ChatGPT for Solicitors: Five Safe Tasks and Three to Avoid

A solicitor can safely give ChatGPT drafting and admin tasks that use no client detail or that are checked line by line: plain-English rewrites, meeting-note structure, client-update drafts, checklists and first-pass summaries of documents already cleared for sharing. Keep legal research, privileged material and anything filed or relied on without independent verification off it.

The dividing line is not how clever the task is, but what happens when the tool is wrong or the data leaks. Safe tasks have a human who checks every word and a text that holds nothing confidential. The three to avoid fail one of those two tests. Below, each task comes with a prompt you can copy, a realistic output and the check a solicitor still has to make.

Follow me on Instagram@sagnikteaches

Before any task: three settings and one habit

Set the ground rules first, because most problems come from the account, not the prompt.

Connect on LinkedInSagnik Bhattacharya

1. Choose the plan on purpose. On personal plans, conversations can be used to improve the model unless you switch that off under Data controls. ChatGPT Business does not train on your business data by default; OpenAI's business pricing page lists a Standard seat at $20 a user a month billed annually or $25 monthly, for teams of 2 to 200, so confirm the current figure there. A firm handling client matters should use a business workspace, not staff members' personal accounts.
2. Check memory and history. Depending on the account, ChatGPT's memory controls include Memory, Reference saved memories and Reference chat history, so look at what your plan stores between chats and switch off what you do not need. Use a temporary chat for anything sensitive; OpenAI says temporary chats are not used to improve models, though a copy may be kept for up to 30 days for safety purposes.
3. Use a Project for shared instructions. A Project holds your house style and checklist so every draft starts from the same rules. OpenAI's help centre says project sharing works on all plans. Do not build on Custom GPTs: OpenAI's retirement FAQ says they are scheduled to retire on 11 December 2026, with creators moved to plugins.
4. The habit: replace names with roles. Write [client], [counterparty], [property] and [date]. A rule this simple removes most of the confidentiality exposure from everyday drafting, and it costs seconds.

None of this replaces your regulator's guidance, your professional indemnity insurer's terms or your client engagement letter. Read them, and record the decision on which tool, which plan and which tasks. The firm-wide questions are covered in can solicitors use ChatGPT without breaching confidentiality.

Subscribe on YouTube@sagnikteaches

Safe task one: rewrite your own draft in plain English

Clients stop reading at the second sentence of dense legal prose. Rewriting your own paragraph for a lay reader is a strong first use, because you wrote the substance and can see instantly whether the meaning survived.

Rewrite this paragraph for a client with no legal background.
Reading age about 14. Keep every condition and deadline exactly as
written. Do not add advice. Mark in [square brackets] anything you
changed that might alter the meaning.

[paste your paragraph with names replaced by roles]

Illustrative before: "Completion is conditional upon the satisfaction of the buyer's lender's requirements as to title." Illustrative after: "The sale can only complete once the buyer's lender is happy with the paperwork showing who owns the property." The useful bit is the brackets: if the tool softens "conditional", it should say so. Your check is to read the plain version against the original and confirm no condition vanished.

Safe task two: turn rough attendance notes into a file note

Take your own bullet points from a call and ask for structure. Because you supply the content, nothing new is invented, and because names are roles, little is exposed.

Turn these bullets into a file note with headings: Attendees (roles
only), Instructions received, Advice given, Actions with owner and
date, Open questions. Use only what is below. If something is
unclear, list it under Open questions instead of guessing.

[bullets]

Realistic output puts "client to send bank statements by Friday" under Actions with an owner, and moves "unclear whether the deposit came from savings or a gift" to Open questions, where it belongs. The check: confirm the dates and amounts match your notes. A model sometimes converts "end of next month" into a specific date, which is a guess. Search the draft for any date you did not write.

Safe task three: draft a client update from facts you provide

Status updates are repetitive and time-consuming. Give the tool the facts, the tone and the next step, and it returns a draft in a minute. See drafting a client care letter with AI for a related, more regulated document.

Draft an email update to a client. Facts: searches have been
returned; one query raised with the seller's solicitor about a
boundary; expected response in 5 working days; no action needed
from the client this week. Tone: reassuring, plain, under 150 words.
Do not predict completion dates. Sign as [your name].

The line to enforce is "do not predict". Left alone, the model fills the silence with comfort such as "this should be resolved quickly". A client who relies on that has an expectation you did not set. Strike any reassurance you cannot stand behind.

Safe task four: build a checklist or timeline template

Checklists contain no client data at all, which makes them the safest task of the five. Ask for a standard-stage checklist for a matter type, then edit against your own file procedures.

Create a stage-by-stage checklist for a residential sale from
instruction to completion, for a small firm. For each stage list the
documents to request, the usual client questions and the common
causes of delay. Present as a table. This is a template for our
internal use; it will be reviewed by a fee earner.

Output is typically a sensible 20-row table. It will miss your firm's own quirks and may reflect general practice that differs from your area, so treat it as a first draft and mark each row "keep", "change" or "delete" in review. Treat anything that reads like a statement of law or a statutory time limit as unverified until you have confirmed it in a primary source.

Safe task five: summarise a document you may already share

A first-pass summary of a long public or non-confidential document, such as a published guidance note or a template contract with all parties removed, saves reading time. The test is whether you would be comfortable emailing the text to an outside contractor.

Summarise this guidance note in 200 words for a busy partner.
Then list the five points most likely to affect a small firm's
conveyancing procedures, quoting the sentence each comes from.
If a point is not clearly stated in the text, say so.

Requiring quotations is the safeguard: each point is traceable, and a point with no quote is suspect. Read the original for anything you will act on. A summary is a map for deciding what to read, not a substitute for reading it.

Avoid one: legal research and citations from memory

This is the clearest hazard. A model can produce a citation that looks perfect and does not exist. Professional-regulator warnings and court sanctions over fabricated authorities are well documented, and a 2026 tax regulator alert on AI cited such cases as a warning to practitioners in another profession. The risk is structural: the tool predicts plausible text, not truth.

If a task needs authority, use a legal research service that cites to real sources, then read the source yourself. If you do ask ChatGPT for a starting list of topics, do not copy any case name or section number into a document until you have found it in a trusted database. Anthropic's own legal plugin pack, for comparison, marks citations that come from model knowledge alone with a "verify" tag, which tells you how seriously the vendors take it; see which plugins a small law firm should turn on.

Avoid two: privileged client material on a personal plan

Do not paste witness statements, medical details, financial records, settlement figures or full names into a personal ChatGPT account. Switching off "improve the model for everyone" lowers one risk, but the data still sits in an account the firm does not control, outside your retention and deletion rules. On ChatGPT Business, OpenAI's help centre says a removed member's chats and files are retained indefinitely and are not visible to the workspace owner, which matters if a fee earner leaves with matter detail in their history.

The practical rule: if you would hesitate to attach it to an email to a stranger, it stays out. For work that genuinely needs the full document, use a tool the firm has assessed, with a contract covering confidentiality, retention and sub-processors. The checks are in the supplier checklist for solicitors.

Avoid three: anything filed or sent to the other side unchecked

Pleadings, witness statements, correspondence with an opponent, advice letters and anything signed off with your name should not carry unchecked AI text. The issue is accountability: a drafting error in a letter to the other side is yours, and a subtle shift in meaning in an advice letter can create a negligence claim. AI can help you improve your own draft, but the thinking and the final wording must be yours.

A workable boundary: ChatGPT may restructure or polish what you wrote, and you must read the polished version against your original word by word. It may not originate the advice, the legal argument or the facts. Record who reviewed the final text.

What a failure looks like: a realistic mistake

Suppose a fee earner asks for a plain-English rewrite of a paragraph about a break clause. The original said the tenant must give written notice "not less than six months and not more than twelve months before the break date". The rewrite reads "the tenant must tell the landlord at least six months before leaving". The twelve-month window has silently gone, and a client who follows the rewrite could serve notice too early and see it rejected.

It looks small and reads well, which is the danger. The defence is the instruction to mark changes in brackets, and a reviewer habit of checking every number, date and "not more than" against the source. Build a final pass into the routine: highlight every figure, time limit and condition in the original, then confirm each appears unchanged in the output. It takes two minutes and catches the failure that matters most.

Reading the plan terms before the first matter

Before anyone uses the tool on client work, someone in the firm should read three things from the vendor: the data-use terms for your plan, the retention rules for chats and files, and the admin controls over who can connect which apps. OpenAI now calls connected services apps (formerly connectors), managed under Settings, then Plugins, with permission options that run from "Always ask" to "Allow all actions". For a law firm, leave connected apps off at first, and enable one only after deciding exactly which folders it can see.

Write down the answers in a short note: which plan, which settings, who administers it, how a leaver's access is closed, and what happens to their chat history. If you cannot answer those questions from the vendor's own pages, ask the vendor in writing, and wait for a written reply. Anthropic's repository for its legal plugins makes a related point from the other side: every output is a draft for lawyer review, and the lawyer, not the plugin or the vendor, is responsible for the positions taken. The checking is on you whichever vendor you choose.

Polishing your own advice letter without handing over the advice

A sixth use sits close to the boundary, so treat it carefully. You wrote an advice letter and want it tighter. Ask for edits and nothing else.

Edit this letter for clarity and concision only. Do not change any
legal statement, figure, date or recommendation. Show every change as
a numbered list: the original phrase, your replacement, and why.
If a sentence is ambiguous, flag it instead of rewriting it.

[your letter, names replaced by roles]

A realistic result lists perhaps twelve changes, mostly cutting filler, with two flagged ambiguities such as "the property" appearing where two properties are discussed. You accept the changes you agree with, resolve the flags yourself and keep your original as the record. The tool has not given advice, only helped you say your own advice clearly.

A worked sample: a four-person firm triages its week

Take an illustrative four-person firm with two fee earners, a paralegal and an office manager. On Monday the paralegal drafts three client updates from facts she supplies, and the fee earner reads each against the file. On Tuesday a fee earner rewrites a dense paragraph in a settlement letter for a client, then compares it with the original. On Wednesday the office manager creates a checklist for a new matter type. On Thursday a partner asks the tool for a summary of a published guidance note.

The same week, three requests come up that the firm declines: a fee earner wants to paste a witness statement to "tidy the English", someone suggests asking for case law on a limitation point, and a trainee wants the tool to draft a letter before claim. The rule for each: tidy the statement yourself, use a research database for authority, and draft the letter then use the tool for polish only. After a month the firm logs which tasks were safe, and adds them to a one-page policy.

For a session you can run with the whole team, the one-hour AI training plan for solicitors' firms gives an agenda and exercises, and ChatGPT or a legal AI tool helps decide when to buy something specialised. The wider hub is the AI guide.

Applying the same tests in a different practice area

The five and three are not specific to property work. Take an illustrative family-law practice with two fee earners. A plain-English rewrite of a paragraph about a financial disclosure deadline passes the test, since the fee earner wrote the paragraph and checks the dates. A checklist of documents to request from a client before a first meeting passes too, because it holds no client data. A first-pass summary of a published guidance note on court procedure is fine when the note is public.

What fails is the same list as before: asking the tool what the courts usually award on a particular fact pattern, pasting a client's bank statements to spot undisclosed assets, and sending a drafted position letter to the other side without a full read. In an employment practice the safe tasks are meeting-note structure and update emails, and the avoid list adds pasting a grievance file with named colleagues. Ask two questions of any new task: what happens if the output is wrong, and what leaves the building if I paste this? If either answer worries you, it belongs on the avoid list.

A short policy you can adopt this week

• Business workspace only; no personal accounts for client work.
• Names, addresses, dates of birth, account numbers and case numbers replaced by roles before pasting.
• Permitted: plain-English rewrites, file-note structure, update drafts, checklists, summaries of public documents.
• Not permitted: legal research from memory, privileged material in unassessed tools, unchecked filed or sent text.
• Every AI-assisted document has a named reviewer who has read it against the source.
• A quarterly review of settings, plan terms and the list of approved tasks.

Adjust it to your regulator's guidance and insurer's terms, and take advice where unsure. The five safe tasks are worth the effort, because they return time on work that never needed a lawyer's judgement. The three to avoid are worth refusing every time, because the cost of one mistake is far larger than the minutes saved.

How to test a task before you adopt it

Pick one safe task and trial it on five real, anonymised examples. Time each one with and without the tool, and have a second person mark errors found at review. If the tool saves time and review finds only small edits, adopt the task and write the prompt into the Project. If review finds a changed meaning even once, tighten the prompt, for instance by adding "mark anything you changed in brackets", and retest.

Keep a simple log with the date, task type, plan used, reviewer and any error caught. After a quarter it shows which tasks genuinely pay off, and it gives you evidence to show your insurer or regulator that use is supervised. Where a task keeps producing errors, drop it from the approved list without fuss.

One last caution on conveyancing-style templates and time limits: the tool tends to state deadlines and fees confidently. Any number in a checklist, such as a notice period or a time limit, should be confirmed against a primary source before it goes in a template your colleagues will trust, because a wrong figure in a template repeats on every matter.

Finally, include clients in the decision where it matters. Some will ask whether you use AI, and a short, honest answer on your website or engagement letter, saying that AI tools may assist with drafting and are always supervised and checked by a solicitor, saves awkward questions later. If a client objects, respect it and note it on the file.

Questions solicitors ask about ChatGPT

Can a solicitor put client names into ChatGPT?

Avoid it. Most drafting tasks work with names replaced by roles such as [client] and [other party]. If you need real matter detail, use a business or enterprise workspace whose terms you have read, check what your regulator and your insurer expect, and record the decision in your file.

Does ChatGPT remember what I type?

Conversations are stored in your account unless you use a temporary chat, and memory features can carry details between chats. Check the data controls and memory settings on your plan, turn off anything you do not want, and do not rely on deletion as your only safeguard.

Is a business plan enough to be safe?

It removes one risk, because business plans do not train on your content by default. It does not stop the tool getting a point of law wrong, inventing a case, or exposing data through a badly scoped connected app. Supervision and checking still apply.

Who is responsible for an error in AI-assisted work?

The solicitor. A tool does not share your professional duties, so any output that reaches a client or a court is yours. Keep a note of what was drafted with AI and who checked it.

Further reads

• AI Time Capture for Solicitors: Recover Hours You Never Billed — Recover unbilled time with careful AI use.
• AI Readiness Checklist for Accountants, Solicitors, Consultants — Check your firm is ready before widening use.
• Law Firms AI UK: What Solicitors Should Check Before Using AI — Pre-use checks for a small law firm.
• Best AI for Solicitors in the UK: Matched to Each Task — Which AI suits each task a solicitor does.
• Per-Seat or Pay-As-You-Go? Legal AI Pricing for Small Firms — Per-seat or pay-as-you-go pricing for small firms.
• Law Firm AI Usage: How to Audit What Your Fee Earners Really Use — A checklist to find out which AI tools fee earners really use: survey, browser and billing checks, then an approved list.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI business pricing page; OpenAI help centre articles on Projects, data controls, memory, temporary chats, connected apps, member removal and the custom GPT retirement FAQ; Anthropic's claude-for-legal repository notes on lawyer review; regulator guidance on confidentiality, which readers should check for their own jurisdiction.