No, not advice. An AI agent can give legal information: what a policy says, what a standard clause usually does, which deadline the organisation's playbook sets. Legal advice applies the law to someone's specific facts and recommends what to do. That belongs to a qualified lawyer, because people act on advice in ways that cannot be taken back.
Where is the line between information and advice?
Information is general and checkable. "Our standard terms cap liability at the fees paid in the previous twelve months" is information. So is "the playbook says we accept a limitation period of no less than one year".
Advice is specific and directive. "You should accept their liability cap because the risk on this deal is low" is advice. It weighs the facts, makes a judgement and tells someone what to do.
The test is reliance. If a reasonable person would act on the answer without checking it with a lawyer, treat it as advice. In many jurisdictions some legal activities are reserved to qualified, regulated people, and giving advice without that standing creates risk for the organisation that deployed the agent. The same line runs through medical advice and financial advice.
What can an in-house legal agent do?
| Task | The agent | Stays with a lawyer |
|---|---|---|
| Answering staff questions on internal policy | Answers from the approved policy, quoting it | Questions the policy does not cover |
| Reviewing contracts against a playbook | Marks each deviation and quotes the playbook's fallback position | Accepting any non-standard term |
| Summarising a contract | Produces a summary with clause references | Advising on what the summary means for a deal |
| Tracking deadlines and obligations | Extracts and diaries them | Deciding how to respond to a breach |
| Legal holds | Identifies custodians and sends standard notices | Deciding scope and release |
| Customer-facing explanations | Explains the organisation's terms in plain words | Any question about a customer's legal rights in a dispute |
Legal holds show the pattern well. The agent can do much of the administration, as described in handling legal holds with an AI agent, but deciding what the hold covers is a legal judgement.
Why does reversal matter for advice?
Wrong information can be corrected: you send a note saying the policy was misquoted. Wrong advice is acted on. Someone signs the contract, misses the limitation period, sends the termination notice or tells a regulator something they should not have. Those actions cannot be recalled.
Some outputs also carry legal weight on their own. A message telling a customer their complaint is rejected, or that a contract is terminated, can itself be a decision with legal effect. The agent drafts those; a person sends them. The general argument is in why autonomy in regulated industries is a question about reversal.
A worked example
Take a software company's commercial team. A salesperson uploads a customer's redline of the master services agreement and asks the legal agent whether they can sign.
The agent compares each change with the playbook. Nine changes fall inside approved positions. One deletes the cap on the customer's own liability for data misuse, and another moves governing law to a jurisdiction the playbook does not list. The agent marks those two, quotes the playbook's fallback language for the first, and states that the second has no approved position.
It does not answer "can we sign". It tells the salesperson which changes are within the playbook and that two need a lawyer, and it holds the review for the named commercial lawyer with the comparison attached. The lawyer decides.
What controls keep the agent on the right side?
- Approved sources only. The agent answers from the playbook, policies and precedents the legal team has approved, and records which version it used.
- A refusal pattern. Questions that ask "should I" or "can I" about a specific situation get information plus a route to a lawyer, never a recommendation.
- Certified refusals. Before the agent goes live, test it on real questions from the team's own history and check that it refuses the advice questions, not just that it answers the information ones.
- Named escalation. Each escalation goes to a specific lawyer or rota, with the question and the agent's working attached.
- Privilege considered. Whether privilege attaches to an agent's output is a question for counsel in your jurisdiction; design the workflow so a lawyer is involved where privilege matters.
Questions people ask
Can an agent tell a customer what their rights are under a contract?
It can explain what the contract says in plain words. If the customer is in a dispute with the organisation, questions about their rights go to a person, because the organisation is not a neutral adviser.
Can a law firm use an agent to answer client questions?
It can handle information and administration: matter updates, document requests, explaining next steps. Advice to a client comes from the responsible lawyer, who remains accountable for it.
What should the agent say when it refuses?
It should say what it can help with, why the question needs a lawyer, and who it has passed the question to. A refusal with no route forward is a poor outcome.