Law firms usually teach lawyers to get better drafts out of an agent, when the skill they need is supervising its work. Partners excuse themselves, training time goes unrecorded because it cannot be billed, and few people learn how the agent's scope protects client confidentiality. The result is fluent work signed by people who never checked its sources.
Why is drafting skill the wrong focus?
Most law firm sessions on AI teach how to ask for a better first draft: what to include in the instruction, how to request a different tone, how to get the precedent right. That is useful for a few minutes.
The harder skill is review. When an agent drafts a letter of claim, a contract mark-up or a completion statement at depth 2, the lawyer who sends it owns it. They need to check every proposition against its source, see which facts came from the client and which from the firm's own precedents, and send it back with a reason when it is wrong. A fluent draft hides errors better than a clumsy one, which is exactly why review needs training. Teach supervision first; drafting tips can follow.
Why can't partners opt out?
Partners often attend the launch, then leave the detail to associates and paralegals. But professional conduct rules hold the lawyer responsible for work done under their name, whatever tools produced it, and partners supervise the people who review the agent's output.
A partner who does not know what depth the agent works at, or what it means that a draft was held because a source was external, cannot supervise effectively. They also set the culture. If the partner signs agent-drafted advice after a glance, associates learn that glancing is enough. Partners need a shorter session than reviewers, but they need one, and it should be assessed.
What happens to training that cannot be billed?
Time spent learning to oversee an agent is non-chargeable, so it competes with targets. Firms that leave it to individuals get uneven results: the conscientious associate trains in the evenings, others learn on live matters.
Treat agent oversight training as part of the firm's risk management, record it, and give it a time allowance that does not count against targets. The case for funding it as a control rather than as optional development is set out in how to get budget approval for AI agent training, and the numbers that drive the cost are in how much AI agent training costs for law firms.
How does matter confidentiality change the reviewer's job?
Law firms hold information for many clients, some of them on opposite sides of related matters. An agent's role scope is one of the five bounds every action must meet, and in a law firm it should keep the agent within the matter it is working on.
Reviewers must understand this well enough to notice when something looks wrong: a draft that cites a fact the matter file does not contain, or a precedent drawn from another client's work product rather than the firm's approved precedent bank. Training should cover what the agent may read on each matter, how scope is set, and how to report a draft that seems to cross a line. A reviewer who simply deletes the stray paragraph has fixed one draft and left the cause in place.
Which sources should lawyers trust least?
Legal work mixes sources of very different strength. The firm's approved precedent bank, at a fixed version, is sealed. A court order or a document from a named counterparty's solicitors is signed. The client's account of events is declared. An email claiming to come from the client with new bank details for a completion is external, and an external claim cannot authorise a payment.
That last case is where training pays for itself. Conveyancing and transactional teams move client money, and payment instruction fraud often arrives as a convincing email at the worst possible moment. When the agent holds a completion payment at depth 4 because the payee changed through an external channel, the reviewer must know to verify through a known channel before releasing anything.
What should a law firm do instead?
- Make supervision the core of training; treat drafting tips as an extra.
- Give partners a short, assessed session on depth, verdicts and provenance.
- Record training time and protect it from billing targets.
- Teach how role scope protects confidentiality, and how to report a draft that crosses it.
- Drill held payment releases on real past transactions.
A firm can also adopt the practice of tying approval rights for client money to a current certificate, depth by depth. The general order is in the guide to training teams for AI agents, and the detail of content is in what an AI agent training programme for law firms should include. Accountancy firms share the problem of juniors reviewing work they never prepared; what accountancy firms get wrong about AI agent training covers it.
Where the Academy fits
Pryme Intelligence Academy gives partners the Foundations track, about four hours and free, and gives associates, paralegals and supervisors who approve held actions the Operator track. Operator certification requires a written exam and two weeks of real review decisions with no missed red line, signed off by a Pryme assessor. Certificates can be checked at source, which helps when a client asks.
Questions people ask
Should trainees review agent drafts?
Yes, under supervision, because it is now a large part of the work they will do. Pair each trainee's decisions with a supervisor's on the same drafts and discuss the differences. Do not give trainees approval rights for client money.
Do clients need to know an agent drafted their documents?
Explain in engagement terms how agents are used and who reviews their output. Clients mainly want to know that a qualified, accountable person checked the work. Be ready to show that the reviewer was trained and current.
How long should a partner's session be?
Long enough to cover depth, the four verdicts, provenance and what the record shows, and to decide a handful of real held cases. An hour or two is usually enough, followed by a short test.