Supervision is already the backbone of legal practice, so build the proof on it. For each matter where an agent drafts, the firm should show who approved the work, that they held a current certificate checkable at source, and that the supervising lawyer reviewed it. That chain is what a client's outside counsel guidelines are likely to ask for.
Who asks law firms for this evidence?
Three groups, with different motives.
Clients. Corporate legal departments write guidelines for their outside firms, and those guidelines can ask how AI is used on their matters and who checks it, sometimes with evidence before work starts.
Regulators and professional bodies. They expect supervision of non-lawyer and junior work to be real. An agent's draft is another form of delegated work.
Insurers. Professional indemnity insurers want to know that a missed deadline or wrong advice cannot easily originate in an unchecked agent output.
All three are satisfied by the same evidence, provided it works at the level of the matter, not the firm. A firm-wide statement that "all fee earners have completed AI training" answers none of them for a specific file.
What should the evidence chain look like?
| Link | What to keep |
|---|---|
| Agent use on the matter | Which agents worked on the file, at what depth, on which tasks |
| First check | The paralegal or trainee who approved each draft, and their certificate status at the time |
| Supervision | The lawyer who reviewed the work, and when |
| Training content | What approvers were trained and tested on, including the firm's own matter types |
| Currency | Renewal dates and briefings when agents changed |
This is close to what firms already keep for supervision. The addition is the certificate status at the time of approval, and the agent's own record of what it drafted and what was changed. Together they show that a qualified person stood between the agent and the client.
What should lawyers and staff be trained on?
Legal agents usually draft chronologies, research notes, disclosure lists, standard correspondence and first drafts of documents. The errors that matter are specific:
- Authority mistakes. A research note citing a case that does not support the proposition, or a superseded rule.
- Date and party errors. Deadlines computed from the wrong trigger, letters addressed to the wrong party.
- Source confusion. A date from the court's sealed order versus one in the other side's letter, which is external.
- Confidentiality. Material from one client's matter appearing in another's draft.
Staff should be tested on drafts with these errors seeded in. For a sense of what goes wrong when firms skip this, see what law firms get wrong about AI agent training.
An example from a client audit
A bank's legal department audits its panel firms. It asks one firm for evidence on three matters where agents drafted disclosure lists. The firm produces, for each, the agent's draft, the approving paralegal and her certificate number and status, the supervising associate's sign-off, and a note of the changes made.
On one matter the paralegal's certificate had lapsed during the matter. The firm's records show that her approval rights on disclosure lists were suspended on the expiry date, the next two lists were approved by a colleague, and she renewed a week later. The client accepts the evidence because the firm's controls caught the lapse before it mattered.
Which controls make the evidence credible?
- Matter-level records of agent use and approvals, not just firm-wide policy.
- Approval rights depend on current certificates, so a lapse suspends rights automatically.
- Supervising lawyers review agent-assisted work on the same terms as other delegated work.
- Confidentiality walls apply to agents as they do to people: an agent working on one client's matter should not draw on another's.
Firms with offices in the Gulf should also consider data rules there, including the UAE Personal Data Protection Law and training staff who use agents. Neighbouring professions face similar questions; see how accountancy firms prove staff are trained and how property management firms do it. The guide to AI literacy and certification sets out the wider approach.
Questions people ask
Do partners need certificates?
Partners who approve agent drafts directly should hold one. Partners who supervise need enough literacy to judge the review beneath them and to answer client questions.
Should we tell clients when agents work on their matters?
Some clients' guidelines require it, and it is good practice. The evidence chain above makes the answer easy to support.
Can trainees approve agent drafts?
Yes, if they hold a current certificate and their work is supervised as usual. Seniority matters less than tested competence on the task.