AI Risk Levels: What CFM 2.454/2026 Says to Record

The question keeping most technical managers awake is not whether their team uses artificial intelligence. It’s simpler and more uncomfortable: "If an inspector opens a resident’s record tomorrow and asks where the AI use is logged, do I have an answer?" You already know a nurse pasted a symptom into a chatbot last week. You suspect the physician used a decision-support tool during a round. And you have no clean, defensible way to show any of it.

Here is the one sentence that captures what CFM 2.454/2026 changes: from now on, using AI in resident care is not the sensitive part — not recording it in a structured, traceable way is. The resolution doesn’t ask you to stop using AI. It asks you to make its use visible in the medical record, tied to a human who takes responsibility for the decision.

That reframes the whole problem. The panic ("we can’t use AI anymore") is misplaced. The real work is triage: knowing which uses have to be written down, which don’t, and how to prove the difference to someone who wasn’t in the room.

What CFM 2.454/2026 actually asks

Read plainly — and please confirm the exact wording and its application with your legal counsel and your regional council before you build policy on it — the resolution rests on a few principles that matter for daily operations.

First, AI is a support tool, never the decision-maker. A human professional remains responsible for every clinical act, and no software changes that. Nothing you record can transfer that responsibility away from the person who signed off.

Second, when AI meaningfully informs a clinical decision about a resident, that use belongs in the medical record — the same record that already governs medication, evolution notes and events under RDC 502/2021. The AI note isn’t a separate logbook off to the side; it’s part of the resident’s continuous, auditable history.

Third, the record has to be traceable: who used the tool, which tool, for what purpose, and the human validation that followed. A vague "we sometimes use AI" satisfies nobody. Traceability is the entire point.

None of this is exotic. It’s the same logic your institution already applies to a controlled medication: name the substance, name the person, name the moment, name the check. AI just joined that list.

The triage: low, medium, high

Not every keystroke near an AI tool becomes a record entry. The useful mental model — again, to validate with your council — is to sort uses by how directly they touch a clinical decision about a specific resident.

Low risk — generally administrative, usually no clinical-record entry. AI that drafts a shift-handover summary from notes staff already wrote, transcribes a meeting, organizes a cleaning rota, or helps compose a routine message to a family member. There’s no clinical judgment about a resident’s body or treatment. These belong to operations, and your ordinary process governs them. The caution: the moment that "summary" starts suggesting a clinical action, it stops being low risk.

Medium risk — informs care, human confirms, record the use. AI that suggests wording for an evolution note that a nurse then edits and signs; a tool that flags a possible drug interaction the pharmacist reviews; a triage assistant that proposes a priority a professional confirms. The decision stays human, but AI shaped it. Because it touched the clinical reasoning, the resident’s record should show that it was used and that a human validated the output.

High risk — AI touches diagnosis, dosing or a direct clinical recommendation; record it fully, with the responsible professional named. A model reading an image or lab trend and proposing an interpretation; a system suggesting a dose adjustment; any output that, if followed blindly, could harm a resident. Here the entry is non-negotiable: the tool, the purpose, the exact human who reviewed it, and the decision they made. This is where an inspector will look first, and where a missing line is hardest to explain.

The line between medium and high will sometimes be a judgment call. That’s fine — the resolution isn’t asking for perfect philosophy, it’s asking that when AI influenced care, the record shows it and shows the human behind it.

A worked example from one shift

A caregiver on the night shift describes a resident’s new confusion into a phone assistant and asks what it might indicate. The assistant lists possibilities, including an infection. The nurse reads it, decides it’s plausible, and escalates to the on-call physician, who orders tests.

Walk it back through the triage. The caregiver’s query touched a clinical concern about a specific resident — that’s at least medium risk. So the record for that resident should carry: AI consulted for symptom interpretation, which assistant, by whom, and that a nurse and then a physician validated and acted. Written that way, the AI use is no longer a liability hiding in someone’s phone history. It’s evidence of attentive, well-governed care — which is exactly what you want an auditor to see.

Over-the-shoulder view of a technical manager writing on a clipboard at an administration desk, an open laptop angled to the side, warm natural light

Why paper and spreadsheets quietly fail this

You can try to capture all of this on paper or in a shared spreadsheet. The problem isn’t good faith — it’s that neither survives the traceability test. A spreadsheet cell can be overwritten with no history of who changed what, when. A paper note lives apart from the resident’s record, so nobody can reconstruct the sequence months later. And crucially, the fields drift: one professional writes "usei IA," another writes nothing, a third writes a full paragraph. Under inspection, inconsistency reads as absence.

Traceability requires three things paper and spreadsheets don’t give you: an entry bound to the specific resident’s record, an immutable trail of who wrote it and when, and the same structured fields every time so nothing depends on a busy professional remembering the format at 3 a.m.

The minimum fields to capture, every time

  • Who used it — the professional, identified, not "the team."
  • Which AI — the specific tool or system, not "an app."
  • Purpose — what clinical or operational question it served.
  • Human validation — who reviewed the output and what they decided.
  • Responsible professional — the person who owns the resulting act.

If those five are present and unchangeable, you have the spine of a defensible record. If any is optional, it will be the one that’s missing when it matters.

Audit-ready: how Doseo structures this

Doseo was built for ILPIs, so this record isn’t a bolted-on afterthought — it’s a structured field inside the resident’s electronic prontuário, alongside the medication and evolution history you already keep. When a professional logs an AI use, the system captures the five fields above as data, not free text, and stamps it with the author and timestamp in an audit trail that can’t be silently rewritten. Because it lives in the same record governed by RDC 502/2021, and because access is controlled under LGPD with data residency in Brazil, the AI entry inherits the traceability and privacy guarantees your institution already relies on. When an inspector asks "show me the AI use," you filter the record and it’s there — consistent, attributed, complete.

To be clear about the boundary: the structure protects the institution’s ability to prove what happened. It does not, and cannot, absolve the professional of clinical responsibility — the resolution is explicit that a human owns every decision, and no software feature changes that.

Checklist to get compliant

  • Confirm the resolution’s exact requirements and your interpretation of the risk levels with your legal counsel and regional council.
  • Map, honestly, where AI is already being used in your house — including the informal phone-and-chatbot uses nobody logs today.
  • Sort those uses into low / medium / high and write the rule for what each triggers.
  • Standardize the five minimum fields so every entry looks the same.
  • Move the record into the resident’s prontuário, not a side spreadsheet, so it’s traceable and immutable.
  • Update your POPs and staff training so the youngest caregiver on the night shift knows when a use must be logged.
  • Rehearse the inspection question: can you retrieve every AI entry for a resident in under a minute?

CFM 2.454/2026 isn’t asking your institution to fear technology. It’s asking you to be able to account for it — calmly, in writing, tied to the humans who care for your residents. If you’d like to see how a structured AI-use field sits inside a resident’s record and holds up to an audit, take a look at how Doseo organizes it. Then go confirm the details with your counsel, and turn a source of anxiety into one more thing your ILPI simply has in order.

Quer ver o DOSEO na sua rotina?

Agende uma demonstração gratuita e conheça o eMAR de circuito fechado que reduz em até 78% os erros de medicação.

Agendar demonstração ↗