Last month our team spent two days at Pri-Med Tampa, and by the end of it we'd had dozens of conversations with clinicians across primary care, psychiatry, internal medicine, and a handful of specialties we don't get to talk with nearly enough. Some had never touched an AI scribe. Most had. What we heard from them about their AI documentation tools was more useful than any market research report could have told us.

My biggest takeaway from these conversations at Pri-Med: the conversation around AI documentation has changed.

For many clinicians, the question is no longer, “Should I use AI?” It's “Is the AI I'm using actually making my day easier?”

And based on what we heard, the answer isn't always straightforward.

What Clinicians Are Really Saying About AI Documentation Tools

Ask a physician who's already using an AI medical scribe how it's going, and more often than not, there's a pause before the answer.

“It's fine, but…”

We heard some version of that again and again.

The “but” varied. The note doesn't sound like me. I have to edit too much. It works well for simple visits but struggles with complicated ones. It includes too much information. It leaves out something I specifically said. It doesn't follow the format I want.

None of those necessarily mean the technology isn't useful. In fact, several clinicians we spoke with liked the idea of using an AI tool to support their documentation. But there was a noticeable gap between having an AI scribe and having one that truly fits the way a physician practices.

After years of supporting physicians through Scrivas, we've learned that documentation is much more personal than it might appear. Two physicians in the same specialty, at the same practice, can have very different preferences for how they document an encounter.

If AI is going to meaningfully reduce documentation burden, generating a technically correct note isn't enough. The note needs to reflect how that individual provider thinks, documents, and works.

Why EHR-Built-In AI Scribes Fall Short

Convenience is important, but it isn’t everything. We heard quite a bit about AI documentation tools built directly into EHRs. The appeal is easy to understand, it is already there, but that alone does not solve the documentation problem. We heard time and time again that clinicians want more customization, better output, and more control over how the AI fits into their workflow. We agree! The technology should adapt to the clinician, not require the clinical to adapt to it.

The PHI and AI Hallucination Questions Every Practice Should Be Asking

One of the conference sessions we sat in on dug into something that doesn't get talked about enough: how the large language models underneath these AI scribes actually handle protected health information, even when the scribing platform layered on top claims to be secure. That session, and the questions that followed it, made it clear this isn't a fringe concern. Clinicians are thinking about it. The same session also reinforced something every AI vendor in this space should say out loud more often: every model can hallucinate. Ours included. The question isn't whether the risk exists, it's what happens next when it does.

We need to be asking how these risks are mitigated? What safeguards are in place? And what happens between AI-generated content and the final clinical record?

A Documentation Gap Hiding in Plain Sight: Multilingual Care

A few practitioners also flagged something we already suspected mattered, and heard confirmed again and again: patients who don't speak English as a first language are getting shortchanged by tools that quietly do a worse job with those encounters. For many practices that is not a minor gap. That's a meaningful share of the patients walking through the door every day.

The challenge is maintaining the meaning and clinical context of the encounter and translating that into documentation the provider can actually use.

Features That Got Clinicians Excited

It wasn't all complaints. Two features that came up again and again as the parts of SAI that made clinicians stop during demos and say, “Wait…show me that again.”

Pre-charting was one. Clinicians leaned in when we showed how SAI puts together a summary of what matters before the visit even starts, using previous context and important patient history, so nobody's walking into a room cold or digging back through the chart themselves.

The other was how SAI Assistant behaves once it's integrated with a practice's EMR. Most of the tools we discussed with clinicians only pay attention to what's said in the current conversation. SAI pulls in a patient's actual history and uses it to inform the clinical insights it surfaces during the visit, not just the transcript from that day. More than one physician told us that was the first time an AI tool had actually seemed to know the patient, rather than just listening to them.

How SAI Addresses What We Heard

A lot of what we heard at Pri-Med validated decisions we've already made in developing SAI.

On the editing burden, this is the whole reason SAI is built the way it is. It's configured per provider and per specialty, not one template stretched across every clinician in a practice, because the fastest way to lose a physician's trust in an AI tool is to hand them a note that doesn't sound like them.

On PHI and security, this isn't something we bolted on after the fact. SAI was built inside the clinical space by a physician who handles PHI every single day, and the security architecture reflects that.

On hallucinations, we don't pretend the risk doesn't exist. No AI model is immune to them. SAI is designed with safeguards that account for that reality, including keeping the physician in control of the final clinical documentation.

On language, SAI documents encounters in languages other than English to the same clinical and billing standard, because a patient's care shouldn't be documented worse because of what language they speak.

At Your Side, At Your Service

And there's one more piece I want to spend a minute on, because it rarely comes up in demos: support. "At Your Side, At Your Service" has been our promise from the beginning and the standard we have built Scrivas around.

As we have evolved to include SAI, that commitment has not changed.

Clinicians are busy. Nobody has an extra hour in their week to sit down and learn a new piece of software, no matter how good it is. Most failed implementations we've seen, and we've seen plenty over the years, don't fail because the tool was bad. They fail because nobody was there to help the practice actually get it running, answer the third question a physician asks on a Tuesday afternoon, or troubleshoot the one workflow that doesn't quite fit. That's what we're there for. Hands-on, from day one, not a support ticket queue you fall into once the sales team moves on.

Pri-Med reminded us why we build SAI the way we do. The questions we heard weren't new to us, but hearing them from dozens of different practitioners, in their own words, back to back for two days, is a different kind of confirmation than any survey could offer. We left Tampa a few new insights and we're already putting them to work.

— Alyssa