AI for lawyers: what is safe to use and where it saves time

What AI for lawyers can and cannot do in legal work

Treat the model as a clerk that is fast, tireless and untrustworthy until a person has checked it. It can read what you put in front of it, draft from that material, group papers, pull a list, and flag a gap such as a date with no owner. It cannot take responsibility for the file. It cannot decide take, refer or decline. It cannot confirm that a date is the deadline. It cannot run a conflict check as a legal conclusion. Those steps belong to a person in the firm.

The useful split is source-in-front-of-you versus judgement. If the email, exhibit or bundle is on the screen, a draft acknowledgement, a summary of that bundle, or a list of parties named in it can save time. If the next step is a professional decision, the model is not the owner of that step. A first draft is not advice. A suggested date is not the date. A list of names is not a conflict search that has been done.

We see the same failure when a tool is asked to work from memory of the internet instead of from the file. The result can look finished: complete sentences, a citation shape, a confident date. Finished is not the test. The test is whether a person can point at the page, the email or the exhibit that supports each claim. If they cannot, the text is not ready to leave the firm. Draft, group, summarise, flag. Then a person reads, matches, and signs.

Confidentiality and client data

Confidentiality is the first filter, not a setting you look at after the demo. Before a client file goes in, the firm should be able to say where the data goes, who can read it, how long it is kept, and whether it is used to train a model. If those answers are not in writing, do not put client material in the tool. A verbal assurance in a sales call is not a control.

Where the data goes means a real location, not a slide. Ask which country the files sit in, which subprocessors see them, and whether a prompt is logged. Ask whether the vendor's staff can open a document to debug a job. Ask whether a copy remains after you delete the matter. Retention should be a number you can state, with a rule for what happens when the engagement ends.

Training on inputs is a yes or no question, and the useful answer is specific. Does this product train on our prompts, our uploads, or our corrections. Is that on by default. Can it be turned off in the contract, not only in a dashboard that someone forgets to check. If the vendor trains on customer content and cannot offer a firm-level off switch in writing, treat the tool as unfit for confidential work.

Access inside the firm matters as much as the vendor. Who can paste a file into the tool. Who can see the output. Who can export a log. A shared personal account is not an access model. If anyone can drop a matter into a consumer chatbot, the policy is already leaking.

We do not list certifications we do not hold, and we do not ask you to trust a badge in place of answers. Lawyers using AI need a paper trail they can show: the vendor's written position on training, storage, staff access and deletion. If that paper is missing, the safe use is not to use it on a live file. For the questions to put to a vendor, see how to choose legal AI tools.

Checking output against sources and hallucination risk

Models invent names, dates and citations. They also invent quieter things: a clause that was not in the lease, a party that was not on the writ, a deadline that was not in the order. The fluent sentence is the hazard. A hallucination check is not a product feature you switch on once. It is a matching step in the process: each claim in the output is true only when a person has matched it to a page in the source.

If the tool can point at the source page, the person still reads that page. A citation mark is a hint, not a verdict. If the tool cannot point at a source, the person has to find the source before anyone relies on the text. That takes time. It is still cheaper than sending a draft that does not stand up. AI for lawyers is not a shortcut around that check.

Write the check so it does not depend on memory. The reviewer should see the output and the source together, tick that each material claim is matched, and record who did the check. An unmatched claim is treated as untrue for the purpose of the file, even if it sounds likely. Likely is not a source.

Dates need a harder rule than prose. A number that looks like a deadline is not a deadline until a qualified person in the firm confirms it against the court rules and the papers. We do not calculate legal deadlines, and a model should not be allowed to either. The process on deadlines and court dates is the same whether the suggestion came from a colleague or from a tool: a person confirms every computed date.

Human review and sign-off

A person reviews every output before it leaves the firm. That is not a preference. It is the difference between a draft and a communication the firm is responsible for. The model does not file, does not advise and does not replace counsel. Sign-off is a named person, on a named version, at a named time.

Review has to be forced by the workflow, not hoped for. A toggle that nobody uses is not a control. The path from generate to send should stop until someone with the file has marked the output as reviewed. If the tool emails, files or posts without that stop, it is not safe for client work, however good the demo looked.

The reviewer needs the source, the output, and a place to record the decision: that the draft matches the papers, unmatched claims were cut, and a date was confirmed. Signing that the tool is usually accurate is not a review. PredictaOps is not a law firm. We do not give legal advice, file in court or replace counsel.

Where AI saves lawyers time: intake, deadlines and documents

Time is saved where the work is repetitive and the source is in front of you. Time is not saved where a person has to decide. That split is the whole of the useful map. Below it sits in the three processes we keep seeing in small firms: how an enquiry becomes a file, how a date gets an owner, and how a document leaves the building.

On client intake, AI can turn a web form, an email or a voicemail note into a structured enquiry: who wrote in, how to reach them, what they asked, and a first-draft acknowledgement for a person to edit. It can put that enquiry on one list instead of leaving it in an inbox. It cannot name the owner. It cannot decide that a conflict check is done. It cannot take the file. The process still needs capture, owner, conflict check, decision, open file, next contact. The tool sits inside those steps. It does not replace them.

On deadlines and court dates, AI can read an order or a letter and list the dates it can see, each with a pointer back to the line it came from. It can flag a date that has no owner, or a reminder that never went out. It cannot calculate the deadline as advice. It cannot confirm the date. A qualified person does that, against the rules and the papers, and the calendar records who confirmed it. Reminders then run off that confirmed date, not off a guess.

On documents, AI can summarise a bundle you uploaded, draft from the pages in that bundle, and group papers by type or date. Every claim in the draft still has to be checked against its source page. A person reviews before the document leaves the firm. That is the whole of the gain: less time spent on the first pass, no less care on the last pass.

Legal operations is the name for keeping those three processes visible when nobody has legal ops in their job title. legal operations is the process page for firms that need the function without a dedicated team. The commercial path, a system built around one bottleneck, is AI for law firms.

Further reading: the Law Society's Generative AI: the essentials guidance.

How to start: one process, one agreed number

Do not start with a transformation programme, and do not start by giving every person in the firm a new chatbot. Start with one process that is already costing you time: intake that sits in an inbox, dates that live in one person's head, or documents that are drafted without a source check. Name an owner for that process. Agree one number you will look at every week, such as time to first response, dates with no owner, or drafts that went out without a recorded review.

Then put the tool at one depth only. If the process is intake, the first job is a single list and a first-draft acknowledgement, not an automated decision to take the matter. If the process is dates, the first job is a list of dates-from-papers for a person to confirm, not a calendaring engine that files reminders on unconfirmed numbers. If the process is documents, the first job is a draft from the bundle with source pointers, not unsupervised send.

We design and build legal AI systems for law firms. Guhan MS, Founder & Legal Engineer, PredictaOps, holds an LLB (Hons) from Sheffield Hallam University and is trained in English law. We are not a law firm, and we do not pretend the tool is counsel. The method is the same one we use on paid work: Look, Build, Stay. Look at how the work actually moves. Build a thin slice around the bottleneck. Stay with it until the agreed number is visible every week.

If you want that commercial path, it is AI for law firms. If you want the vendor questions before a file goes into a product, that is how to choose legal AI tools. Either way, pick one process. One number. A person on every output.