Start with the people whose names go on the work. A law firm's programme should train supervising partners and senior associates before anyone else, then paralegals and support teams by the tasks they review. It must include practice on real matter files, drills on privilege and hostile documents, sign-off checked by someone outside the team, and renewal.
Why does a law firm train differently from other businesses?
Three things set it apart. Supervision is personal: a partner answers for the work of everyone on the matter, and an agent is now one of them. Confidentiality and privilege attach to almost every document the agent reads. And a large share of what the agent reads was written by the other side.
That last point shapes the whole programme. Disclosure bundles, opponent correspondence and documents uploaded by clients are external provenance. They can contain errors, and they can contain text written to steer an agent, the risk the OWASP Top 10 for LLM Applications calls prompt injection. Every reviewer must learn that an instruction found inside a document is data, not an instruction.
Who takes which track?
- Partners and the risk and compliance team: a short common grounding on depth, held actions, the four verdicts and the record. They set the red lines.
- Senior associates and team leads who approve agent drafts: the daily reviewer track, with an exam and a supervised practicum on their own practice area.
- Paralegals and legal assistants: the reviewer track if they release agent work; otherwise the common grounding.
- Knowledge lawyers and the legal technology team: the builder track, because they write agent briefs and choose the sources.
- Whoever leads deployment across several practice groups: the architect track.
What may the agent do, and what stays with a lawyer?
| Work | The agent | Depth | Stays with a lawyer |
|---|---|---|---|
| Matter intake and conflict search | Searches records, drafts the conflicts note | 1 to 2 | The decision to act |
| Document review | Tags and summarises documents | 2 | Privilege calls and relevance disputes |
| Drafting | Produces first drafts of letters and standard clauses | 2 | Advice, and anything sent to a court or opponent |
| Time and billing | Drafts narratives from time entries | 2, then 3 | Write-offs and client disputes |
| Client account | Prepares transfer requests | 4, held | Every release of client money |
Giving advice stays with a qualified person at every depth. The training should say this in one line and then spend its time on the harder question: what to check before signing an agent's draft.
An example: a conflicts note that looked complete
Take a mid-sized firm's intake team. A new instruction arrives; the agent searches the client database and drafts a conflicts note saying no matches. The trained reviewer notices the verdict log shows the search used only the current client list, because the archive was not connected as a source. The note was accurate for what it read. It was not a clearance.
An untrained reviewer signs it. The programme should replay cases like this from the firm's own past matters, so reviewers learn to read what the agent looked at, not only what it concluded.
How should practice, sign-off and renewal work?
Practice should use anonymised or closed matters from the reviewer's own practice group, with each decision recorded as approve, correct or reject and a reason. Include refusals: an agent that declines to summarise a privileged folder outside its scope is working properly. Include withheld actions too, so lawyers understand that some actions are blocked for every agent at the platform floor, not by the firm's settings.
Sign-off should be checked by someone outside the team and lapse after a year. Firms can tie the right to release agent work to a current certificate, depth by depth. Start with a readiness assessment in a law firm, then roll the training out across the firm one practice group at a time. What law firms get wrong about AI agent training lists the usual mistakes, and the same structure suits accountancy firms. The overall method is in training teams for AI agents.
Where the Academy fits
Pryme Intelligence Academy offers a free readiness assessment for teams of up to 25 people, which suits a single practice group. The team takes the Foundations test set in its industry, and the report shows scores by concept, the gaps and the first agent to start with. Certificates are issued through VerifyDoc.ai, so a client can check a reviewer's status by number.
Questions people ask
Do trainees need the full reviewer track?
Only if they release agent work without a supervisor. Most trainees benefit from the common grounding and then learn review under a certified associate.
Can a firm show clients that its reviewers are trained?
Yes, if certificates can be checked at source. A client panel review can then confirm who is current rather than relying on a training spreadsheet.
What should partners learn first?
What a held action is, what the record shows, and how depth is earned. With those three, a partner can supervise agent work as they supervise people.