Four artefacts, in order
- Saved note: what the person wrote and is accountable for.
- Summary: condensed text produced from one or more saved notes.
- Extracted item: a candidate action a model believes it found.
- Created task: a real assignment with an owner and a date, made by a person.
What summarization is genuinely good at
Compressing a long, messy history into something readable before a call. Pulling together six visits' worth of notes so the person picking up the phone is not reading for four minutes while the customer waits. Restating a rambling dictated note into three clear lines.
In all of those, the model is serving a reader who still has the original available. That is the safe shape: summary as convenience, note as truth.
Where it goes wrong
- Summaries treated as the record, with the original never read again.
- Confidently smoothing over an ambiguity that mattered.
- Inventing a specific — a number, a date, a name — that was never written.
- Silently creating tasks or messages so nobody knows what a model decided.
- Summaries of summaries, where the detail is now three steps from the observation.
The rule for anything the model wants to do
Reading and summarising is low risk. Writing to a record is medium risk. Anything that leaves the company or cannot be undone — a message to a customer, a financial change, a deletion — is high risk and belongs behind an explicit human approval every time.
Applied to notes: suggest the follow-up, show the person, let them create it. The five seconds you save by skipping confirmation is not worth a system that acts on a misread sentence.