Get the supervising lawyer's written permission and keep client data out of tools that train on it, then use AI on three jobs: first drafts from the firm's own precedents, summaries that cite the page, and billing narratives written from your real time notes. Verify every fact and citation yourself, and tell the lawyer which tool touched which task.
The lawyer you work for answers for your output, so their rules become yours before any tool does. Courts have sanctioned lawyers over citations an AI tool invented, which is why a paralegal who verifies is worth more with AI than without it. Rules on what a non-lawyer may do differ by place, and everything here assumes you work for lawyers rather than for the public directly.
Whose rules apply to you: the supervising lawyer's
A 2024 ethics opinion on generative AI from a national bar association is the best-known guidance, and bar bodies in several places have published their own. It is written for lawyers, but its duties flow straight down to anyone doing their work. Its text says lawyers must:
- give AI output used in client work an appropriate degree of independent verification or review;
- assess who could access client information before entering it into a tool, and obtain informed client consent for tools that learn from what they are given, with boilerplate in an engagement letter not being enough;
- set clear AI-use policies for lawyers, non-lawyers and staff, train them and supervise their use;
- bill hourly work at the time actually spent, which includes entering information and reviewing the output, never charge for learning a tool they will use regularly, treat a tool that works like ordinary office equipment as overhead unless the client agreed otherwise in advance, and keep flat fees reasonable when a tool makes the work much quicker.
Read that as the standard the lawyer is held to and you are working under. The exact rules depend on where the lawyer practises, so ask which guidance applies to them rather than assuming. If you serve the public directly where that is permitted, the confidentiality and scope duties are yours, so check your professional body's rules and your insurance terms.
Before the first AI-assisted job, send each supervising lawyer a short permission note with five questions, and keep the reply with your engagement papers:
- Which AI tools, if any, may I use on your matters?
- Which matters or document types are off limits?
- May I enter party names, or must I use placeholders?
- What do you want me to report each time AI touched a task?
- Who checks citations before a document leaves the office, me or you?
The answers become your operating rules for that lawyer, and they will differ. One may allow placeholders in a business-plan tool, another may allow only the firm's own system, and a third may say no AI at all. The note also gives them a chance to say what they have told their clients, which is their duty to decide and yours to respect. The wider view of dividing work between a firm and its AI is in how to divide the work between paralegals and AI.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowThree tiers of material, and where each may go
Confidentiality is where a freelance paralegal's AI use most often goes wrong, usually through speed and not malice. A three-tier rule keeps the decision to a few seconds per document.
| Tier | Material | Where it may go | Example |
|---|---|---|---|
| 1 | Generic or public | Any tool you are permitted to use | A blank precedent letter, a public rule's text |
| 2 | Matter facts with names and identifiers replaced by placeholders | A business-tier assistant that does not train on inputs, and only if the lawyer's note allows it | A chronology written as "Claimant A" and "Company B" |
| 3 | Privileged, identifiable or sensitive material | Only a tool the firm has approved for that matter, inside the firm's system, or no AI at all | Unredacted medical or financial records |
Tier 2 needs a placeholder key. Replace each real name with a code using find and replace, keep the key in a separate file, and put the names back only after the AI's output has been checked. Placeholders are not anonymisation, though. A rare combination of dates, places and amounts can still identify a client, so the lawyer decides what counts as Tier 2 for each matter.
Which tool may take which tier
The tier says how careful to be, and the tool decides whether being careful there is enough. Business plans don't train on your content by default, individual plans carry a model-training switch in their privacy settings, and a firm's own tools sit under the firm's contract and settings. This table maps each kind of tool to the tiers it can sensibly take.
| Type | Fits | Cost and terms shown |
|---|---|---|
| Business-tier general assistant | Tier 1, and Tier 2 where the lawyer's note allows | ChatGPT Business and Claude Team: $20 a user a month billed annually, $25 monthly, two seats minimum, so $40 to $50 for one person. Microsoft 365 Copilot Business: $21 annually, $25.20 monthly. None trains on business content by default |
| Individual plan | Tier 1 only, unless the lawyer allows more and the training switch is off | ChatGPT Plus $20 a month; Claude Pro $17 billed annually or $20 monthly |
| The firm's own assistant | Any tier the firm approves, inside its own system | Licensed and priced by the firm; practice-management systems such as Clio now build AI into intake, drafting and bill preparation |
| Legal-specific products | Tier 3 work only through a seat and contract the firm holds | CoCounsel Legal: no price on the product page; its page says data is never used to train underlying models and describes review of up to 10,000 documents against 100 questions with answers traceable to source. Spellbook: a Word add-in for review and drafting, zero-data-retention agreements that prevent use for training, priced by team members on a licence, 7-day free trial |
| Your own spreadsheet | Not a tier question | $0; keeps time sums, invoice totals and the citation log under your control |
Most freelancers need only the first row and the last, and many start with neither, using only what each lawyer provides. The risk sits in the data and the checking, not in the product name. If a firm offers you a seat on a legal product, what a small firm should test in paralegal software before paying gives the questions to ask.
Workflow one: a first draft from the firm's own precedent
The safest drafting job is adapting a precedent the lawyer already trusts, because the AI is rearranging approved wording instead of inventing it. Take a routine letter asking the other side's representative for documents. You have the firm's precedent and a facts sheet, with placeholders where the lawyer's note requires them.
You are helping a paralegal prepare a first draft for a supervising lawyer's review.
Use ONLY the attached precedent letter and the facts sheet below.
Task: adapt the precedent to the facts. Keep the precedent's structure and
wording wherever the facts allow.
Rules:
- Do not add legal conclusions, statute names, case names or deadlines.
- Where a fact is missing, write [CONFIRM: what is needed] instead of guessing.
- Do not change names, dates or amounts from the facts sheet.
- Under the draft, list every sentence you changed or added that was not in
the precedent.The output has two parts worth reading. The draft will carry markers such as "[CONFIRM: response date agreed with supervising lawyer]". And the list of changes might include, for instance, "Added: 'pursuant to the applicable rules'". That second one is the catch. The precedent had no such phrase, and it is a legal reference that nobody has checked. You delete it, or take it to the lawyer, and do not let it sit in the letter because it sounds right.
Do not rely on the model's own list alone. Run your word processor's compare function between the precedent and the draft, so every difference is visible as a redline, and clear each change against the facts sheet. Anything that is neither a fact nor the precedent's own wording goes. On a letter of this kind, imagine 45 minutes by hand becoming 15 minutes to prepare the facts sheet and run the prompt, plus 15 for the comparison and checks. That saves 15 minutes a letter, or three hours a month at 12 letters, and the lawyer receives a draft whose changes are already marked. How small law firms use AI to draft letters and routine documents shows the same idea from the firm's side.
Workflow two: summaries that cite the page
Summaries of bundles, transcripts and record sets are the job where AI saves the most time and invents the most. A summary that sounds right and cites the wrong page is worse than none, because the lawyer will trust it. So the prompt demands evidence in the shape the lawyer already uses, and you treat every citation as unverified until you have opened the page.
Summarise the attached transcript for a supervising lawyer.
Output a table with these columns: event, who said or did it (one person per
row), what was said or done, page and line.
Rules:
- Every row needs a page and line. If you cannot give one, leave the row out.
- One speaker per row. Never combine two witnesses in one row.
- Put any statement that could matter to a dispute in quotation marks,
exactly as written.
- Do not infer motives and do not draw conclusions.
After the table, list: pages you could not read, topics the transcript
mentions but does not explain, and any inconsistencies within the transcript.This is the shape of the output, from an invented transcript of a witness hearing:
| Event | Who | What | Page and line |
|---|---|---|---|
| Delivery of the goods | Witness A | "The pallets arrived on the Tuesday, not the Monday" | 41:12-14 |
| Call to the supplier | Witness B | Says they phoned the supplier twice before the delivery date | 58:3-9 |
| Inspection | Witness A | Says the inspection took place with the supplier's driver present | 63:20-24 |
Now the verification, which is the real work. Sample a third of the rows first. In an illustrative run the table has 18 rows, and checking 6 finds two wrong page references. That is too many to sample, so you check all 18. The full check turns up four wrong page references in total (22%), one wrong date and one row that merged two witnesses' statements into one. Twelve rows were clean. A workable rule is to sample a third and, if more than one row in six is wrong, to check every row. Record what you did, and tell the lawyer plainly: "All 18 rows checked against the transcript; six corrected."
Two further habits help. Ask for the list of pages the tool could not read, and open those pages yourself, because a scan with no text layer can fail quietly. And never ask a general assistant for case law. Research belongs in the firm's research service, and if the lawyer asks you to use an AI research tool, you pull up and read every authority it names. The cost of skipping that step is public. One researcher's database of court decisions dealing with AI-invented material listed more than 2,200 cases by early November 2026, and the outcomes it records run from warnings and struck filings to costs orders, fines and referrals to professional regulators. The longer routine for bundles is in summarising bundles and transcripts with AI safely.
Workflow three: billing narratives that match the time you worked
Your time is the one thing the AI must never touch. It can turn rough notes into narratives in the lawyer's style, but durations come from your timer or notes, rounding happens in a spreadsheet formula, and the total is checked against the raw minutes. Here is an illustrative morning, with the lawyer's convention being tenths of an hour at $55 an hour:
| Raw note | Minutes | Narrative drafted from the note | Units |
|---|---|---|---|
| call w/ lawyer re missing docs | 18 | Telephone attendance on supervising lawyer regarding outstanding documents | 0.3 |
| letter to other side asking docs | 40 | Drafting letter to other side's representative requesting documents | 0.7 |
| read bundle 2 for dates | 50 | Reviewing document bundle 2 to identify key dates | 0.8 |
| file note | 12 | Preparing file note of instructions and next steps | 0.2 |
| Total | 120 | 2.0 |
The check is simple. 120 minutes is 2.0 hours, and the four rounded entries add to 2.0, so the invoice line is 2.0 hours at $55, which is $110. If the rounded entries ever differ from the raw total by more than 0.1, find out why before the invoice goes out. The prompt that produces the narratives should say: use only the verbs and objects in my notes, never add a task, never change a duration. Without that rule a note saying "read bundle 2" can come back as "Reviewing and analysing the supply contract", which describes work that didn't happen and inflates the entry.
The billing rules come from the same guidance that governs the lawyer. On hourly work, bill the time actually worked, and that includes the time you spend checking AI output. Don't bill for learning a tool. Treat your subscription as overhead unless the lawyer has agreed otherwise. A fixed fee per deliverable, such as $60 for a first draft of a standard letter, avoids arguments about hours, but set it for the work as you now do it. The same guidance warns lawyers that charging an unchanged flat fee when AI makes a task much quicker may be unreasonable, and the lawyer will apply that test to your charge before passing it on. Because the lawyer may pass your charge on to their own client, tell them you use AI in the work and agree the fee knowing that. For the lawyer's side of time recording, AI time capture for solicitors covers the same ground.
Before each invoice goes out, run a short check: totals recomputed in the sheet, one narrative per task with no blocks of mixed work, the rate matching your agreement, narratives specific enough that the lawyer could explain them to their client, matter numbers correct, and disbursements listed apart from fees. If drafting the narratives cuts three hours of month-end admin to one and a half, you save 90 minutes, and you keep it, because that time was never billable.
Telling the lawyer: a filled-in reply and wording for your terms
A permission note is only useful if the reply is specific. This is an invented reply from one supervising lawyer, and it is the kind of answer to aim for:
Tools: a business-tier assistant only, never a free account. Names: placeholders required, key kept separately. Off limits: family matters and anything involving medical records. Reporting: put the tool and the task in one line at the top of each deliverable. Citations: you pull every authority from our research service and I check the key ones before filing.
Then put your side in your terms of engagement, in plain words the lawyer can forward if a client asks:
I may use AI-assisted tools to produce first drafts and summaries. I use only business-tier tools that do not train on my inputs, and I do not enter privileged material into any tool unless you have approved that tool for that matter. I review and verify all output before it reaches you, I will tell you which tools I use and for which tasks, and I will stop using any tool you object to. Time is billed at actual time worked.
Each deliverable then carries a one-line report, for example: "AI-assisted first draft adapted from the firm's precedent; all changes compared and checked; no authorities cited." Whether the client is told, and how, is the lawyer's decision and duty. Your job is to give them what they need to make it, which is exactly what the report line and the terms provide. The wider question of which legal tasks should stay human is covered in legal tasks a small firm should never hand to AI.
Six failures and the sign that shows each
- An invented authority. The sign is a case or rule you can't find in the research service. The fix is the rule that AI never supplies case law, and that any authority a tool names is pulled and read in full.
- Another matter's details in a draft. A stray name or date from a previous job survives in a reused chat or project. The sign is a detail the facts sheet never contained. Open a new chat for each matter, clear project files between matters, and search every draft for the placeholder key and for other clients' names.
- A merged speaker. Two witnesses' accounts end up in one summary row. The sign is a row that quotes someone saying two things they could not both have said. The fix is the one-speaker-per-row rule and the citation check.
- Privileged text in the wrong tool. It happens when documents arrive unlabelled and you are in a hurry. Write the tier at the top of every file before you open it, so the decision is made before the paste.
- An inflated narrative. "Read" becomes "analysed". The fix is the instruction to use only the verbs and objects in your notes, plus a read-through before each invoice.
- The lawyer hears about your AI use from someone else. That ends relationships faster than any error. The permission note and the report line prevent it at almost no cost.
A four-week start
Week one is paperwork: send the permission note, label your files by tier, and write the one-line report template. Week two is billing narratives for one lawyer's time, using matter numbers and placeholders only, because it is your own notes and the lowest risk. Week three is one drafting job from a precedent, with the compare check. Week four is one summary with the full citation check. Record three numbers for each workflow: minutes saved, corrections made, and questions the lawyer asked. After a month the numbers say which workflows earn their place, and the lawyer's questions say which of your safeguards to tighten before you extend anything.
Questions independent paralegals ask about AI
Do I need the lawyer's permission before using AI?
Yes, and in writing. The lawyer you work for is responsible for your output and for client confidentiality, and the guidance they answer to expects clear AI policies for non-lawyers as well as lawyers. A short permission note, listing tools, off-limits matters and what you will report, protects both of you.
Can I put client names into ChatGPT or Claude?
Not into a consumer plan, and only into a business plan if the supervising lawyer's note allows it. Use placeholders for names and identifiers where you can. Privileged or sensitive material should go only into a tool the firm has approved for that matter, or into none. Placeholders reduce risk, but dates and rare facts can still identify a client.
Can I bill for the time AI saved me?
On hourly work, bill the time you actually worked, including writing prompts and checking AI output. Learning a tool you use regularly is not billable, and a subscription that works like office equipment is overhead unless agreed otherwise in advance. A fixed fee per deliverable avoids counting hours, but set it for the work as AI now makes it, because guidance for lawyers warns that an unchanged flat fee can become unreasonable.
Does the client need to be told I use AI?
Whether and how the client is told is the supervising lawyer's duty and decision, so give the lawyer what they need: which tools, for which tasks, with what safeguards. Guidance for lawyers says they may tell clients how they use AI even where it is not required, and calls the engagement agreement a logical place to do it; the wording is theirs.
Further reads
- AI Contract Review for Small Firms: What It Catches and Misses — What AI contract review catches and misses
- Can Solicitors Use ChatGPT Without Breaching Confidentiality? — Confidentiality questions from the lawyer's side
- AI Billing Software for Lawyers: Which Features Earn Their Cost — Which billing features earn their cost
- AI App for Lawyers: Score Accuracy, Privacy and Cost in a Short Trial — Score accuracy, privacy and cost in a short trial
- ChatGPT or a Legal AI Tool: Which Should a Small Firm Use? — General assistant or legal-specific tool
- Best AI Chatbot for Lawyers: General Models vs CoCounsel — General models against CoCounsel for lawyers
- AI for Expert Witnesses: Report Drafting With Defensible Checks — A solo expert's defensible-report checklist: what AI may do, how to disclose its use, and the verification log that stands up in cross-examination.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: the text of a July 2024 national bar association formal ethics opinion on generative AI (archived copy), CoCounsel Legal product and plans pages, Spellbook pricing page, Clio legal AI page, a public database of court decisions involving AI-invented material, and the pricing pages for ChatGPT, Claude and Microsoft 365 Copilot Business, all read on 4 November 2026.