Redlining against a playbook: when the AI's issue is not your client's issue
The situation
Contract review is the work my students are most likely to be handed on day one, and it looks most automatable until you watch someone do it well. The difference between a junior reviewer and a good one is not spotting the indemnification clause — AI does that fine — it is knowing which of fifteen flagged issues this client actually cares about, and what the fallback is.
Steps
-
Build the playbook before you see the contract
For a defined client, write preferred positions, acceptable alternatives and walk-away lines for six to eight key terms — liability caps, indemnification, IP ownership, termination, payment, confidentiality — each with a rationale. Written in advance it is a commitment device; written afterward it is a rationalisation.
What you only learn by doing it: Require the fallback column for every term and do not accept “negotiate.” A playbook with preferred positions and no fallbacks is what most students submit and it is useless in practice — the entire value is knowing in advance how far you will move, because in a live negotiation you will not compute that calmly.
-
Run the AI review against the playbook as explicit context
Feed the counterparty's contract and the playbook together and ask for deviations from the playbook rather than for “risks.” A generic risk review returns the model's average of all contracts it has seen; a playbook-anchored review returns deviations from this client's position.
What you only learn by doing it: Have students run it both ways on the same contract and diff the outputs. The generic run produces a longer, scarier, more impressive-looking list that is mostly irrelevant to the client. Seeing those two lists side by side teaches materiality faster than anything else.
-
Triage the flags into material, immaterial, and wrong
The contract itself; a triage table
Every flagged item gets sorted: material to this client, present but immaterial, or not actually in the document, with a section number cited for each. The third bucket justifies the whole exercise.
What you only learn by doing it: The “not actually in the document” bucket is never empty, and the errors are asymmetric in a dangerous way. Models are better at flagging clauses that exist than noticing ones that are missing — an absent limitation-of-liability provision generates no flag at all. Add an explicit checklist pass for absent terms.
-
Draft the redlines and write the cover note
Word track changes — the format the receiving side expects
Actual tracked changes plus a short cover note explaining each material change in business terms and what the fallback is if refused. The cover note is what turns a marked-up document into a negotiating position.
What you only learn by doing it: Redline in Word regardless of what generated the suggestions. Students who paste AI output as clean replacement text destroy the change history, and opposing counsel will notice immediately. The tracked-changes artifact is the professional deliverable; the clause language is only half of it.
-
Defend the redlines against pushback
A live session with you or a practitioner as opposing counsel
Students defend each material change, get pushed back on, and must either hold or go to their documented fallback. Changes they cannot explain in business terms get struck.
What you only learn by doing it: Ask “why does your client care about this one?” on a clause the AI flagged and the student kept without thinking. The silence is instructive, and it is the moment students stop treating the AI's list as a to-do list.
Where this breaks down
For paralegal students this sits close to the unauthorised-practice line and the line must be said out loud: drafting and flagging under attorney supervision is permitted work, but advising a client that a term is acceptable is a legal opinion and is not. The contract does not become a business document because the AI framed its output in business language.
AI contract review misses absent provisions far more often than it misstates present ones, so a clean AI report is not evidence of a clean contract.
Models will invent plausible clause language, cite standards that do not exist, and describe market norms with no basis. Treat any claim about what is “standard” as unsupported unless the student can point to a real comparable.
Never upload an actual client's or employer's contract to a consumer account. Use classroom fact patterns or publicly filed agreements — SEC EDGAR exhibits are free and real.
Provenance: documented practitioner practice, extrapolated to the classroom. The playbook structure — preferred positions, acceptable alternatives, fallbacks — and the AI-against-playbook review pattern are documented in published contract-playbook guidance and in a state bar association write-up of the tooling. The two-run diff, three-bucket triage and absent-terms checklist are ours.