Law firm AI savings: clients want a number the timesheet can't produce

Clients now want law firm AI savings in writing. According to Financial Times reporting in early September, Citigroup asks law firms competing for its work to explain how much they save through AI. Morgan Stanley plans more competitive bidding and more fixed fees. Goldman Sachs is looking at how AI efficiency should show up in its legal bills. On 26 September the New York Times DealBook put the question in a headline: clients are asking “Where’s my discount?” What nobody in that fight says out loud is that most firms can’t answer. They aren’t hiding the savings. They never measured the work the AI changed.

TL;DR: Law firm AI savings are hard to quantify because time entries record hours against a matter and a billing code, not the steps inside the work, and clients are now asking for the number anyway. When AI takes the first pass on document review, the entry still reads “review/analyze”, only shorter. A blanket discount doesn’t fix that. It hands the client a number that isn’t attached to anything. The firms that come out ahead will map one repeatable matter type step by step, time it before and after, count the verification work, and price it on a fixed fee they can defend.

Clients are asking. Firms are mostly not answering

The best current read on the firm side is BigHand’s 2026 Legal Pricing and Budgeting Trends Analysis, published on 14 September. It surveyed more than 800 law firm leaders, pricing professionals and finance managers across the UK and North America. The gap is right there in the numbers.

What BigHand found (2026, 800+ respondents)	Share of firms
Report growing demand for AI-driven efficiencies or transparency around AI use	56%
Report increased client demand for financial transparency	51%
Report improved profitability per matter from AI	31%
Keep existing pricing and use AI efficiency to improve profitability	30%
Have adjusted pricing structures for AI-augmented work	29%
Track matter budgets against actuals continuously through the matter	23%
Say lawyers most commonly default to hourly rates	35%
Say partners are uncomfortable discussing AI with clients	35%

Read the middle of that table as a single sentence. About a third of firms say AI made matters more profitable, and about a third changed nothing in how they price. That isn’t greed. Under hourly billing, the efficiency either stays with the firm as margin or disappears as lost hours, and nobody wrote down where it went.

The last row is the one I’d pay attention to. Partners are uncomfortable talking about AI with clients because they’re being asked to defend a number they don’t have.

The timesheet records hours, not steps

Look at what a time entry actually is. A matter, a timekeeper, a duration in tenths of an hour, a free-text narrative. For clients on e-billing, a task code and an activity code from UTBMS, the Uniform Task-Based Management System. The American Bar Association and the corporate counsel association helped create its litigation code set in 1995 to give clients and firms meaningful cost information.

In the litigation code set, L320 Document Production covers identifying and reviewing documents for production and reviewing them for privilege. The activity list runs from A101 Plan and prepare for to A112 Travel. A104 is Review/Analyze.

Now picture an AI first pass in that workflow. Say a model reads a review set overnight and flags a few thousand documents as possibly privileged, and an associate spends the next day checking the flags. The time entry says L320, A104, 7.5 hours. A year earlier, a similar matter might have produced L320, A104, 60 hours spread across three people.

The hours dropped. That’s all the record can tell you. It can’t tell you which step went away. It can’t separate the tool’s effect from a matter that was simply smaller. It doesn’t show that “review” now means checking a model’s judgment rather than forming your own, which is a different job with its own error rate. Nothing in A101 through A112 tells reading a document apart from checking what a model already flagged in it.

So when Citi asks for the savings, the honest answer from most firms is a comparison of totals between matters that were never alike. It’s the same problem anywhere work is billed or budgeted by the hour. The system of record was built to describe cost, not work. You can’t pull a process measurement out of it after the fact, because the process was never recorded.

A discount is the wrong answer to the right question

If you can’t isolate the saving, you negotiate a percentage. Ten off, fifteen off, whatever the relationship partner thinks the client will accept. Both sides lose something. The client gets a number that isn’t tied to any step, so there’s no way to know if it’s fair. The firm pays for the tools and gives up part of the gain without knowing whether it can afford to, and it will have the same argument again next year.

The clients are asking the right question in the wrong form. “How much did AI save?” really means “which parts of this work are cheaper now, and why am I still paying for them by the hour?” That question has an answer, but only at the level of steps.

I’d put it the way I put it for any automation project. A saving you can’t point to a step for is a negotiating position, not a number. The only reason I can defend the 71+ hours a week the PDF-parsing automation took out at HomeWorks Energy is that the work it replaced was a countable step: documents in, data out, a known volume. Legal work is messier than that, but not as much messier as partners think. Privilege review, NDA markup, diligence checklists and deposition summaries are repeatable enough to count.

What measuring it actually looks like

This is process work, and it’s the kind I do in services engagements outside law as well. None of it requires a new platform:

• Pick one matter type. High volume, repeatable, already partly on AI. First-pass privilege review or NDA review are good candidates. Don’t start with bet-the-company litigation.
• Write down the steps as they happen now, not as the practice manual describes them. Who touches the document, in what order, and what they decide at each point.
• Time a baseline on a sample of real matters, by step. Twenty or thirty matters is enough to see the shape.
• Tag AI-assisted steps separately inside your own time system. Keep submitting L320 and A104 to the client’s e-billing platform, but carry an internal sub-code underneath so you can tell first-pass review from verifying a model’s flags.
• Count verification and rework as steps. If checking the model’s privilege calls takes a senior associate four hours, that’s part of the price. Firms that leave it out will quote savings they can’t deliver.
• Price the matter type on a fixed fee built from the measured steps, and show the client the step map rather than a discount.

The fourth item is the one that fails quietly. Internal sub-codes drift the same way clause labels do. One team tags “AI-QC”, another “verify”, a third doesn’t tag at all, and six months later the data can’t be compared. I wrote about that failure in legal AI agents and clause taxonomy drift. The fix is the same: one definition per code, owned by one person, with entries that don’t match rejected at submission rather than cleaned up at month-end.

For the in-house side of the table

If you’re the client, “how much did AI save” is a question that rewards the firm that makes up the most confident number. Ask for the step map instead. Which steps in this matter type are done with AI now, which are still done by a lawyer, and how is the AI’s output checked? A firm that can answer that can give you a fixed fee it will stand behind. A firm that can’t answer it has no real savings figure to give you, only a discount.

The step map also answers a question that matters more to a bank’s general counsel than the discount does: whether the firm’s AI use is under control. A firm that knows which step the model does, and who checks it, has control. A firm that only knows the hours went down doesn’t.

Microsoft found the same thing inside its own business this month. The gains in its Frontier Playbook came after a team mapped and simplified the work, not from the licenses. Law firms are about to find out that clients want that redesign done before the tool, and they want to see it on the invoice. If your firm, or your legal department, wants to start with one matter type, that’s the conversation to have.