The Audit Cost of an AI Scribe: The Number Nobody Puts on the Landing Page

Every AI scribe demo ends at the same moment. The visit wraps, the draft appears, everyone nods. What the demo never shows is what happens next: a physician reading that note line by line, deciding whether to put their name on it.

I'm a practicing physician and I built one of these products, so I'll say the quiet part: generation was never the hard problem. Generation is measured in seconds and printed on every landing page. Verification is measured in your evening and printed nowhere. Call it the audit cost of the note. It's the time between "draft ready" and "signed," and for a while the industry acted like it didn't exist.

Then the research started coming in.

What the Published Evidence Found

Four results from the last two years are worth reading in full. None of them are mine, and I'm going to describe them the way I'd describe them to a colleague: factually, without spin, because each vendor involved deserves credit for the data existing at all.

The hallucination study. In October 2025, a blinded study in Frontiers in AI (DOI: 10.3389/frai.2025.1691499) compared ambient AI notes against physician-written notes across 97 encounters, 194 notes, and 388 blinded reviews. Reviewers found hallucinated content in 31% of the ambient AI notes versus 20% of the physician notes (p=0.01), and rated the AI notes lower on accuracy and succinctness. The study was funded by Suki, an AI scribe vendor, and they published the unfavorable result anyway. That's worth respecting. It's also worth sitting with: roughly one in three ambient notes carried something that wasn't said in the room.

The null RCT. NEJM AI published a head-to-head randomized trial (DOI: 10.1056/AIoa2501000) covering 238 physicians across 14 specialties. Nabla reduced documentation time by 9.5% (p=0.02). DAX Copilot, now sold by Microsoft as Dragon Copilot, reduced it by 1.7%, a result that was not statistically significant (p=0.66). An earlier 2024 evaluation of DAX in JAMIA had already reported no quantifiable effect on patient safety and no productivity benefit. Time savings is the industry's headline claim, and in its first rigorous head-to-head test, DAX couldn't demonstrate it.

The deployment safety data. A 2025 preprint from Dai and colleagues (ML4H, arXiv: 2512.04118) analyzed 50,123 encounters from 470 physicians using a deployed ambient scribe. In the safety feedback physicians submitted, medication errors were the largest cluster at 18.5%, and hallucinated or nonexistent conditions in the HPI and assessment-and-plan accounted for 15.4%. It's a preprint, so hold it loosely until peer review. But the error taxonomy matches what the blinded studies keep finding: the failures cluster exactly where clinical risk lives.

The measurement problem. A scoping review on medRxiv (2025.01.29.25320859) reached a conclusion that should reframe every vendor conversation you have: there is no standardized way to measure hallucination rates across products. Vendor accuracy numbers are not comparable to each other. When a scribe company quotes an accuracy number, the honest follow-up is: measured how, and verified by whom?

Audit Cost Is Becoming Revenue Cost

For most of us the audit cost lands as time. Since late 2025, it also lands as money. Cigna's downcoding policy, in effect since October 2025, automatically downcodes level 4 and 5 E/M claims when the documentation doesn't support the complexity billed. Read that alongside the hallucination data and the problem sharpens: a note that sounds fluent but can't show its evidence isn't just a liability risk you have to proofread. It's a claim that pays less.

The payer trend and the research point the same direction. What gets rewarded now is not a fast draft. It's a defensible note.

The Architecture the Trials Actually Support

Here's the part of the literature that gets less airtime. Go back to that NEJM AI trial (Lukac and colleagues, the same paper) and look at what the winning arm actually was. Both arms used the same workflow: the scribe drafts, the physician reviews and edits before signing. The draft was never consumed raw in either arm. Same trial, same workflow expectations, two products, one significant reduction and one null result. The lesson isn't that ambient scribes work or don't. It's that the architecture behind the draft, and the review it has to survive, decide whether the category delivers. The draft alone isn't the product. The draft plus a disciplined way to check it is.

That distinction matters when you read study designs, too. In the Frontiers hallucination study, the hallucination detection was performed by human reviewers as part of the research protocol. Detection was the study's method. It wasn't the product's feature. When you evaluate a scribe, ask which one you're getting.

How We Build for Time-to-Signed-Note

Scribeable's answer to all of this is an architecture choice we made early: two passes, not one. A first pass drafts the note. A separate second pass verifies it against what was actually captured, and when something is ambiguous, it asks you a clarification question instead of guessing. That interruption is deliberate. A guess that reads smoothly is exactly the failure mode the hallucination literature describes.

• Unsupported statements get flagged. Recommendations in the plan that aren't tied to anything in the encounter are marked, with an integrity banner on the note, so your eye goes to the sentences that need it most.
• Structured data is traceable. Labs, vitals, and medications in the note are highlighted back to their sources, so verifying a value means glancing, not hunting.
• Calculations are code, not vibes. 236 clinical calculators are scored deterministically in code and validated against regression tests, because a risk score should never be a language model's recollection.

And the honest scope, because this article would be hypocrisy without it: our unsupported-statement flagging is a deterministic safety net. It catches recommendations with no grounding in the encounter. It is not semantic hallucination detection, no product I'm aware of has credibly solved that, and per the scoping review above there isn't even an agreed way to measure it yet. You still read your note before you sign it. The goal of the design is to make that read fast and focused, not to pretend it away.

The Tradeoffs, Stated Plainly

• A verification pass takes longer than a single generation pass. We spend more compute and more seconds per note than a draft-only pipeline would. That's the cost of the design, and we think the evidence supports paying it.
• Clarification questions interrupt you. Sometimes the model asks when you'd rather it had guessed. We chose the interruption over the silent guess, and not everyone will prefer that.
• No independent, standardized hallucination benchmark exists for us either. The measurement critique in the scoping review applies to every vendor in this market, including this one. What we can show you is the mechanism, in the product, on your own notes.

Questions to Ask Any Vendor, Including Us

• What is your hallucination rate, measured how, and verified by whom?
• Does the product itself flag unsupported statements, or did human reviewers do that in your study?
• What happens when the model isn't sure: does it ask, or does it guess?
• What is your time-to-signed-note, not your time-to-draft?
• Show me the verification workflow, not the generation demo.

You can compare how we stack up against the enterprise products on the Abridge, Nuance DAX / Dragon Copilot, and Suki comparison pages, where every competitor claim is sourced and dated. Or skip the reading and test the mechanism directly: the trial is 14 days or 15 notes, no card, and pricing is published at $39 to $79 a month.

The market sells time-to-draft. The research, the payers, and your own evenings are all pricing time-to-signed-note. A note isn't done when it's generated. It's done when you'd stake your signature on it.