Legal 500 Submission: Draft Work Highlights and Client References

Write a Legal 500 submission one practice area at a time: a short overview, team details, dated work highlights, a list of active clients, and a separate referee spreadsheet. AI can draft the highlights from redacted facts, but client names, values and anything confidential never go into the prompt, and a partner signs off every line before upload.

Most firms lose time in two places: chasing partners for matter details, and arguing about what is confidential. AI helps with the first only if you give it a clean, anonymised fact sheet, and it makes the second worse if someone pastes a live client file into a chat window. The routine below keeps the real names in your own documents and only the draft wording in the AI.

Follow me on Instagram@sagnikteaches

What the directory actually asks you to send

Read the current guideline for your market before you open any AI tool. Formats, caps and dates differ by edition and region, so what follows is drawn from recent published guidelines and must be checked against the file for your own practice area. The structure is consistent enough to plan around.

Connect on LinkedInSagnik Bhattacharya

• A separate document for each practice area, uploaded through the directory's submissions site. The guidelines say plainly that emailed submissions are deleted and not actioned.
• Practice overview: brief bullet points such as hires and departures, what the team is best known for, and one very significant matter. The guideline asks for no lengthy paragraphs and no unsubstantiated claims.
• Feedback on the previous edition: do you agree with the ranking, and if not, what you suggest and the evidence for it.
• Contact details for the person the researchers should contact to arrange their follow-up: name, position, email and phone.
• Team information: head of team, leading partners, next-generation partners, leading associates, arrivals and departures, headcount. No extensive biographies.
• Detailed work highlights: in the guideline I read, up to 20 matters inside a defined twelve-month window, each with the firm's role, the partner who led, dates and the client. At least one matter should be cleared for publication.
• Active clients: a list that marks new clients and the length of long-standing relationships.
• Referees, in a separate Excel file built on the directory's own template, with a hard deadline of its own (in the guideline I read, the same day as the submission). The same guideline called 20 referees per practice area "a good minimum guideline", asked for a mix of long-standing contacts and recent client gains, and noted that some junior referees can be more responsive than senior figures.

Treat every number in that list as "check this year". The point for AI drafting is the shape: highlights are structured records, which is exactly what language models handle well, as long as the record is yours and the confidential fields have been removed.

Subscribe on YouTube@sagnikteaches

Why confidentiality is the whole design problem

The guideline says that anything you send may be published by the directory's publications unless you mark it confidential, and it asks you to label confidential client names, values or other details with the word CONFIDENTIAL in bold red upper case, and to set the confidential text itself in red. It also asks you not to put a blanket confidentiality notice at the front, and to be clear about whether the case, the client name or the value is the secret.

That gives you three tiers of fact for each matter, and your AI workflow should follow them:

1. Public: already reported, or the client has cleared it. Safe to draft from, subject to your firm's tool policy.
2. Confidential to the directory: may go to the researchers, marked as confidential, but must never leave your systems for a consumer chat tool.
3. Not for anyone: privileged material and anything your engagement terms bar you from describing. It does not go in the submission or the prompt.

Which AI tool is acceptable is a firm decision, not a drafting trick. If you have not settled it, read ChatGPT or a legal AI tool: which should a small firm use? and which legal tasks a small firm should never hand to AI first. ChatGPT's and Claude's business plans say they do not train on your content by default, and their consumer plans have a model-training switch in privacy settings, but a privacy setting does not change your duty to the client.

Stage one: build a matter register in your own system

Before any drafting, collect matters into one spreadsheet or your practice-management tool. One row per matter, columns filled by the fee-earner, not by AI. This is the master record and the only place the real client names live.

The two columns that do the heavy lifting for AI are "anonymised descriptor" and "issue that made it hard". Partners resist writing a paragraph, but they will answer two focused questions. A five-minute voice note per matter, typed up by a colleague, is often enough to fill them.

Stage two: draft highlights from redacted facts

Now the AI step. You paste only the anonymised columns: never the client name, never a matter ID that contains the client's initials, and never a value the client has not cleared. Say a mid-sized commercial litigation team has twelve candidate matters. Here is a prompt that works on one row at a time.

You are helping a law firm draft one "work highlight" for a legal directory submission.
Use ONLY the facts below. Do not add any fact, figure, date or party that is not here.
Write 70 to 90 words in plain, factual English. Start with the firm's role.
Leave the client as the descriptor given. Do not use words such as "landmark",
"pioneering" or "leading". If something needed is missing, list it under MISSING
instead of guessing.

Role: acted for the defendant
Client descriptor: a regional logistics group
Lead partner: [Partner A]; team: [Associate B], [Associate C]
Dates: March 2025 to May 2026
Issue: claim over a delayed warehouse fit-out; counterclaim for unpaid sums;
disputed expert evidence on cause of delay
Outcome: claim dismissed after trial; counterclaim succeeded in part
Value band: mid seven figures
Disclosure tier: confidential to directory

A realistic reply, labelled illustrative, looks like this:

Acted for the defendant, a regional logistics group, in a claim over a delayed
warehouse fit-out, led by [Partner A] with [Associate B] and [Associate C] (March 2025
to May 2026). The team also pursued a counterclaim for unpaid sums and challenged the
claimant's expert evidence on the cause of delay. The claim was dismissed after trial
and the counterclaim succeeded in part. Value: mid seven figures.

MISSING: whether the court's decision is public; the trial length.

What to fix: the output is sound because the input was tight, but a partner still needs to confirm that "dismissed after trial" matches the court's decision and decide whether "mid seven figures" is acceptable in the confidential box. The MISSING line is the most useful part, because it turns the AI into a checklist generator rather than a creative writer. Once the facts are verified, a person re-inserts the real client name in your own document and applies the red upper-case marking the guideline asks for.

Three small edits that stop most bad drafts

Ask for one highlight per prompt and per matter. Batching twelve matters in one prompt invites the model to borrow phrasing and even outcomes from one matter to fill gaps in another. Give it a word cap, because directory researchers read hundreds of these. And forbid superlatives in the prompt itself, since the guideline explicitly warns against unsubstantiated claims.

A before and after on a weak highlight

Here is a typical partner draft and the version after a redaction-first AI pass. Both are illustrative.

Before: "We were instructed in a highly complex, ground-breaking dispute for a major client, achieving an excellent result in difficult circumstances."

After: "Acted for the claimant, a family-owned manufacturer, in a dispute over a supply contract terminated after a change in specification. Led by [Partner A] (January to November 2025). The tribunal upheld the challenge to the termination and awarded damages. Outcome: claim succeeded in full."

The "after" has a role, a descriptor, a lead, dates and an outcome. A researcher can ring a referee about it, which is what the directory does next.

Stage three: the referee list, where AI should do less

Referees are real people at real clients who will receive an email from the directory, so this is the stage where restraint pays. AI is useful for logistics and wording, not for choosing or contacting anyone.

1. Source names from the matter register. The relationship partner proposes a referee per matter, preferring a mix of long-standing contacts and recent client wins, as the guideline suggests.
2. Ask for consent personally. A short, human request from the partner beats a form email. AI can draft that request for the partner to edit.
3. Log consent in the register before the name goes on the directory's spreadsheet.
4. Use the directory's own template unchanged. The referee notes I read ask you not to delete any columns, add headings or notes, type in capitals, or send referees in any other format, and warn that referees may not be contacted otherwise. The directory generates its outreach emails from the spreadsheet, so a typo in the file goes straight into the email.
5. Brief the referee, lightly, on what the researcher may ask. Do not script their answer.

A draft the partner can adapt, with placeholders:

Subject: A small favour, if you are willing

Hi [first name],

Each year the Legal 500 asks clients for feedback on the law firms they use. We would be
grateful if we could put your name forward as a contact for our [practice area] work,
including [matter description kept generic].

If you agree, one of its researchers will email you asking for your feedback, and
you can be as candid as you like. If you would rather not, that is
absolutely fine and it will not affect how we work together.

Thanks either way,
[your name]

Notice what is absent: any suggestion of what to say, and any pressure. A referee who feels coached sounds coached, and researchers hear the difference.

A consent tracker for twenty referees

Say a four-partner employment team is aiming for twenty referees. Build a simple tracker with a status for each name: not asked, asked, yes, no reply, declined, listed. Each Friday, the practice manager counts yes and listed and chases the asked column once. If only twelve have said yes by the internal deadline, you submit twelve and add the stragglers only if they reply before the directory's deadline. The guideline says late referees may be contacted much later or not at all, so there is no value in a name that arrives after the cut-off.

Stage four: practice overview and the feedback section

The overview and the "feedback on last year's ranking" box are where firms over-claim. Ask AI to turn your bullets into tighter bullets, nothing more. A prompt that keeps it honest:

Tighten these overview bullets for a legal directory. Keep each bullet under 25 words.
Remove adjectives that cannot be proved. Do not add facts. Flag any bullet that makes a
claim needing evidence, and say what evidence a partner should attach.

- We are widely regarded as the best employment team in the region.
- Two new partners joined from larger firms in 2026.
- We acted in several high-profile cases this year.

A sensible reply flags the first bullet as unsupported and asks for client or peer evidence, and flags "several high-profile cases" as vague. The second bullet survives. Fixing a vague bullet means going back to the matter register, which is the point: the register is the evidence.

Feedback on the ranking

If you disagree with a ranking, the guideline invites a short evidence-based case. Give AI your facts (matter counts, new instructions, hires) and ask it to organise them as a three-bullet argument. Do not ask it to argue harder. A researcher will discount enthusiasm and weigh numbers.

A second worked example: the first-time submission

A small property-finance practice has never submitted. A tip in the guideline I read for first-time areas was to include several examples from the previous two years as well as at least ten new ones, so the team builds a register of eighteen matters: ten inside the window and eight older. Partners rate each matter's disclosure tier; five turn out to be cleared for publication, nine can go to the directory marked confidential, and four are off limits.

The associate drafts fourteen highlights from redacted facts, using the prompt above, in about ninety minutes of prompting and reading. Each draft is checked against the register by a second lawyer, and two highlights fail the check: one gave the wrong end date, and one said the firm "acted for the borrower" when the firm acted for the lender. Both errors came from a partner's ambiguous note, not from the AI, which is a good reminder that the register's quality caps the draft's quality. The lender and borrower slip is exactly why a partner reads every line before it goes anywhere near the directory.

Stage five: the active-clients list and new-client gains

The active-clients list looks like the easy part, and it is where confidentiality slips are most likely, because a list of names is the one thing a researcher can publish at a glance. Keep the real list in the register and give the AI only categories when you want help with wording. For example, a prompt can turn three facts per client into one line without ever seeing a name:

Turn each row into one line for an "active clients" list. Keep the descriptor exactly as
given. Say "new in 2026" only where the row says so. Do not guess sector or size.

1. Descriptor: a family-owned manufacturer; relationship: since 2014; tier: public
2. Descriptor: a regional logistics group; relationship: new in 2026; tier: confidential
3. Descriptor: a venture-backed software company; relationship: since 2022; tier: confidential

The reply is three short lines, such as "A family-owned manufacturer, client since 2014". The value is consistency, not creativity. A partner then swaps descriptors for real names in your own document, with the red marking on anything confidential.

A mistake that looks harmless: the helpful model fills a gap

A realistic failure shows why the MISSING instruction matters. Say a team leaves the outcome blank on a matter that settled before trial. Without a MISSING rule, a drafting tool may write "the matter settled on favourable terms", which sounds plausible and may be false. With a confidentiality settlement attached, it may also be a statement the firm is not allowed to make. The fix is procedural: the prompt forbids invented outcomes, the register has an explicit "outcome cleared for use" field, and the second reader checks every outcome sentence against that field rather than against memory.

A fifteen-minute review meeting that catches the rest

Once drafts exist, run one short meeting per practice area rather than circulating a document for comments. The practice head, the practice manager and one associate read the highlights aloud in order. Anything that makes a listener ask "which client?" or "when?" goes back to the register. Anything that sounds like marketing is cut. Anything the partner cannot prove is removed. Reading aloud finds the repeated openings ("Acted for... Acted for...") that language models produce when every prompt has the same shape, so vary the first words by hand where the role differs, for example "Advised the lender" and "Defended a claim brought by".

Keep the meeting's output in the register, not in the document. If a partner changes a date or an outcome in the room, the register changes first and the draft is regenerated or edited from it, so next year's team starts from a corrected record instead of an annotated Word file nobody can trace.

Checks before upload

• Fact check against the register: roles, dates, lead partner, outcome. A second person, not the drafter.
• Disclosure pass: every confidential item marked as the guideline asks, and no blanket confidentiality banner at the top.
• Window check: every matter inside the stated period, or explained if older.
• Placeholder sweep: search for square brackets, "MISSING" and "[Partner".
• Tone sweep: remove "leading", "landmark", "unrivalled" and similar words.
• Referee file: template unchanged, consent logged, email addresses checked character by character.
• Delivery: upload through the submissions site, not by email, ahead of the deadline.

When a highlight will not survive anonymising

Some matters are recognisable even without a name: a one-off, industry-shaking dispute where the descriptor gives it away. You have two options: ask the client for permission to describe it, or leave it out of the AI step and write it by hand with the partner. Do not weaken the redaction to get a better draft. A drug developer's only patent case in a given year is identifiable from "a mid-sized drug developer" plus a date range.

How to tell whether the process worked

Judge it by internal measures, not by the ranking, which you do not control. Did every highlight reach the partner with all the register fields filled? How many corrections did the second reader make? Did every referee on the file know they were coming? Did the final document need red-marking fixes after review? Track these across years and the routine gets faster. The following year you re-use the register, update it and re-run the prompts.

If you want the matter register and the redaction habits designed for your own firm, that is the kind of job my AI implementation consultation covers. For the broader rules a firm must follow when describing results, see AI marketing for small law firms, and for summarising large documents safely, how to summarise bundles and transcripts with AI safely.

Questions lawyers ask before using AI on a directory submission

Can I paste a client's name into ChatGPT or Claude for a highlight?

Only if your engagement terms, your confidentiality duties and your firm's AI policy allow it. The safer habit is to strip names, values and dates into placeholders first, draft on the placeholders, and put the real details back by hand inside your own document system.

Do I need a referee's consent before listing them?

Get it. The directory emails every referee you list, so a surprised client reflects badly on the firm. Ask the relationship partner to make a short personal request first, and record the yes in your tracker before the name goes on the spreadsheet.

How many work highlights should one practice area include?

Follow the current guideline for your market and practice. One recent regional guideline allowed up to 20 matters and asked for at least one cleared for publication. Quality and breadth of the matters matter more than reaching the cap.

Should AI write the practice overview as well?

It can tighten bullet points you supply, but the overview carries the claims the researchers will test. Keep every claim tied to something a partner can evidence, and cut the adjectives that the directory's own guideline warns against as unsubstantiated claims.

What if the directory changes its format this year?

Download the current guideline and templates for your practice area before drafting, and rebuild your prompt headings from them. Formats, deadlines and limits differ by edition and market, so last year's file is a starting point, never the rulebook.

Further reads

• How Small Law Firms Use AI to Draft Letters and Routine Documents — Routine drafting a small firm can hand to AI.
• AI Readiness Checklist for Accountants, Solicitors, Consultants — Check the firm is ready before wider AI use.
• Can You Use AI to Write Testimonials? Where the Legal Line Is — Where client quotes and AI writing collide.
• Per-Seat or Pay-As-You-Go? Legal AI Pricing for Small Firms — What legal AI tools cost per seat or per use.
• AI Literacy Requirements: What Your Staff Need to Know — What your staff need to know before using AI.
• Turning Hours Saved by AI Into Advisory Work Clients Pay For — Turn the time saved into work clients pay for.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: the Legal 500 submission information pages, a 2027-edition editorial submission guideline and its referees guidance (checked 4 November 2026), and OpenAI's business-data page, Claude's pricing page and the two vendors' consumer training-setting help pages.