The intake summary that stays on the paralegal side of the line
The situation
My students will spend careers doing initial client interviews and writing them up for a supervising attorney, and firms are already dropping AI notetakers into those calls. What I assess is whether they can produce an attorney-ready memo that spots issues while never once giving a legal opinion, and whether they can articulate why the AI summary needed changing.
Steps
-
Set the consent and confidentiality posture before the call
Platform recording settings; a written consent script
Students draft the on-the-record consent language, demonstrate how to disable default auto-recording, and write a paragraph on where the transcript will be stored and whether the vendor trains on it. This runs first because in practice it is the step that gets skipped.
What you only learn by doing it: Have them actually find the toggle in platform settings, not just describe it. Students who have only read about it cannot find it under time pressure — and default-on recording of a privileged conversation is not a mistake you get to make twice.
-
Conduct the mock intake and capture it
A fact pattern you control; any transcript tool
A role-played client delivers a fact pattern seeded with deliberate traps: a date that shifts mid-conversation, an offhand mention of a prior lawyer, a question a paralegal must not answer. The student discloses their status at the outset.
What you only learn by doing it: Seed exactly one moment where the client asks “so do I have a case?” The correct answer is a referral to the supervising attorney, and the number of students who instead answer it — kindly, helpfully, and in violation of the guidelines — is the most useful data the exercise produces. Score it.
-
Generate the summary, then reconcile it line by line
Any free tier, with the raw transcript beside it
Produce an AI summary, then mark it up: every factual assertion checked against the transcript and annotated accurate, distorted, or absent. The reconciliation markup is the graded artifact, not the summary.
What you only learn by doing it: The failure mode is not invention, it is smoothing. AI summaries resolve the client's contradictions into one clean version of events — the date given twice differently becomes a single confident date. That flattening destroys exactly the inconsistency the attorney most needs to see.
-
Write the attorney memo with an issues section that asks, never answers
A memo template you provide
Facts (reconciled, with transcript citations), chronology, documents to request, and an issues-for-attorney-review section written entirely as questions or flags — “client references a prior representation; conflicts check needed” rather than any conclusion about merits.
What you only learn by doing it: Ban the word “should” from the issues section and watch what happens. Nearly every unauthorised-practice slip in student memos travels on that one verb — “client should file by…” is a legal opinion wearing a procedural costume. A hard lexical rule catches it better than a lecture.
-
Document the AI use and hand off for supervision
A disclosure block appended to the memo
Which tool, what was submitted to it, what was corrected, and confirmation that a human reviewed the full transcript. Students then present to a supervising attorney — you, or a visiting practitioner — who asks what was changed and why.
What you only learn by doing it: This is where students discover that “the AI wrote it” is not an available answer. The supervising attorney is on the hook, which means they will ask, and a paralegal who cannot say what they corrected has made supervision impossible. Run it live if you can; the discomfort is the lesson.
Where this breaks down
Anything AI produces about law must be treated as unverified. A summary that quietly attaches a wrong legal label to a client's facts is harder to catch than a fake citation, because there is nothing to look up.
The professional-responsibility boundary is the core of the exercise. Paralegals may conduct interviews and draft under supervision but may not establish the relationship, set fees, or give legal opinions — and an AI-generated summary that reaches a conclusion does not launder that conclusion into permissibility. The paralegal who transmits it owns it.
Recording a client conversation without consent is deceptive conduct. Consent, storage disclosure and independent human review of every transcript and summary are requirements, not best practices.
Never upload identifiable client information to a consumer account whose terms permit training on inputs. Use a classroom fact pattern with fully fictionalised names.
Provenance: documented for the ethics scaffolding, extrapolated for the classroom sequence. The consent, storage-disclosure, independent-review and supervision requirements come from New York City Bar Formal Opinion 2025-6, which addresses AI recording and summarisation of client conversations specifically. The UPL boundary is NALA's Model Standards and the ABA Model Guidelines.