The short version
Use it to find and summarise, never to assert. Every output cites its source document and paragraph, because an unsourced legal statement is worthless to a solicitor and dangerous to the firm. Start with disclosure review and first drafts of standard documents. Settle privilege, client consent and the billing question before the pilot, not after.
What are the highest-value uses in a firm?
Three, in rough order of how quickly they pay:
- Disclosure and document review. Large volumes, relevance judgements against defined criteria, and a task that is currently expensive precisely because it is mechanical enough to delegate and consequential enough to check.
- First drafts of standard documents. The forty-first shareholders' agreement of the year, assembled from your own precedent bank rather than from general knowledge.
- Summarising a matter. Bringing someone up to speed on a bundle, a chronology, or three years of correspondence — work that is pure cost and gets billed apologetically.
What sits below the line is advice. Not because the technology cannot produce something advice-shaped, but because a firm’s product is a considered opinion for which it accepts responsibility, and that is precisely the thing you cannot delegate to a system.
Why does citation matter more here than anywhere else?
Because an assertion without a source is not usable output in a law firm. A solicitor cannot act on “the contract permits termination on thirty days’ notice”; they need to know it is clause 14.2 on page 31 so they can read it themselves.
This flips the design. The system’s job is to find and point, not to conclude. A tool that says “three documents appear relevant to this issue — here they are, with the passages highlighted” is enormously valuable and carries almost no risk. One that summarises the legal position in confident prose with no provenance has produced something nobody can rely on and everyone must re-do.
It also makes the failure mode safe. An incorrect citation is immediately obvious when clicked. An incorrect summary reads perfectly.
What about privilege and client confidentiality?
The question to answer before anything is connected is where client material is processed, who can reach it, and how long it is retained. Legal professional privilege is not automatically waived by using a supplier — firms have used external providers for decades — but the arrangement has to be deliberate, documented, and consistent with your retainers.
Three practical steps. Check whether your client engagement terms contemplate third-party processing and update the template if not. Establish whether the deployment can run inside your own environment rather than sending material outward. And confirm the position on whether anything you submit could be used to train a model, in writing, for the specific plan you are on.
Conflicts deserve a mention too. If a single deployment holds material from opposing parties in different matters, the information barriers your firm maintains between teams need to exist inside the system as well, and that is an access-control design decision rather than an afterthought.
What does this do to billing?
This is the question that stalls more legal deployments than any technical issue, and it is usually discussed too late.
If a disclosure exercise that billed two hundred hours now takes seventy, an hourly firm has just cut its own revenue on that matter. The technology worked and the P&L got worse. Firms that resolve this tend to do it by moving the affected work to fixed fees — which clients generally prefer anyway — and competing on turnaround rather than on the hours consumed.
The uncomfortable version is that a firm which does not resolve it will find a competitor quoting a fixed fee for the same exercise and winning the work regardless. The billing question is not avoidable, only deferrable.
What does a firm have to supply?
Four things no tool arrives with:
- Your precedent bank. A first draft assembled from your own approved documents is useful; one assembled from general knowledge is a liability.
- Your review threshold. What a fee earner must read personally, and what may be relied on after a spot check. This is a supervision question and it needs a partner's name on it.
- An audit trail. Which documents were reviewed, by what, and what a person looked at before signing off — because disclosure obligations can be tested afterwards.
- Training that is honest about limits. Fee earners need to know the failure modes, not just the interface.
How should a firm start?
On a closed matter. Take a disclosure exercise you have already completed, run the review again, and compare against the answer you know is correct. It costs a few days, risks no client, and produces a real accuracy figure rather than a vendor’s.
Then pick one practice area and one task rather than rolling out firm-wide. The rules, the precedents and the tolerance for error differ enormously between, say, a corporate team and a personal injury team, and a deployment tuned for both at once is tuned for neither.
Will this replace junior solicitors?
It removes a category of work that trainees have traditionally learned on, which is a real problem the profession has not solved. Document review is tedious and it is also how people learn what matters in a bundle.
The firms thinking about this properly are treating the saved time as capacity rather than headcount — juniors reviewing the system’s exceptions and its uncertain calls, which is more demanding than reading everything and probably better training. That is a deliberate choice, though. It does not happen by default, and a firm that simply takes the saving will find in three years that it has no mid-level lawyers.
What should we not use it for?
Anything that constitutes advice to a client without a solicitor’s judgement between the output and the recipient. That is the bright line, and it is worth writing into the firm’s policy in those terms rather than leaving it to individual discretion.
Two more that catch firms out. Do not use it to check whether a case or authority exists — verify citations against the actual source, always, because a plausible-looking but non-existent reference is the single most publicised failure mode in this profession and the reputational cost is severe. And be cautious with anything where the client’s own words matter precisely: a summary of a witness statement is a paraphrase, and paraphrase is where meaning shifts.
How do we handle supervision obligations?
Treat output as the work of an unqualified assistant, because that is functionally what it is. The supervision framework your firm already operates — who reviews what, to what standard, and what is recorded — extends to it without needing to be reinvented.
What does need writing down is the threshold. Which categories of output require a fee earner to read the source material personally, and which may be relied on after a spot check. A partner should own that document, and it should be specific enough that a junior knows which side of the line a given task falls on without having to ask.