AI Legal Bill Review: Fix Time Narratives Before Invoices Go Out

Export the pre-bills with matter codes instead of names, then put each time entry through three checks before a partner sees it: a rewrite that gives every narrative a task, subject and purpose; a privilege screen for advice or strategy a third-party reader should not see; and a test against the client's billing guidelines. The partner reviews only what is flagged.

The assistant edits wording and nothing else. It never adds work, changes time or invents detail, so a vague entry such as "Review documents" goes back to the timekeeper as a question, not forward as a better-sounding sentence. Billing for work that was not done is the failure that matters most here, and the routine is built to prevent it.

Follow me on Instagram@sagnikteaches

Why the firm should find the problem lines before the client does

Corporate clients and their e-billing systems check invoices line by line, and most published guidance on legal bill review is written for that reader. The same checks are open to the firm sending the invoice, and running them first changes what a flagged line costs. Block billing is the plainest example: it means putting an entire block of time into one invoice line, so the reader cannot tell how long each task took. The task and activity codes that e-billing invoices carry assume task-based entries, so that the client's system can test each line against its billing guidelines; a multi-task line defeats that, and review software can flag it for the client to reject or pay at a discount.

Connect on LinkedInSagnik Bhattacharya

Vague wording draws the same treatment. One published outside-counsel billing guideline, for example, rejects a bare "Research" or "Telephone calls" as a narrative, bans block billing, asks for time in tenths of an hour, closely reviews lawyer or paralegal time spent on clerical-level tasks, and refuses to pay for invoice preparation or review. Guidelines differ from client to client, so nothing here replaces reading the one that applies. What the firm can do is make the review mechanical enough that every invoice gets it. Take one illustrative line: 3.4 hours of mixed work at $350 an hour is $1,190 on a single line. If a client's reviewer discounts it by half, one careless narrative has cost $595, and the partner finds out when the payment arrives short.

Subscribe on YouTube@sagnikteaches

Four stages on one export

Passes 1 to 3 run on the same spreadsheet, and the partner meets the invoice only at the end, through an exception report.

Set up the export so the assistant never needs a client name

A narrative review needs the wording of each entry, not the identity of anyone in it. Six set-up steps keep it that way:

1. Export the pre-bill as a spreadsheet with these columns: matter code, date, timekeeper initials, hours, task code and narrative. Add three empty columns named proposed, action and final.
2. Run a find-and-replace list over the narrative column that swaps client, party and witness names for tags such as [CLIENT], [SUPPLIER] and [WITNESS 1]. The mapping stays on your own system.
3. Write a two-line matter note with no names, for example "Contract dispute over late deliveries between a wholesale distributor and its supplier, pre-trial". The assistant uses it to check that proposed wording fits the matter.
4. Work in a tool that does not train on your content by default. ChatGPT Business and Claude Team each cost $20 a seat a month billed annually ($25 monthly) with a two-seat minimum, and the assistant inside your practice-management software qualifies only if its terms say the same.
5. Read three documents before the first run: the engagement letter, the client's billing guideline and your firm's AI-use policy, for which the AI use policy clauses tutorial is a starting point. Ethics guidance on generative AI (Formal Opinion 512, July 2024) says informed client consent is required before client information goes into a self-learning tool, one that can carry what it is given into later output, that general boilerplate in an engagement letter is not enough, and that output needs an appropriate degree of independent verification.
6. Keep the timekeeper's original text in its own column and never overwrite it.

Pass 1: the narrative clean-up prompt, run on eight lines

The main example is one illustrative invoice: a month of pre-trial work on the supply dispute in the matter note, with 38 entries, 31.4 hours and three timekeepers. The prompt below ran over all 38 entries; eight of them appear here because together they show every outcome the prompt can produce.

You are editing time-entry narratives for a law firm's pre-bill. You change wording only.
For each entry return one row with: ID, ORIGINAL, PROPOSED, ACTION, QUESTION.
Rules:
1. Keep the hours and the task exactly as recorded. Never add work, a subject, a document name or an outcome that is not in the entry or in the matter note.
2. A good narrative starts with a verb and names the task, the document or subject, and the purpose, in 12 to 30 words.
3. If an entry is too vague to improve without guessing (for example "Review file"), set ACTION to QUERY and write one question the timekeeper can answer in a line. Leave PROPOSED empty.
4. If an entry contains more than one task, set ACTION to BLOCK and list the tasks in QUESTION. Do not split the time.
5. Never accept attention, analysis, conference, work on, review or miscellaneous as the whole description.
6. ACTION values: OK, REWORD, QUERY, BLOCK.
Matter note: [two lines, no names]
Entries: [paste the narrative column with ID and hours]

The reviewer's job is to read the table against the originals, and four things stand out. Row 2 is the commonest failure: the entry said "docs", and the assistant wrote "delivery records", a plausible noun nobody recorded. The reviewer reverts it to "outstanding documents". Compare every proposed line with its original and treat any new noun or outcome as a defect.

Rows 1, 3 and 7 are the behaviour you want, because the assistant declined to guess. A looser prompt without rule 1 would have turned "Call w/ client" into "Telephone conference with client regarding settlement options", and if settlement was never discussed, that sentence goes to the client as a statement of work done. Rows 6 and 8 pass this stage, but pass 1 is not designed to catch a clerical task or a revealing line. Passes 2 and 3 are.

The query note and the shape a finished narrative takes

Queries only work if the timekeeper can answer in a line. Send each one with the entry's date, hours and original text, and say what happens if there is no answer. A filled-in version reads:

[first name], entry 3 on 14 March (2.5 hours, "Research"): which issue did you research, and did you produce a note? Please answer in one line by Thursday, from your notes or the documents. If the work cannot be described, the time will be written off before the invoice goes out.

The last sentence matters. Without it, a query becomes an invitation to reconstruct from memory, which is how inflated narratives get written. With it, the timekeeper knows an honest "I can't say" costs the firm a small write-off and nothing else.

Give timekeepers a target shape as well, so fewer entries need rescuing next month. A good narrative opens with a verb, names the object and says why, and carries a count or a document name where one exists:

These are what a timekeeper writes from knowledge, not what the assistant should invent. Paste the five-row pattern into the firm's house style sheet, add a list of approved opening verbs (draft, revise, research, review, prepare, confer, telephone, correspond with, attend), and include the sheet in the pass 1 prompt so every timekeeper's output lands in the same shape.

Pass 2: screen for privilege and confidential detail

A narrative can tell a stranger what the client fears, what the firm plans or what it has found. That matters most when the invoice travels beyond the client's legal team: to an insurer paying under a policy, a litigation funder, a parent company's finance function or an auditor, and occasionally to a court in a fee dispute. Whether a given narrative is protected depends on the court and the matter, so the firm's standing rule belongs to your ethics partner, not to the assistant. The assistant flags a line and proposes a neutral version, and a lawyer accepts or refuses it.

You screen law-firm time narratives before an invoice goes to a client's finance team and possibly to third parties.
Flag any narrative that would let a stranger learn: the legal theory or strategy being pursued; a settlement figure or authority; what the firm has found in the evidence; a witness's weaknesses; or personal or commercial secrets of anyone.
For each narrative return: ID, RISK (HIGH, MEDIUM, LOW or NONE), WHY in one clause, NEUTRAL REWRITE.
The neutral rewrite names the task and the type of document or subject only. It must stay specific enough for a billing reviewer: a verb, a subject, a purpose.
Never add facts. If you cannot rewrite without guessing, answer QUERY.

On the worked invoice, the screen flagged 5 of the 38 lines: 3 high and 2 medium. A lawyer accepted four of the neutral rewrites and rejected one, because "Correspondence with witness" was too vague for the client's guideline; the partner rewrote it by hand as "Draft email to witness requesting a call", which names the task and gives nothing away. That tension is the core of this pass. A line can be too vague for the guideline and too revealing for privilege at the same moment, and the way out is the task and the document, never the advice and never the finding.

Pass 3: turn the client's billing guideline into checks

Each client's guideline differs, so build one rule sheet per client. Paste the guideline into the assistant and ask it to list every rule with its clause number, the guideline's own wording, a test and a type (countable or wording). Then a lawyer checks the sheet against the guideline itself. The extraction step is where a misread threshold or a missed exception creeps in, so this human check happens once per client and again whenever the client reissues the guideline.

Read the billing guideline below and list every rule that affects time entries or invoices.
For each rule return: CLAUSE, RULE quoted exactly from the guideline, a TEST a spreadsheet or reviewer could apply to a time entry, and TYPE: COUNTABLE (numbers, dates, names) or WORDING (the text of the narrative).
Include every threshold and every exception. If you are unsure whether something is a rule, list it under UNSURE.
Guideline: [paste the text]

The last row connects to the ethics guidance mentioned earlier, which says a lawyer cannot bill a client to learn a generative AI tool the lawyer will use regularly. If you are billing AI-assisted work at all, read the tutorial on billing clients for AI-assisted work first, because the rule that hourly bills must reflect actual time cuts both ways.

Countable rules belong in formulas, because a spreadsheet does not tire or approximate. With hours in column D, dates in B, timekeepers in C and narratives in E, four checks cover most of the table:

Increment (tenths):    =IF(ABS(D2*10-ROUND(D2*10,0))>0.0001,"CHECK INCREMENT","")
Daily cap (12 hours):  =IF(SUMIFS($D:$D,$B:$B,B2,$C:$C,C2)>12,"DAILY CAP","")
Attendees (limit 2):   =IF(COUNTIFS($B:$B,B2,$E:$E,"*conference*")>2,"ATTENDEES","")
Too short (under 5):   =IF(LEN(TRIM(E2))-LEN(SUBSTITUTE(TRIM(E2)," ",""))+1<5,"TOO SHORT","")

Rule flags from the assistant must carry the clause number they rely on. A flag with no clause is noise, and the reviewer ignores it. Keep the rule sheet and the formulas together in one workbook per client, so the next month's export drops into the same checks.

The partner step: an exception report in place of a full re-read

By the time the partner opens the invoice, every flag has an owner and a result, so the review is a read of the exceptions plus a small sample of the untouched lines. Here is the report for the worked invoice, with the figures it produced.

PRE-BILL EXCEPTION REPORT   Matter [CODE]   Period [month]
Entries 38 | Hours 31.4 | Value $9,720

Pass 1 (narratives)
  21 OK | 8 reworded (7 accepted, 1 corrected by the reviewer)
  6 queried: 5 answered and reworded, 1 unexplained and written off
  3 block-billed: time split by the timekeepers, all billable

Pass 2 (privilege screen)
  5 flagged (3 high, 2 medium): 4 neutral rewrites accepted, 1 rewritten by the partner

Pass 3 (guideline check)
  3 flags: 2 clerical lines and 1 third attendee, written off
  Increment, daily-cap and approval formulas: no hits

Write-offs: clerical $315 | attendee $60 | unexplained $225 | total $600
Invoice after review: $9,120

Integrity: entries 38 in, 38 out | hours 31.4 in, 31.4 out (before write-offs)
Sample: 4 of 21 untouched lines read by the partner, no changes
Originals kept in column [X].   Reviewed by [partner]   Date [ ]

The arithmetic is worth following. The partner billed 7.2 hours at $450, the associate 19.0 hours at $300 and the paralegal 5.2 hours at $150, which is $3,240 plus $5,700 plus $780, or $9,720. Before the review, the lines a client could reject or discount added up to $2,370, about a quarter of the invoice: block-billed time of $1,770, clerical lines of $315, a third attendee at $60 and an unexplained half hour at $225. After it, the three block entries stay billable because the timekeepers supplied the real split, and the firm has written off $600 itself, before any client had to ask.

The timings are illustrative too. The three assistant runs took about ten minutes in all. The timekeepers spent roughly 35 minutes answering six queries and three block-billing questions, and the partner spent 20 minutes on the report and the four sampled lines. Nobody should read those as a promise of savings. What changes is where the partner's attention goes: from scanning 38 lines for trouble to deciding on nine flagged ones (five privilege flags, two clerical lines, the third attendee and the unexplained half hour), with the integrity line proving that no hours moved.

How the routine fails, and what catches it

A four-week rollout that leaves month-end alone

Do not switch the routine on during a billing run. Introduce it in four steps and keep the old process until the new one has earned its place:

1. Week 1: run pass 1 on last month's invoice for one client, alongside your biller's normal review. Compare what each found. Nothing goes to the client.
2. Week 2: add the privilege screen and test it on three invoices that went to third-party payers. Have the partner mark which flags were worth raising.
3. Week 3: build the rule sheet for your largest client, get the lawyer's check on it, and add the four formulas to the workbook.
4. Week 4: run live on one invoice with the exception report. Afterwards, list everything the partner changed and feed the pattern back into the prompts.

If you are earlier in the journey, the broader tutorial on legal billing with AI and pre-bill checks covers the case for doing this at all; this one assumes you have decided to run it. Catching hours nobody recorded is a separate job, and the law firm time-tracking tutorial handles that earlier in the month, because a pre-bill review can only improve what is already on the page.

Questions firms ask before running invoices through AI

Should the AI see client names?

Not if you can avoid it. Replace client, party and witness names with tags or matter codes before export, because the review only needs the wording of each entry. If a narrative still names a person, treat the file like any other confidential client document, and check your engagement letter and the client's guidelines for limits on AI tools.

Can the assistant split a block-billed entry?

It can show where the tasks divide, but only the timekeeper knows how long each one took. Ask the assistant to list the separate tasks in the entry and send the entry back with a question. Dividing the time by guesswork creates a record that does not match what happened, which is worse than the block entry.

Do we have to tell clients we use AI on invoices?

Check two places: your engagement letter and the client's own billing or outside-counsel guidelines, which may carry clauses on AI tools. Ethics guidance on generative AI says informed consent is required before client information goes into a self-learning tool, and boilerplate wording is not enough. Ask your ethics partner what your situation needs.

Which AI tool should a small firm use for this?

Use a business plan that does not train on your content by default, such as ChatGPT Business or Claude Team (each $20 a seat a month billed annually, with a two-seat minimum), or the assistant in your practice-management software if its terms say the same. The routine relies on the prompts and the checks, so it works with any of them.

Can we bill the client for the time this review takes?

Pre-bill review is normally internal administration, some published billing guidelines refuse to pay for invoice preparation or review at all, and ethics guidance says a lawyer cannot bill a client to learn an AI tool the lawyer will use regularly. Time spent on the client's matter is billed as actually worked; any efficiency from AI should show up as lower hours. Ask your ethics partner where the line sits under your fee agreements.

Further reads

• AI Billing Software for Lawyers: Which Features Earn Their Cost — Which billing-software AI features earn their cost for a small firm.
• Law Firm Accounts Receivable: AI-Drafted Chasers for Overdue Bills — Chasing overdue invoices once the bills are out.
• AI Time Capture for Solicitors: Recover Hours You Never Billed — Recover hours that never reach the pre-bill.
• ChatGPT Business for Law Firms: Seats, Privacy Settings and Cost — Seats, privacy settings and cost for a firm on ChatGPT Business.
• Best AI for Solo Lawyers: A Stack That Fits One Person — Assistant, practice-management AI and contract tool for one lawyer: three priced stacks from $120 to $275 a month, and what to leave out.
• Clio Manage AI Pricing: What It Costs and How to Switch It On — Manage AI (formerly Clio Duo) is bundled into Core and above. What Clio publishes, what needs a quote, and the admin steps to switch it on.
• AI Tools for Lawyers in Australia: A Safe Starting Checklist — A checklist for Australian law practices: client data, checking output, supervision, billing and court AI rules, with a filled-in one-page policy.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Formal Opinion 512 on generative AI tools (July 2024), a published outside-counsel billing guideline, the UTBMS code-set overview, an e-billing vendor's explainer on block billing, and the ChatGPT Business and Claude Team pricing pages. Checked 4 November 2026.