Outside Counsel Bill Review Prompt: Catch Padding Before You Pay

Legal departments approve invoices they never actually read. A partner bills 9.4 hours in one line that reads "review documents; calls; strategy," a paralegal's clerical filing shows up at $250 an hour, and the associate rate quietly jumped $75 since the engagement letter. Nobody has time to police it. An outside counsel bill review prompt puts the line-by-line audit back in reach: paste the invoice and your billing guidelines, get back a scored table with a defensible write-down per flagged line.

The pages that rank for legal bill review are almost all vendor marketing. Brightflag's own explainer on how AI invoice review works never enumerates the violation categories and routes you to "Book a Demo" (brightflag.com). Milliman's comparison of manual, rules-based, and AI review promotes its Datalytics-Defense product and shows no flag categories, no rubric, no sample memo (milliman.com). The concept is everywhere. The copyable tool is nowhere.

That's the gap. You don't need a platform to run the logic on one firm's invoices. You need the rules written down and a model told to apply them mechanically.

What you can do with a bill review prompt

The job is narrow and repeatable, which is exactly what makes it a good fit for a reusable prompt:

• Flag block billing where one time entry bundles several discrete activities with no per-activity hours.
• Catch clerical or administrative work (filing, scheduling, formatting) billed at an attorney rate.
• Spot duplicate and expanded entries where two timekeepers bill the same internal conference.
• Detect unapproved rate increases against the rates in your engagement letter.
• Surface vague narratives ("attention to matter," "review file") that don't justify the time claimed.
• Produce a firm-ready write-down memo with the dollar amount and the rule each deduction cites.

Every one of those is a rule you can state in a sentence. The model's job isn't to have an opinion about whether a bill is fair. It's to apply your rule and show its work.

Why the model needs a ruleset, not a vibe

Ask ChatGPT to "review this legal invoice for overbilling" and you get soft language: this entry seems high, this might be padded. Useless in a conversation with the firm. The fix is to hand the model an enforced ruleset first, then have it score against that.

This is why the strongest version of the job runs in two moves. First, convert your Outside Counsel Guidelines into a numbered, checkable ruleset with a pass/fail test and a default write-down per rule. Then score each invoice line against that ruleset. Splitting the two keeps the audit consistent across every invoice from every firm, because they're all measured against the same rules rather than the model's mood that session.

Anatomy of the prompt

The reusable shape is a ruleset builder feeding a line-item auditor. Here's the skeleton:

Variables
  {{billing_policy}}    -> your Outside Counsel Guidelines, pasted as-is
  {{rate_caps}}         -> approved rates by timekeeper level (optional)
  {{prohibited_tasks}}  -> work you never pay for (clerical, first-class travel...)
  {{invoice_text}}      -> the invoice lines to audit

Prompt
  Role: senior legal-bill auditor enforcing OCGs
  Step 1: turn {{billing_policy}} into a numbered ruleset
          (Rule ID, pass/fail test, default write-down)
  Step 2: score each invoice line against the ruleset
  Rule: cite the exact Rule ID on every flag; no vague "looks high"

Output contract (locked)
  | Line | Timekeeper | Hours | Rule ID | Violation | Write-down $ | Rationale |
  then: a firm-ready deduction memo totaling the recommended write-downs

Notice the output contract sits at the end. Models weight the most recent tokens, so on a long pasted invoice the format instruction gets ignored if it's buried up top. Put the table spec last and the audit lands in the shape you asked for.

Naming violations the UTBMS way

If your firm bills against UTBMS codes (the L-series for litigation, E-series for expenses), tell the model to reference them. A clerical entry billed under an attorney activity code is a stronger flag when the prompt names the mismatch than when it just says "this seems administrative." The vocabulary is your audience's native language. Use it.

Variables you'll set

The {{billing_policy}} variable is the one that earns its keep. Paste your actual guidelines and the ruleset reflects your policy, not a generic template. Leave it as Not stated and you still get a defensible default, which is better than the zero-review most invoices get today.

Model behavior worth knowing

Claude tends to honor a ## Output format heading and keep the audit table columns stable across a long invoice. GPT-4o is faster but benefits from the column spec restated on the final line of the prompt, especially past a few dozen line items. Both models will hedge if you let them. The single instruction that stops the hedging: "For every flagged line, cite the Rule ID and state the write-down in dollars. Do not use the phrase 'looks high.'" Take the vague escape hatch away and the output gets specific.

One honest caveat. The prompt flags candidates; a human confirms the write-down before it goes to the firm. Treat the output as a first-pass audit that turns a two-hour manual review into a ten-minute confirmation, not as an unattended approval. Anything a model reads off an invoice can be misread, so the last step is always your eyes on the flagged lines.

Getting started

1. Copy your Outside Counsel Guidelines into {{billing_policy}}. If you don't have formal ones, write Not stated.
2. Add rate caps and prohibited work if you track them; otherwise leave them Not stated.
3. Paste one firm's invoice into {{invoice_text}} from the PDF or your e-billing export.
4. Run it and read the audit table. Confirm each flag against the invoice line.
5. Send the deduction memo to the firm, or fold the numbers into your e-billing rejection.
6. Reuse the same ruleset on the next invoice so every firm gets measured identically.

Run this on a quarter of invoices before you approve them and the recovered spend usually pays for the review many times over. The point isn't to nickel-and-dime good firms. It's to enforce the guidelines you already agreed on, consistently, without burning a day on it.

For adjacent finance and legal work, the same rule-first structure powers a contract review pass and a vendor review. If you're wiring AI into a legal or finance team, read the AI usage policy prompt for the guardrails, and the contract risk review prompt for the sister workflow on the agreement side.