How Small Law Firms Use AI to Draft Letters and Routine Documents

Start from your own precedents, not a blank prompt. Give the AI an approved template plus the matter facts and ask it to fill and adapt, flagging anything it had to assume. Routine letters such as acknowledgements, client care letters, chasers to the other side and file-closing letters suit this well; anything setting out advice, rights or deadlines needs full fee-earner review.

The trap is the general chatbot drafting from nothing. Ask one for "a letter to the other side's solicitors chasing a reply" and you get something fluent and plausible, with a deadline you didn't set, a threat you didn't authorise and phrasing no one at your firm would use. Every one of those has to be found and removed. Starting from a precedent flips the review: instead of checking a whole letter, the fee earner checks the handful of places where the draft differs from an approved text. That is where the time saving actually comes from.

Follow me on Instagram@sagnikteaches

Sort your letters into three risk bands first

Before choosing a tool, list the letters the firm sends most and put each in a band. Here's a filled-in version for an illustrative family law practice:

Connect on LinkedInSagnik Bhattacharya

Most small firms find that green and amber letters make up the bulk of outgoing correspondence by count, even though red letters take the most time each. Start automation in green, prove the review routine in amber, and treat red as "AI helps with phrasing" rather than "AI drafts". The legal tasks a small firm should never hand to AI covers the red band in more detail.

Subscribe on YouTube@codingliquids

Bands attach to the letter as it will actually be sent, not to its type. A file-closing letter is green when the matter is finished. In a financial remedy matter where the pension sharing order still has to be implemented by the pension provider, the same closing letter has to tell the client what remains outstanding and who is responsible for chasing it, and that moves it to amber at least; if it mentions a time limit, red. Put one question on the drafting screen before anyone uses a green precedent: "Does this letter tell the client they still need to do something, or by when?" A yes moves it up a band.

Three ways to set it up, and what each suits

Document automation filled from your case management system

If most of your routine letters are standard with variable fields, document automation beats generative AI for green letters: it fills the same approved words every time. Clio Draft (formerly Lawyaw) converts Word documents into templates with conditional logic that adjusts pronouns and clauses, and pulls client and matter details from Clio Manage, or runs on its own if you don't use Clio. Your existing case management system may have something similar; check before buying.

The conditional logic is what makes a template cover real variety. A probate practice's letter notifying a bank of a death needs to read "I am the executor" for one executor and "we are the executors" for two, "the late [[DECEASED_NAME]]" throughout, and an extra paragraph asking for interest to be calculated to the date of death only where the account was a savings account. Set those three conditions once, from fields the case system already holds (number of executors, account type), and the same approved text produces the right letter for every estate without anyone editing a pronoun. A generative model given the same job will usually get the pronouns right too, but it will also rephrase the interest paragraph slightly differently each time, and every rephrasing is something to read.

Copilot in Word, pointed at your precedent

For amber letters, where some adaptation is needed, Microsoft 365 Copilot in Word lets you reference a specific file as a source by typing "/" in the Copilot prompt box and choosing the document. Point it at the precedent and the attendance note, and ask it to draft from those. Microsoft 365 Copilot Business is $21 per user a month on annual billing; bundled with Business Standard, the combined price is $23.50, less than buying the two separately. For contract-heavy practices, Spellbook works inside Word as a contract drafting and review add-in; its pricing is by quote. Drafting from your own files with Copilot in Word shows the referencing in practice.

Referencing the attendance note brings its own risk: the note contains things said in passing. An illustrative case from a client care letter drafted this way. The fee earner's note from the first meeting read "costs discussed, probably 6-8k to first appointment, will confirm". Copilot, asked to draft the client care letter from the precedent and the note, filled the estimate field with "$6,000 to $8,000 to the first appointment", which read as a firm estimate the fee earner had never confirmed. The fix is a field rule in the precedent itself: [[COST_ESTIMATE: from the approved estimate only; never from notes]], and a line in the prompt telling the model to leave the field as missing when there's no approved figure.

A general assistant with the precedent pack in a shared Project

ChatGPT Business and Claude Team both offer shared Projects where you can store precedents and standing instructions, and neither trains on business content by default. This suits firms without Microsoft 365 Copilot, and it works well for amber letters if the precedents are clean. Whether a general assistant or a legal-specific tool is the better fit overall is a separate decision; ChatGPT or a legal AI tool for a small firm works through it.

The Project route is also the cheapest to test. For the family practice's four users on Claude Team Standard, that's four seats at $25 a month on monthly billing, $100 in all, or $80 a month billed annually; ChatGPT Business Standard costs the same. Against the time figures in the first month below, the subscription is covered by well under an hour of fee-earner time a month, so the real cost to watch is the six hours of precedent cleaning up front.

A simple rule: green letters in high volume, use document automation. Amber letters, use Copilot in Word or a Project with the precedent attached. Red letters, the fee earner writes and AI only tidies phrasing.

Build the precedent pack the AI works from

The quality of AI drafts depends almost entirely on the precedents. Most firms' "precedents" are old letters from real matters, which is the root of the most embarrassing errors. Clean ten to fifteen of your most-used letters into proper templates:

• Replace every matter-specific detail with a named field in double brackets.
• Mark sections that must never change with a note, such as [[FIXED: complaints wording]].
• Add a one-line purpose at the top: when to use this letter and when not to.
• Keep one version per letter type. Three versions of the client care letter means three sets of drafts drifting apart.

Cleaning is mostly spotting the details that don't look like details. One sentence from an old request-for-documents letter, before and after:

"Your husband", "Tuesday", "joint", "March" and "three payslips" are five facts from someone else's matter, and none is in brackets, so a model filling fields would leave every one of them in place. Read each old letter once with a highlighter for anything that would be untrue for the next client, and turn each highlight into a field or delete it.

Here's the opening of a filled-in precedent for an amber-band client care letter, as it might sit in the pack:

PURPOSE: Client care letter for new family matters (financial
remedy). Not for children-only matters; use CC-02.

Dear [[CLIENT_TITLE]] [[CLIENT_SURNAME]],

Thank you for instructing us in connection with [[MATTER_DESC]].
[[FEE_EARNER_NAME]], [[FEE_EARNER_ROLE]], will be responsible for
your matter, supervised by [[SUPERVISOR_NAME]].

Our estimate of costs for [[STAGE_DESC]] is [[COST_ESTIMATE]],
based on the assumptions set out in section 4.

[[FIXED: complaints procedure paragraph, do not edit]]

A realistic mistake shows why this cleaning matters. In a firm using an old real letter as its "precedent", a draft for a new client went out with the previous client's child named in paragraph three. The AI had filled the obvious fields and left the rest, because nothing marked that sentence as variable. The fee earner skimmed the fields that looked like fields. With a cleaned template, there is no previous client in the document to leak.

The fill-and-flag drafting prompt

Whether in Copilot or a Project, use a prompt that forbids invention and makes every change visible:

Draft a letter using the precedent [[precedent name]] and the
matter facts below.
Rules:
1. Fill every [[FIELD]] from the facts. If a fact is missing,
   leave the field as [[MISSING: field name]].
2. Do not change text marked FIXED.
3. You may adapt other sentences only where the facts require it.
   List every adapted sentence after the letter, with the reason.
4. Do not add deadlines, consequences, legal propositions or
   case references that are not in the facts.
5. Keep our house style: spelling and tone as in the precedent,
   no "please do not hesitate to contact us".
Matter facts: {paste from attendance note or case system}

An illustrative result for an amber chaser to the other side, in an employment matter where the respondent's solicitors haven't replied to a request for documents:

"Dear Sir or Madam, We refer to our letter of 3 September requesting copies of our client's personnel file and the disciplinary notes of 12 June. We have not received a reply. Please provide the documents within 14 days of the date of this letter, failing which we will make an application to the tribunal without further notice. Yours faithfully..." Adapted sentences: "Please provide the documents within 14 days" (facts mention a previous request; standard follow-up period assumed).

What the fee earner fixes: the 14 days and the threat of an application were both assumptions, and the model admitted the first but not the second. Neither was in the facts, and the client hadn't authorised an application. The fee earner deletes the threat, sets the period they actually want, and adds a line to the prompt: "never state what we will do next unless the facts say so". That is the value of the "list every adapted sentence" rule; even when the list is incomplete, it points the reviewer at the paragraph where the model was improvising.

Employment and probate practices weight the bands differently

An illustrative two-solicitor employment practice mostly sends amber letters: requests for documents, chasers to respondents, updates to claimants on procedure. It gets the most from Copilot in Word pointed at precedents, and almost nothing from document automation, because few of its letters are fixed enough to template fully. A wills and probate practice is the opposite: its letters to banks, registries and beneficiaries are highly standard, so document automation filled from the case management system does most of the work, and generative AI is used mainly to summarise replies from banks before the fee earner acts.

The family law practice's first month

Back to the family practice from the risk-band table, with illustrative numbers. Three fee earners, one assistant, about 45 outgoing letters a week, most of them green or amber.

• Week 1: the assistant and one solicitor clean twelve precedents (about six hours between them) and store them in a shared Project. Nobody drafts with AI yet.
• Week 2: green letters only. The assistant drafts acknowledgements, requests for documents and file-closing letters from the precedents; a solicitor reviews. Drafting time per green letter falls from about ten minutes to four, including review.
• Week 3: amber letters start, beginning with chasers and client updates. Review is slower than expected at first, around seven minutes a letter, because the fee earners are reading the adapted-sentence lists carefully. Two drafts contain assumed deadlines; both are caught.
• Week 4: the prompt gains two lines prompted by week 3's catches, and the client care letter moves to amber drafting. Across the week, the firm estimates roughly five fee-earner hours saved, most of it on green letters.

Five hours a week across three fee earners isn't dramatic, but it compounds, and it came without a single letter going out with an invented obligation. The firm's own honest measure is the catch log: every error found in review, what caused it and what changed. If the log stops growing, the review is working or has stopped happening; spot-check which.

Three illustrative entries from the practice's log in weeks 3 and 4:

Each row ends in a change to a precedent or the prompt, which is the point. An entry with "reviewer caught it" and no change will happen again next week.

What the reviewing fee earner checks on every draft

1. Names and parties: client, other side, their representatives, children or third parties. Check every name against the file, not against the draft's own consistency.
2. Dates and periods: any date or deadline must come from the file or the fee earner, never the model.
3. Defined terms: "the Property", "the Agreement" used consistently and matching the precedent.
4. New obligations or threats: anything the firm or client is now committed to that wasn't instructed.
5. Recipient: if the other side is represented, the letter goes to their representative, not to them directly.
6. Enclosures: everything listed is attached, and nothing attached is from another matter.
7. [[MISSING]] fields: none left in the final version.

Keeping client confidentiality intact while drafting

Drafting means pasting matter facts, so the tool matters. Use a business tier with no training on your content and admin controls over sharing, keep precedent Projects separate from anything holding live client documents, and include only the facts the letter needs. A consumer account used from a fee earner's phone is where confidentiality problems usually start. The professional duty is the same whatever the tool; whether solicitors can use ChatGPT without breaching confidentiality sets out the settings and questions to ask, and your regulator's guidance is the final word for your firm.

If your routine letters touch any legal authority at all, even a single case reference in a chaser, route that part through a citation check; the small-firm checking routine for AI-invented case law is written for exactly that.

Further reads

• AI for Small Law Firms: What to Automate First — Where letter drafting sits among a small firm's first automations.
• AI Implementation Plan for a Small Law Firm: The First 90 Days — A 90-day plan that puts drafting in the right order.
• Per-Seat or Pay-As-You-Go? Legal AI Pricing for Small Firms — How per-seat and pay-as-you-go legal AI pricing compare.
• How Small Law Firms Use AI to Answer Client Status Questions — The next routine correspondence job worth handing to AI.
• How to Set Up Human Review for AI Work Without Slowing Down — Design review so it catches errors without slowing fee earners.
• How Conveyancers Use AI to Cut Admin on Each Transaction — Stage-by-stage admin savings for a conveyancing file, a title-summary prompt with sample output, chaser wording and the tasks AI must never touch.
• AI Contract Review for Small Firms: What It Catches and Misses — What AI contract review reliably catches, what it misses and why, with a worked review, a test method and a prompt that demands evidence.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Microsoft Support, Draft and add content with Copilot in Word; Microsoft 365 Copilot Business pricing; Clio Draft product pages and help centre; Spellbook product and pricing pages.