Can ChatGPT check a law firm invoice?
It is an obvious thing to try: paste the guideline, paste the invoice, ask a chat assistant what is wrong. Sometimes the answer looks impressive. This guide looks at what that approach does well, where it quietly fails, and what a safe way to use it looks like. It also says when a dedicated tool is a better fit and when it is not.
What the job needs from any tool
A guideline check is not a summary task. It needs a tool that:
- Reads the whole guideline, including the clauses buried in appendices.
- Reads every invoice entry, not a sample.
- Does the arithmetic correctly: increments, totals, counts of attendees.
- Applies each rule the same way on every entry, on every run.
- Shows which clause each flag comes from, in the clause's own words.
- Tells you what it could not decide, instead of guessing.
A general AI chat assistant can do some of this, some of the time. The question is which parts you can rely on without checking.
What a chat assistant is genuinely good at
- Explaining a clause. Paste one clause and ask what it means in plain words, or what examples would break it.
- Drafting a rule list. Ask it to list the billing rules in a guideline, one line per rule with the clause number. This is a useful first draft, as long as you then read the guideline yourself and correct the list.
- Rewriting a narrative. Give it a vague entry and the facts of what was done, and ask for a clearer description. It is useful as long as the rewrite stays true to the facts you gave it.
- Brainstorming edge cases. Ask what a client might object to in a given entry.
For a single lawyer polishing a handful of entries, this is a fair use of such a tool.
Where it breaks down on a real invoice
These are the failure modes to test for, because they are hard to see from the answer alone.
- Missed content. A long guideline or a long invoice may not be handled in full in one pass. Entries and clauses can be skipped without notice.
- Paraphrased clauses. A chat answer may restate a rule in its own words or cite a clause number that does not match. A paraphrase you cannot trace to the source is not something you can show a client.
- Arithmetic. Adding hours, applying an increment, or counting how many timekeepers attended is exactly the kind of work where a language model can look right and be wrong. Do these checks in a spreadsheet.
- Inconsistency. Run the same invoice twice and the flags may differ. A check you cannot repeat is hard to trust.
- No statement of doubt. A confident answer about an entry whose rule is ambiguous is worse than a note saying "check this one".
- Retyping and copying. Pasting an invoice from a PDF into a chat can scramble columns, which changes what the assistant sees.
None of this means a chat assistant is useless. It means its output is a draft that someone still has to verify entry by entry, and that verification may take as long as doing the check yourself.
Confidentiality comes first
Invoice narratives describe client work, often including litigation strategy or a client's business, and many clients treat their billing guideline as confidential too. Before using any online assistant:
- Read the tool's terms on how your inputs are stored and whether they may be used to improve the product.
- Check your firm's own policy on client information and software.
- Check whether the client's guideline or engagement terms say anything about using AI tools on its matters. Some clients address this, and it is better to find out before than after.
- Remove or generalize names and details you do not need to share. A check on wording rarely needs a party's name.
If you are not sure, ask the client, or whoever at your firm handles confidentiality questions. This guide describes practical precautions; it is not legal or ethics advice.
A safer way to use a chat assistant for this
If you do use one, treat it as an assistant to a checklist, not as the check:
- Build the rule list from the guideline yourself, with clause numbers. Use the assistant to draft, then correct it against the source.
- Give it the rules and a small batch of entries at a time, not the whole invoice.
- Ask it to quote the clause text for every flag, then search the guideline for that quote. If the quote is not there word for word, discard the flag.
- Ask it to say "unclear" when a rule does not clearly apply, rather than forcing a verdict.
- Do arithmetic and counting in a spreadsheet.
- Run the batch twice. Where the two runs disagree, read those entries yourself.
- Have the lawyer responsible decide on each flag.
That is a lot of manual work, which is the honest cost of using a general tool for a specific job.
Side by side
| What matters | General chat assistant | Purpose-built check on the draft |
|---|---|---|
| Setup | You write and maintain the instructions | Upload two files |
| Whole invoice read | Depends on length and how you prompt | Every entry is checked, up to 2,000 on one invoice |
| Clause quoted | Only if you ask, and you must verify it | Quoted word for word in the result |
| Arithmetic | You verify it yourself | Number and pattern rules evaluated by code |
| Same result on re-run | May differ between runs | Number and pattern rules are code, so they repeat; narrative verdicts are model output |
| Says what it could not settle | Only if you ask | Entries it cannot settle show as To check |
| Best for | Drafting rules, rewriting a narrative, explaining a clause | Checking a whole draft invoice against one client's guideline |
When a chat assistant is enough
It is a reasonable choice when:
- You are checking a few entries, not a full invoice.
- You mostly need help rewriting narratives you already know are weak.
- You are happy to verify every output yourself.
- Your firm's rules and the client's terms allow it.
It is also a good way to understand a confusing clause before you decide how to treat it.
When a dedicated check makes more sense
Consider one when you check full invoices every month, when the time spent verifying the assistant's output starts to exceed the time saved, or when you need a result you can show someone: each flag with its clause. Even then, the same principle applies. The tool does a first pass. The lawyer decides. See how to check an invoice against billing guidelines for the full routine, and what block billing is and how to avoid it for the wording rule that most often needs a careful read.
How BillPrecheck does this
BillPrecheck is built for exactly one job: checking a draft invoice against one client's guideline. You upload the guideline and the invoice, as a LEDES 1998B file, a CSV or Excel export, or a Word or PDF table. A scan or photo of an invoice is read too, and every value read from it is marked to check.
Every entry comes back as Pass, Flag or To check. Number and pattern rules are evaluated by code. Narrative rules are judged by an AI model, one verdict per entry. A flag quotes the guideline clause word for word with a plain reason and a suggested rewrite. If a quote is not found word for word in the guideline, the entry shows as To check instead. A second, stronger AI model reviews every result, fixes what it can, and lists what is left for you to decide. Results are model output checked by code and by a second model; they are not verified by a person.
On confidentiality: your guideline and invoice are used only to make your report, and they are never used for training. Unused uploads are deleted after 72 hours and paid work after 365 days, and you can delete sooner. Everything is hosted on AWS. The questions in the confidentiality section above apply here too: read the privacy policy and check your firm's policy and the client's terms before you upload.
You download an annotated invoice (CSV or Excel), a report with the entries flagged and the billed value of those entries (Markdown and PDF), and a rules file to reuse on the client's next invoice.
What it does not do: it does not connect to Clio, Tabs3 or TimeSolv, so you upload a file. It does not check rate caps or daily hour caps today. It does not decide what is billable, and it is not legal or ethics advice. It has been tested only on invented guidelines and invoices so far.
One invoice check costs 23 credits, with packs from $25 for 25 credits and no subscription. Create an account to try it on a draft invoice.
Questions
- Is it safe to paste a client's billing guideline into a chat assistant?
- That depends on the tool's terms, your firm's own policy and anything the client's guideline or engagement terms say about it. Check all three before pasting anything. If you are unsure, ask the client.
- Can a chat assistant replace the rule list?
- No. Whatever tool you use, a rule list with clause numbers is what makes flags traceable. A chat assistant can draft the list, but you should check each line against the guideline before relying on it.
- Why not just ask it to quote the clause for each flag?
- You should, but then you have to search the guideline for each quote, because a quote that looks right can still be worded differently from the source. A flag you cannot match word for word is not safe to act on.
- Does the same caution apply to other AI tools?
- Yes. The questions are the same for any tool: does it read everything, does it show the clause word for word, does it do the arithmetic correctly, and where do your documents go.
Catch the write-down before your client does.
Create an account, upload the guideline and an invoice, download the marked-up report.