ChamberOne

AI agents in law firms: what they can and cannot do today

By the ChamberOne team 4 minute read

The word agent now appears on nearly every legal technology website. It can mean a simple chatbot with a new name, or a system that plans several steps, uses tools and carries out work with little supervision. Partners and practice managers need a way to tell the difference, because the risks and the controls are different. This article gives a plain definition, a sober view of what current systems do, and a list of controls to ask for before one touches a client matter.

Key points

  • An agent chooses some of its own steps, which is why it needs tighter controls than a chat assistant.
  • Bounded, checkable jobs suit current systems; long unsupervised chains do not.
  • Require human approval for anything outward-facing or irreversible, shown with the content to be approved.
  • Insist on a log that records instructions, models, material read and who accepted the output, and that users cannot switch off.

What an agent is, in plain words

A chat assistant answers one prompt at a time. An agent is given a goal and works towards it in steps. It may search a database, read a document, draft a paragraph, call another program and then decide what to do next. The defining feature is that the system, not the person, chooses some of the intermediate steps.

That freedom is both the appeal and the risk. A chat answer can be read and rejected before anything happens. An agent that has permission to send an email or change a record can act before anyone reads its reasoning.

What current systems do reasonably well

Agent-style tools are most useful on bounded jobs with a clear finish line and a checkable result. They tend to work best when the steps are similar each time.

  • Gathering: pulling the documents on a matter that mention a named party, and listing them with page references.
  • Drafting from a template: filling a standard letter or schedule from matter details and flagging the fields it could not fill.
  • Triage: sorting incoming enquiries or emails by type and proposing where each should go.
  • Preparation: building a first chronology from the dates in a set of documents.
  • Suggesting follow-up tasks after a document or meeting, for a person to accept or reject.

What they cannot be trusted to do alone

Errors compound across steps. If each step has a small chance of going wrong, a chain of ten steps has a much larger one, and the final output may look polished while resting on a bad early step. That is why unsupervised long chains are a poor fit for work with legal consequences.

Agents also lack what lawyers supply without noticing: knowledge of the client's real objectives, judgment about what to leave out, awareness of an unspoken conflict, and accountability to a regulator. They do not know that a letter will be read by an angry opponent. They can be steered by text in the material they read, a risk covered in another article in this series. And they do not file with courts or give advice that a lawyer can adopt without reading.

Suits current systems
  • Bounded jobs with a clear finish line
  • Results that are easy to check
  • Steps that are similar each time
Needs a person
  • Long unsupervised chains of steps
  • Knowing the client's real objectives
  • Accountability to a regulator
Where agent-style tools fit

Cutting through the marketing

When a vendor says its agent handles a task end to end, ask which steps the system performs, which steps a person performs, and what happens when the system is unsure. A demonstration on a clean example proves little. Ask to see a failure and how it is surfaced.

  • Can I see every step the agent took on a real run, in order?
  • What actions can it take that change something, such as sending, deleting, filing or billing?
  • Which of those need a person's approval, and can I change that setting?
  • What data does it read, and does that include matters it has no need to see?
  • What does it do when it cannot complete a step: stop, guess or carry on?

Where approvals belong

A workable rule is that anything outward-facing or irreversible needs a human decision. Sending a document to a client or another party, recording time against a client, issuing an invoice, closing a matter and deleting a record all fall in that group. Internal drafts and suggestions can be looser, as long as they are labelled as drafts and nobody can mistake them for approved work.

Put the approval where the lawyer can see what they are approving. A button that says 'approve all' without showing the content is theatre. The reviewer should see the output, the source material and any uncertainty the system flagged.

Logs: the record you will want later

If an agent touches a client matter, you need to be able to reconstruct what it did. At a minimum, a log should show who started the task, what the instruction was, which model or provider was used, what material was read, what output was produced and who accepted it. It should be kept for as long as the firm keeps the matter file, and it should be searchable when a client, insurer or regulator asks a question.

Ask whether users can switch logging off. A log that its subject can disable is not much of a safeguard.

Questions for the partners' meeting

Before any agent-style tool is switched on, a short internal discussion saves arguments later. Who in the firm is allowed to start an automated task on a client matter? Who reviews the results, and how quickly? What is the firm's position if a client forbids AI use on its matters, and how will the tool respect that? Who owns the decision to switch a feature off after an incident? Writing down the answers, even in half a page, turns a vague enthusiasm into a policy that a new joiner can follow.

A sensible way to start

Begin with one narrow job, such as drafting the first version of a routine letter or sorting enquiries. Run it alongside the existing process for a few weeks and compare results. Count the corrections reviewers make. Expand only when the corrections are few and the log shows nothing surprising. If a tool will not let you limit what it can do, that is a reason to wait.

How ChamberOne handles AI output

  • Drafts the matter brief, analyses documents and suggests tasks when you ask for them.
  • Keeps each output as a draft until someone accepts it.
  • Records every AI call in the audit log with the model and token counts.
  • Lets you choose the model for each AI feature, GPT-5 or Claude, or switch AI features off for the firm.

What it does not do

  • It does not give legal advice, and AI output can be wrong or incomplete.
  • It does not replace a lawyer's review.
  • It does not file documents with courts.

This article is general information about the law, legal practice and technology, current when it was published. It is not legal advice. Laws and guidance change: check the sources linked above and take advice on your own circumstances from a lawyer qualified where you are.

Read next

AI agents in law firms: what they can and cannot do