You Signed an AI-Generated Note With an Error: The Recovery Playbook

The moment you don't want

You signed the note last week. Today, looking at the chart again because the patient called with a question, you notice the AI captured the medication wrong. The dose, the frequency, the drug name, something. The error is in the chart and the chart says you signed it.

This article is what to do next. Not how to prevent the error (covered elsewhere); the recovery playbook from the moment the error is discovered.

The first 10 minutes

The immediate response sets the tone for everything that follows. Three things to do, in order:

1. Don't edit the original note silently. It's tempting to fix the error directly. Don't. Charts have integrity requirements; editing a signed note silently can be construed as falsification, even when the edit was correcting a legitimate error. The right pattern is to add an addendum, not to alter the original.

2. Assess whether the error caused or could cause patient harm. Is the prescription with the wrong dose still in effect? Has the patient picked it up? Has the patient taken it? The clinical urgency drives everything else.

3. Stop adding new content to the chart until you decide your next move. Don't sign other notes for this patient yet. Don't add unrelated entries. Pause on this chart while you figure out the recovery.

These three actions take 10 minutes. They preserve your options.

If the error is medication-related and active

If the patient may be taking the wrong medication, dose, or frequency:

Call the pharmacy immediately. If the prescription hasn't been picked up, intercept it. If it has been picked up, the pharmacist needs to know so they can advise the patient.

Call the patient. Don't email or send a portal message. Call. Voice contact for medication errors is the standard of care. The message:

"Hi [patient]. I'm calling because when I was reviewing your chart, I noticed there's an error in the prescription I sent you last week. The correct dose is [correct dose], not what was on the prescription. Please stop taking [wrong dose] if you've started, and the corrected prescription will be at your pharmacy by [time]."

The conversation should be clear, not defensive. The patient's trust matters more than the explanation of how the error happened. They don't need to hear about AI scribes; they need to hear that you caught it and you're fixing it.

Send the corrected prescription. Make sure it goes out before you end the call with the patient or pharmacy. The corrected prescription should be available before they get off the phone.

Document the phone call. A new note (separate from the addendum on the original error note) capturing the patient communication, the pharmacy communication, and the corrected prescription.

If the error is not medication-related but is in the chart

Other errors (wrong history, wrong physical finding, wrong assessment, wrong plan element that isn't a prescription) follow a similar but less urgent recovery:

Determine whether the error affected care. A wrong allergy in the chart that wasn't relied on hasn't yet harmed anyone but could. A wrong physical finding that wasn't acted on may or may not matter. Be specific about what was at stake.

Determine whether the error affects the patient's ongoing care. Is the wrong content being relied on by other providers, by the patient themselves, by labs or imaging that were ordered based on it?

Determine whether the error affects the chart's medico-legal integrity. A chart with errors is harder to defend if questioned. The correction is part of restoring defensibility.

The urgency calibrates to these answers. Some errors warrant immediate action; some can be corrected at the next chart review without harm.

Writing the addendum

The addendum is the formal correction. Every EHR has an addendum or "late entry" feature. Use it.

The addendum should include:

• The date the addendum is being written
• A clear statement of what was incorrect in the original note
• The correct information
• A brief, factual explanation of why the correction is being made

Example wording:

"Addendum [today's date]: During chart review today, I noted that the medication section of the [original visit date] note incorrectly recorded 'lisinopril 40 mg daily.' The correct medication and dose is 'lisinopril 20 mg daily.' I have contacted [pharmacy] and the patient to communicate the correction and send the corrected prescription. The error has been resolved. Original note is preserved per chart integrity standards."

A few principles:

Don't blame the AI explicitly. "AI-generated note contained error" is acceptable; "AI scribe made a mistake" is unnecessarily fault-attributing. The note has your signature on it. The error is in your chart.

Don't blame yourself defensively. "I should have caught this on review" is honest but doesn't help the chart's defensibility. Stick to facts.

Don't omit information. The addendum should be complete enough that a future reviewer can understand what happened without needing additional explanation.

Don't editorialize. No "I am very sorry" or "this should never have happened" in the addendum. Save the apology for the patient communication, where it belongs.

The internal incident report

Most practices have an incident reporting system for patient safety events. A signed note with an error that affected or could have affected care is reportable.

The internal report should include:

• Date of the original error
• Date of discovery
• Nature of the error
• Patient impact (if any)
• Corrective actions taken
• Root cause assessment (if known)
• Prevention recommendations

Even if no patient harm occurred, the report is useful: it tracks the practice's overall error rate, identifies patterns, and demonstrates a safety culture.

For practices using AI scribes, an additional consideration: identify whether the error was uniquely AI-generated (a mishear, a hallucinated content, a diarization error) versus a category of error that could occur in any documentation method. The distinction matters for prevention work going forward.

Notifying your malpractice carrier

When to notify your carrier depends on the severity:

Notify immediately: any error that caused patient harm, any error that the patient has complained about, any error that's been identified in a chart-review process by a payer or regulator.

Notify proactively: any pattern of errors (multiple errors in a short time, or errors in a specific category) even without harm. Patterns are what concern carriers more than individual events.

Don't necessarily notify: a single, caught-in-review, corrected, no-harm error. This category may be handled internally as an incident report without external notification. Check your carrier's specific reporting threshold; some require notification of any reportable event regardless of harm.

If in doubt, notify. Carriers prefer over-reporting to under-reporting. Late notification of an error that becomes a claim is much worse than early notification of an error that doesn't.

See what your malpractice carrier actually says about AI-generated notes for the broader carrier framework.

The vendor notification

If the error was AI-generated, the vendor needs to know:

Send them the specific note and (if available) the audio recording. They can investigate whether their model behaved as expected. They can sometimes identify a fixable issue.

Ask for their root cause analysis. A vendor that takes patient safety seriously will provide one. A vendor that doesn't will deflect; both responses tell you something.

Document the vendor notification in your incident report. "Vendor notified on [date]; awaiting their analysis."

The vendor's response time matters. A vendor that takes a week to acknowledge a patient-safety report is signaling that they're not treating these reports as urgent. A vendor that responds within 24 hours with substantive engagement is signaling the opposite. Both are useful signals for your continued relationship.

The patient relationship after the error

Patients generally respond well to honest, prompt error disclosure. The relationship is at most risk from the cover-up, not from the error itself.

A useful framing for the patient conversation:

• Acknowledge the error briefly and directly
• Explain the fix (the corrected prescription, the corrected information)
• Apologize sincerely
• Don't blame the technology specifically (it can come across as deflection)
• Invite questions
• Document what you said

The patient's trust often increases after a well-handled error. The pattern is that providers who fear disclosure usually find that disclosure went better than they expected. The patient was worried about being misled or about getting bad care; the prompt correction reassures them that the system caught the error.

Patients who escalate to formal complaints after a well-handled error are rare. Patients who escalate after a poorly-handled error are not.

The post-event review

A week or two after the event, do a brief review:

What was the root cause? The audio was bad; the AI mishears this drug consistently; the review process was rushed because the day ran long; the workflow doesn't catch this specific category.

What would prevent recurrence? A vocabulary customization; a different workflow; a habit change in review; a technology change.

What's the action? A specific change with a specific owner and a specific date.

Did the change work? Check back in 30 days.

This brief loop is what turns a single event into systemic learning. Practices that do it have decreasing error rates over time. Practices that don't have stable error rates indefinitely.

When the error reveals a deeper problem

Sometimes one error reveals a pattern:

• Multiple errors discovered in the same chart-review window
• The vendor's accuracy has been quietly degrading
• The review workflow has slipped over time
• A specific provider has been signing rushed notes

In any of these cases, the response is bigger than one corrected note. The pattern needs to be addressed at the practice level: vendor escalation, workflow change, peer support for the rushed provider, additional review for a calibrated period.

The single-event response and the systemic response are different. Both may be needed.

The honest framing

Signing an AI-generated note with an error is, in 2026, going to happen occasionally to any provider using these tools. The rate is comparable to or lower than typed-note error rates, but the error categories are different.

What separates safer practices from less safe practices isn't the error rate; it's the recovery. Prompt acknowledgment, prompt correction, prompt communication, addendum that preserves integrity, internal report, carrier notification when warranted, vendor notification when relevant, patient relationship preserved.

Done this way, an error is a contained event with a clear closure. Done poorly, it's the start of something larger. The playbook is the difference.

---

For prevention work specifically focused on medications, see what to do when your AI scribe mishears a medication. For the malpractice context, what your malpractice carrier actually says covers the carrier side.