. You’re running behind. The patient is uncomfortable. Your assistant needs you in another room. The referring dentist wants to know whether the tooth is savable. Your next patient arrived 15 minutes early, which somehow makes you feel even later. 

In that environment, having something ask, “Did you think about this?” can be useful. 

I have started thinking of an LLM less as an oracle and more as a checklist that talks back. 

Use It When Something Weird Happens 

The same concept can be useful during treatment. 

Let’s say you are treating No. 30 and the distal canal continues to drain despite instrumentation and irrigation. 

I would not ask: 

“Should I obturate?” 

Instead, my query would be: 

“I’m getting persistent clear drainage from the distal canal despite instrumentation and irrigation. Give me the most important things I should consider before deciding whether to obturate today.” 

That’s a very different relationship with the technology. 

Or maybe the preoperative CBCT suggested possible middle mesial anatomy. You have found MB and ML and see a developmental groove between them. An appropriate ask would be: 

“Give me a quick checklist of what I should evaluate before concluding whether a separate middle mesial canal is present.” 

Maybe it tells you something useful. 

Maybe it tells you five things you already know. 

Either outcome takes about 15 seconds. And you still make the decision. 

Then Let It Do Something We Actually Hate Doing 

At the end of the appointment, AI becomes considerably less philosophical and considerably more useful: documentation. 

Before leaving the operatory, hit the microphone: 

“Treatment completed. Four canals. MB, ML, DB and D. Working lengths 20, 21, 19 and 19.5 millimeters. Final sizes 30/.04 mesials, 40/.04 distals. Sodium hypochlorite and EDTA irrigation. Obturation with gutta-percha and sealer. Patient tolerated treatment well. Draft my treatment note and postoperative instructions.” 

One practical trick makes this even easier: Keep the same LLM account open on both your phone and your operatory desktop. 

I can dictate the encounter on my phone, and the response is waiting for me on the computer almost immediately. From there, after reviewing it for accuracy, I can copy the relevant text into the patient’s record. 

For SOAP notes, consultation reports and treatment summaries, this can turn what used to be several minutes of typing into a remarkably quick workflow. Now you have a draft while everything is still fresh in your head. 

Read it. Correct it. Add anything important. Then put the appropriate information into the record. 

I cannot emphasize the “read it” part enough. LLMs have an impressive ability to occasionally say something completely incorrect in language that makes you briefly question whether you are the one who is wrong. 

Never confuse eloquence with accuracy. 

AI Is Also Pretty Good at Speaking Patient 

Another place I think LLMs can be genuinely helpful is patient communication. 

We spend our days explaining irreversible pulpitis, necrosis, cracks, retreatment, resorption, apical disease and postoperative pain. Eventually, you forget that most people don’t use the phrase “symptomatic apical periodontitis” before lunch. 

Try: 

“Explain irreversible pulpitis to a nervous patient in 30 seconds without using dental terminology.” 

Or: 

“Give me a simple analogy for why a root canal sometimes needs retreatment even though the patient already had one.” 

Or: 

“Explain why this tooth may remain sore for several days without making it sound like the treatment failed.” 

You don’t necessarily have to read the response to the patient. Sometimes it simply gives you a better way of saying something you’ve explained a thousand times. 

And Then We Have to Talk About Privacy 

This is where “using AI” becomes too broad a phrase. 

There is a big difference among these three scenarios. 

Level 1: De-identified assistant 

You intentionally tell the LLM: 

“45-year-old female, No. 19, percussion positive …” 

No patient identifiers. No ambient listening. 

That’s where I would start. 

Level 2: PHI-enabled assistant 

Now the system receives identifiable patient information or interacts with the patient’s record. 

At that point, you need to understand the particular product you’re using, its data handling, security, contractual arrangements, applicable privacy requirements and your own institutional policies. 

A familiar logo on the screen does not automatically make every use of that product appropriate for PHI. 

Level 3: Ambient AI 

Now the AI listens to the actual patient encounter, potentially creates the note, accesses records or imaging, or becomes integrated into the clinical workflow. 

This is exciting technology, but it is also a very different privacy and consent conversation. 

Patient disclosure, consent, recording laws, data retention, transparency, autonomy and professional responsibility all become much more relevant. 

So when someone asks, “Do I need an AI consent form?” I don’t think there is a useful one-word answer. 

The better question is: What exactly is the AI doing? 

An LLM rewriting a de-identified explanation of postoperative discomfort is not the same thing as a system recording an entire patient encounter and placing information into the health record. 

We shouldn’t treat them as though they are. 

Keep the Human in Command 

In its position paper on responsible AI integration, the American Association of Orthodontists used a phrase I really like: “Human-in-Command.” 

It describes the relationship I want with these tools. 

AI can organize my findings. It can challenge my thinking. It can draft my note. It can remind me of something. And it can occasionally make me look smarter than I am when I need a good analogy. 

But I need to be capable of evaluating everything it produces. 

That matters because there is a danger more subtle than simply getting a wrong answer: accepting an answer because a computer delivered it confidently. 

If AI gives you a citation, verify it. 

If it writes your note, read it. 

If it suggests something clinically unexpected, investigate it. 

And if the answer doesn’t make sense, don’t torture your clinical findings until the computer becomes right. 

If You’ve Never Done This, Try One Thing Tomorrow 

Don’t buy anything. 

Don’t redesign your workflow. 

Don’t integrate AI into your practice-management software. 

Take your phone. 

After your next consultation, de-identify the case and spend 20 seconds dictating your findings. 

Then ask: 

“What doesn’t fit? What information am I missing?” 

Read the answer. 

At the end of the procedure, dictate what you did and ask it to draft your note. 

Then read that too. 

Try it for a few cases. 

You will figure out very quickly where the technology saves you time, where it helps you think and where it is more trouble than it is worth. 

And that may be the best way to learn AI — not through another webinar explaining what a large language model is, but by actually using one. 

Eventually, once you have streamlined the process for yourself, you may find new, creative ways to use the tool in your practice. You may also identify appropriate tasks that can be delegated to dental assistants, making the process even more efficient. 

You, the endodontist, still do the endodontics. 

The Endodontist Still Does the Endodontics 

Our specialty has never exactly been afraid of technology. 

We work through microscopes. We interpret CBCTs. We use electronic apex locators, ultrasonics and increasingly sophisticated irrigation technologies. 

We can also spend an astonishing amount of time debating .04 versus .06 taper. 

LLMs are simply another kind of tool. 

They aren’t going to negotiate MB2 for you tomorrow morning. But they might organize the information surrounding the case. They might challenge an assumption. They might help explain something to a patient. They might turn 20 seconds of talking into the first draft of a clinical note. 

And maybe the best question isn’t, “Can AI do this for me?” 

Maybe it’s: 

“Can AI help make sure I haven’t forgotten something while I do it?” 

That is the role for AI in endodontics that interests me most right now. 

And unlike most new technology in our specialty, you can find out whether it’s useful without asking your practice owner for another $80,000.