An attorney fee audit is a sequence, and the order is the method. Run the twelve checks out of order and you produce a confident document that will not survive cross-examination. Run them in order and each step constrains the next, so by the time you reach an opinion on reasonableness you have ruled out the ways these analyses usually go wrong.
The order below came out of a working session where I re-ran a half-finished fee analysis produced with a general-purpose AI assistant and no legal training layer. A full day had gone into that draft. What was done all day could in reality have been done in about an hour. The speed is not the interesting part. The order is.
Every attorney fee audit starts before you open a document
First come the project instructions, and you always want to do your project instructions first. Two things go in every one of mine: the link to the matter folder, and the posture. Posture is what the tool is in this matter. On a mediation my instructions say this is a mediation, you are a neutral, do not make assumptions, do not add provisions. On a representation matter they say we are the advocate. On an audit, narrower still: the first report appeared largely accurate, the second did not, and the job is audit and quality control. Without that sentence you get a summary of what you handed over, and a summary of a defective report is a defective report with better formatting.
Second, and this is where most attempts fail, ingest the file rather than the work product. I pulled the earlier AI drafts out of the project and attached the whole case folder in their place.
Mine didn’t just go through what you gave the other tool. It went through your whole file.
That distinction is measurable. An earlier pass against partial data produced a total off by fourteen hundred twenty-five dollars, and I threw the number out rather than report it. A figure derived from an incomplete record is not a finding. It is an artifact of whatever was in the folder that day.
Build the ledger, then prove the invoices before you test the report
Third comes the ledger, entry level, not summary level. Every line item from every source invoice, transcribed into one table. Here that was four hundred twenty-four line items across twenty-five invoices. No abbreviated version works. A legal billing review built on category totals instead of individual entries can only tell you whether the report agrees with itself, which was never the question.
Fourth, before testing a single claim the report makes, validate the source documents. Every one of the twenty-five invoices tied to its own printed subtotal. Twenty-five out of twenty-five, no exceptions. That step locates the problem. The invoices are internally sound, so whatever does not add up is upstream, in what the report says the invoices owe. Skip it and every later discrepancy is ambiguous.
Re-derive every headline figure, and rebuild every exclusion category
Fifth, re-derive each headline number from your own ledger. Do not check the report’s arithmetic against the report. Compute each figure independently from the entries. Three headline figures went through that test and none of the three reconciled. Same discipline as running the income number twice in a support case, for the same reason. A number computed once has not been tested.
Sixth, take each exclusion category the report claims and rebuild it from the invoices. Two could not be reproduced at all. That gives you one testable standard, and it is the most useful sentence here: if you cannot reproduce the category from the invoices, it is not a category. It is a label. A fee reasonableness review that accepts labels it cannot rebuild is description calling itself analysis.
Then test the report’s reasoning against the record, and test it for double counting
Seventh, test the stated rationale against the record. The fee motion, the affidavit, the orders. Reasoning is where these reports are softest, because the arithmetic gets checked and the story never does. Here the record showed two summary judgment motions where the report’s reasoning assumed one. That falsified the rationale behind an entire year of entries. Not the entries, which were billed. The reason given for treating them as the report treated them.
Eighth, check for double counting, the check most fee work fails. Non-duplication gets asserted constantly and demonstrated almost never. With an entry-level ledger you can test it, and several entries here sat inside two or three exclusion categories at once. An attorney fee affidavit swearing to non-duplication with no line-level demonstration behind it is one question from trouble, and a competent opponent asks it.
Authority, billing rates, and your own weak points come last
Ninth, verify authority, and be precise about what verify means. Eleven state authorities came back flagged as unverified, delivered as a verification table rather than as citations woven into the text. It is taught to tell you to go verify the question in a research database rather than assert it, which is why I wrote a separate piece on a research library that refuses to do research.
Tenth, check rate consistency across the life of the case. Rates actually billed on the invoices did not match the rates sworn in the motion, and the discrepancy ran opposite to what the motion swore. Rate drift across a multi-year matter is ordinary. Rate drift that contradicts a sworn statement is not.
Eleventh, state the method’s own weak points inside the deliverable. These invoices were scanned, with visible character-recognition corruption in places. That goes in the report in writing, next to the countervailing evidence: the ledger reconciled to every printed invoice total, which is strong evidence the transcription came through clean. Then the rule that follows, also in writing. Every figure landing in the final report gets checked against the scanned source by eye.
Twelfth, quantify the recommendation against named non-duplicative categories rather than as a percentage off the top. A number tied to specific rebuildable categories can be defended entry by entry. A percentage cannot be defended at all.
What the attorney still owns, and why the value is memory rather than speed
An attorney fee audit run this way delegates no judgment. The appointment and the opinion are the attorney’s. Eyeballing every final figure against the scanned source is the attorney’s. Independently verifying every authority in a research database is the attorney’s. So is the exclusion recommendation, because reasonableness is an opinion and no software holds one. What the system contributed was holding four hundred twenty-four line items without getting tired at line three hundred.
Sometimes we have opinions about things, and then when you start looking at the details, you haven’t always tracked everything. And I love that it just tracks all of those details.
That is the reframe. We all form views about a fee dispute, and those views are usually directionally right and specifically unsupported, because nobody tracked all of it while the case was moving. The compounding version is better. As the case was going and things were landing in the matter folder, the ledger could have been building the whole time instead of in one sitting. The audit would have been a document that already existed the day the motion was filed, and the twelve steps a review rather than a reconstruction.
If you do fee work and want to watch the attorney fee audit order run on your own file structure, book a custom build strategy session at calendly.com/leisawintz/lal-custom-ai-strategy and bring one closed matter with the invoices in it. We build the ledger live. Nothing here is legal advice, and every attorney remains responsible for verifying their own authority and their own numbers before anything is sworn.