Pasting a meeting transcript into a chat and asking for "minutes" seems to solve the paperwork in seconds. The text comes out well-written, organized in bullet points, and with a professional tone. The problem is that well-written is not the same as correct, and in clinical records, that difference is significant.

Text that only looks good

In English, AI-generated content that looks polished but is generic, long, and empty is called slop. In clinical practice, it appears in some well-known ways:

  • Sentences that apply to any child: "the child showed significant progress in various developmental areas." It does not state which area, how much, or compared to what.
  • Too much text: three paragraphs where three lines would suffice. No one reads it again, and what matters gets lost in the middle.
  • Certainty about things not decided: an idea raised in the middle of a conversation becomes "it was decided that."
  • Fabricated facts: a number of trials that no one recorded, a professional who was not in the meeting, a statement attributed to the mother that she did not say.
  • Relative dates: "tomorrow" and "next week" copied from the conversation, which lose their meaning when the minutes are read a month later.

The most dangerous part is that these errors do not look like errors. The fluid text builds confidence, and those in a rush sign it.

The risks of generic AI

An open chat tool knows nothing about the case. It does not know the child, the plan, the ongoing goals, or the team. Therefore, it fills in the gaps with what is likely, not with what is true.

There is also the privacy risk. Copying a child's data to a tool outside the electronic record takes that information out of the clinic's control: who accesses it, for how long, and for what purpose. A child's health data is sensitive information.

How much time generic AI actually saves

The promise is to write in seconds. In practice, without integration, the path is usually different:

  1. Copy the transcript, or part of it, from where it is.
  2. Explain to the tool who the child is, who the therapist is, who the mother is, and what is in the plan.
  3. Read the result and realize it is long, generic, or has something wrong.
  4. Ask again: "shorter," "without that part," "the decision was not that," "the date was different."
  5. Repeat until it is acceptable.
  6. Check line by line against the transcript, because the tool does not show where it pulled each sentence from.
  7. Copy the final text back to the electronic record, adjust the formatting, and define who can see it.

Each iteration of this back-and-forth costs minutes, and the final check is the slowest part: without a reference to the source excerpt, reviewing generated text is almost as laborious as writing it. At the end of the month, the time saved in drafting was spent explaining, asking again, and checking.

With integrated AI, most of these steps disappear. The context is already in the system, the format is already that of a clinical record, each suggestion points to the excerpt that supports it, and the reviewed text is saved directly in the right place. Reviewing becomes attentive reading, not a rewrite.

What changes when AI is integrated

AI makes more sense within the system where clinical work already happens. Integrated into the electronic record, it works with context, limits, and review:

  1. Real context: it knows the child, the associated team and family members, and the ongoing plan. It does not need to guess who is who.
  2. Visible evidence: each suggestion or answer shows the transcript excerpt that supports it. The reviewer checks it in seconds, without rereading everything.
  3. Clinical format: the text follows a format designed for records: third person, absolute dates, only the final decision, actions grouped by person responsible. It comes out shorter and more objective.
  4. Nothing becomes a record without review: the AI suggests and writes, but the professional reads, adjusts, and decides what to save.
  5. Same access rules: the generated text stays in the electronic record, with the same visibility as any note: team, family, or only the author.

How to review text generated by AI

Even with a good tool, the responsibility for the record lies with the person who signs it. Before saving, check:

  • Are the names and roles correct? Was someone mentioned who did not participate?
  • Do the numbers match what was recorded in the session?
  • Are the dates absolute and correct?
  • Was what is marked as a decision actually decided?
  • Can anything be cut without losing information?
  • Who will be able to see this note?

How Proaba uses AI

At Proaba, the AI works within the clinical record:

  • Voice note: after transcription, Proaba suggests what to write, such as supervision notes or a plan update for the family, and each suggestion shows the excerpt that justifies it.
  • Minutes: absolute dates, clinical record in the third person, only the final decision, actions by person responsible, without attributing statements to those who did not speak.
  • Session recording: the summary cross-references the transcript with the plan and shows what was planned, what was recorded, and what happened but has not yet been recorded.
  • Ask the session: each answer comes with the transcript excerpts that support it, and one click leads to the point in the video.

In all cases, the professional reviews before saving and chooses who can view the record. The AI saves the drafting effort, and the clinical judgment remains with the team.