In a firm that sells hours, the first measurement question is not how many hours AI saved. It is which hours.
An hour cut from billable drafting on an hourly matter is revenue you no longer invoice. An hour cut from a non-billable internal memo is margin you keep. Same stopwatch, opposite sign, and almost no return model written for professional services separates the two.
That omission is why measuring AI for professional services goes wrong in a way it does not at a software company. The hours-saved number is easy to produce and sometimes negative in disguise. Below is the method we use with firms of 50 to 500 people whose unit of sale is time: where to measure first, which numbers you already hold, and the two results that should end the spend.
An hour saved has a sign
Take a matter type where research and first-draft work runs six hours, and a trained associate now does it in five. At a $300 billed rate and the 88% realization rate Clio reports as the 2025 law-firm average, you have removed about $264 of invoiced work from that file.
Across four hundred matters a year, that efficiency gain is a six-figure revenue cut wearing a stopwatch. Nothing about the training failed. The hour had nowhere to go.
Now move the same hour. The associate writes a non-billable pitch summary or an intake recap in half the time. That hour carried employer cost and no invoice line: compensation in professional and business services ran $59.80 per hour worked in the second quarter of 2026 by BLS reckoning. The saving is worth at least that, and a billed rate if anyone resells the time.
So before a figure gets multiplied by a rate, split every claimed saving into hours that were invoiced and hours that were not. The second bucket is a gain on contact. The first is only a gain if you can say where the hour went. The general method for showing that capability moved is in our guide to proving an enablement program worked; this page is the version for a business that sells the thing AI consumes.
Start with the hours nobody sold
Clio’s 2025 benchmarks put average law-firm utilization at 38%, about 3.0 billable hours inside an eight-hour day. Apply the same source’s 88% realization and 93% collection rates and roughly two and a half hours of that day become collected revenue. Five and a half do not.
That unsold majority is where to start. The sign is never in doubt, because every minute returned on the non-billable side is cost removed with no conversion story required. It is also where hands-on fluency shows up first: drafting, summarizing and first-pass reading on a person’s own documents, which in a services firm means scoping notes, pitch material, status updates and intake summaries. Those are the tasks people do badly and late because nobody bills for them.
The baseline already exists, too. Your practice-management system has collected non-billable time entries for years, so nobody can accuse you of a measurement built to flatter the program. Pull non-billable hours per person per week by category for the twelve months before training, then for the twelve weeks after. One table, one existing source, no new dashboard.
The conversion test comes before the spreadsheet
A saved billable hour has three destinations. Which one applies is a demand question, not a training question, and it decides whether a return model is worth building at all.
| Where the hour lands | This works when | The number that proves it | What kills it |
|---|---|---|---|
| Refilled with new billable work | You are turning work away or have a real backlog | Billable hours per person flat or up while hours per matter fall | A pipeline that cannot absorb capacity this quarter |
| Kept as margin on fixed-fee work | Part of the book is flat fee, capped or contingent | Effective hourly rate rises on the same fee | A book that is entirely hourly |
| Lost | Neither of the above | Hours per matter fall, revenue falls with them | Assuming a saved hour converts itself |
The third row is the common one, and the sector’s own numbers show why. In the 2026 report on the US legal market from the Thomson Reuters Institute and Georgetown Law, average firms posted 13% profit growth and worked rates grew 7.3%, growth the report attributes to pricing rather than to doing the work differently. A market that has spent three years converting rate increases is not practiced at converting capacity. Capacity nobody redirects is the standard failure of a license-led rollout, which is the case made in why buying licenses is not the same as building capability.
Answer the conversion question in front of the people who own the pipeline, before anyone opens a spreadsheet. If the honest answer is “we could not sell another forty hours this quarter,” the gain is margin on unsold time. Book it there instead of calling it revenue.
Four numbers your time system already holds
Firms keep being sold new AI enablement metrics. The four that convince a partnership are the ones it already argues about.
| Metric | What it tells you about the program | The trap |
|---|---|---|
| Utilization rate | Whether recovered non-billable time became sellable time | Rises if people work longer, so hold hours worked constant |
| Realization rate | Whether AI-assisted time survives client review | A quiet fall means clients are writing down the hours you billed |
| Effective hourly rate | On fixed-fee and capped work, the cleanest signal there is: same fee, fewer hours, higher rate | Needs good time capture on work nobody bills hourly, where capture is worst |
| Write-downs and work-in-progress age | Whether faster first drafts shorten the cash cycle | Moves for a dozen reasons, so read it as support only |
Effective hourly rate deserves the emphasis. If any part of your book is flat fee, capped or contingent, start there. The fee is fixed, so fewer hours drop straight into margin and the arithmetic needs no assumption about pipeline. A firm with a third of its revenue on fixed fees has a clean experiment available today and usually does not know it.
Those four feed the return-on-investment case a finance team will accept, and our walkthrough of a finance-ready model sets the inputs out line by line.
The hours-saved claims in circulation are too big
Search this topic and you will be told firms recover 15 to 25 hours per person per week and lift utilization 12 to 18 points. Those pages carry no sample size, date or source. Here is the same question with evidence attached, sorted by the weight each tier deserves.
| Tier | Figure | Source |
|---|---|---|
| Measured | Mean saving of 5.4% of work hours among people using AI at work, 1.4% across all workers | Bick, Blandin and Deming, NBER working paper 32966 |
| Self-reported | About 60 minutes per employee per day, among nearly 600 accounting professionals of whom 98% use AI daily | Karbon, State of AI in Accounting 2026 |
| Predicted | 5 hours per week in the coming year, about 240 hours annually, roughly $19,000 of value per professional | Thomson Reuters Future of Professionals Report, 26 June 2025 |
| Claimed | 15 to 25 hours per week, utilization up 12 to 18 points | Vendor and agency blog posts, no sample disclosed |
The measured figure, 5.4% of a forty-hour week, is a little over two hours. The quoted five hours is a forecast respondents made about themselves, and the $19,000 attached to it is that forecast priced at a billed rate. Useful as a direction, neither is a result, and a partner who finds you presented a prediction as a measurement will not trust your next number.
Expect the adoption-to-revenue gap too. Clio’s 2026 survey found 71% of solo practitioners and 75% of small firms using AI, while only about a third reported any revenue increase. That is the normal outcome, not a failure signature. Our benchmarks for what good looks like at six and twelve months set the bands for use and depth; this page covers turning use into money inside a pricing model that punishes speed.
Your clients are setting this number in parallel
While you measure the gain, your clients are deciding how much of it they intend to keep.
Everlaw and Wakefield Research surveyed 300 in-house attorneys and legal operations professionals, publishing on 20 October 2026. Among those who engage outside firms, 60% expect lower fees or fewer billed hours when AI is used and 62% expect clear disclosure that it was. Eighty-one percent expect outcome-based pricing to pass the billable hour within five years, 47% within three. Fifty-four percent say AI has already cut the volume of work they send out.
The contract language is further behind, which is your window rather than your reprieve. A September 2026 review of 1,054 outside-counsel guidelines found only 20 that mention AI at all, 12 requiring disclosure, three expecting savings passed on as fewer hours, and none dated before 2024. The clause is coming; it has not reached most engagement letters.
Accounting is working the same arithmetic out loud. Bloomberg Tax reported in March 2026 that an AICPA survey of more than 1,000 firms put the median net hourly rate up nearly 7% over two years to $170, with the report urging firms to move faster away from hourly billing. An RSM partner in the same piece put the question plainly: is the time saving all margin for the firm, or all price reduction for the client? Decide your split before a client picks it for you, and make sure the measurement supports the half you keep.
What a defensible before-and-after looks like
Six steps, in order, none needing new software.
- Pick one matter type, not the firm. One recurring piece of work with enough volume to produce a median.
- Pull the baseline from time entries. Twelve months, median hours by phase. A firm that sells hours holds the best baseline data of any business trying to measure this, and most never use it.
- Split the target work into invoiced and non-invoiced hours. You now have two experiments with different success conditions.
- Train the people who do that work, on their own documents. A hands-on session runs roughly $2,000 to $15,000 in the market by group size and scope. At a $170 to $300 realized rate, the top of that range breaks even on 50 to 90 recovered hours, under five hours each across twenty people. Write that threshold down first.
- Run eight to twelve weeks, then report four things: median hours per matter phase, non-billable hours per person per week, effective hourly rate on fixed-fee work, and realization. Keep assumptions on a line of their own.
- Ask every trained person to name the task that changed, in one sentence. The answers predict the second year better than any dashboard. “The first-pass review memo takes 40 minutes instead of two hours” survives a partner meeting. “Everyone loves it” will have no defender at renewal.
Two results that mean stop
Write both down before the program starts, because a target with no failure condition is a hope with a percentage attached.
Hours per matter unchanged after a follow-up session. One pass can miss. If a second pass aimed at the work people are stuck on still moves nothing, the constraint is not fluency and more sessions will not find it. Our piece on diagnosing an enablement push that did not land separates the failure modes that look alike from a distance.
Hours down, revenue down, and no fixed-fee book to catch it. The result nobody plans for: the training worked, hours fell, the firm invoiced less. That is a pricing decision wearing the costume of a measurement result, and the fix is a conversation about which share of the book moves to a flat fee, not another training day.
Frequently asked questions
How do you measure AI gains when the firm bills by the hour?
Split every saving by whether the hour was invoiced. Non-billable time recovered is a straight cost gain, read from existing time entries. Billable time recovered is a gain only if it is refilled with new work or banked as margin on a fixed fee. Track that side as effective hourly rate on flat-fee matters and billable hours per person on hourly ones.
What is the single best metric for AI enablement in a professional services firm?
Effective hourly rate on fixed-fee, capped or contingent work: fees divided by hours worked. The fee does not move, so fewer hours drop straight into margin, and the math needs no assumption about whether your pipeline can absorb freed capacity.
Why did our hours fall but revenue not rise?
Because a saved hour does not convert itself. Refilling it needs unsold demand you can bill this quarter, and most firms do not have it on the timescale the efficiency arrived. With no backlog and no fixed-fee work you created unsold capacity rather than revenue, and the honest place to book that gain is cost.
How many hours a week should we expect AI to save?
Plan on roughly two hours per week per regular user and treat anything above five as an outcome to be proved. The best measured evidence puts the mean saving at 5.4% of work hours among people using AI at work. The five-hours figure in circulation is a forecast, and the 15-to-25-hour claims carry no sample size.
Should we tell clients we used AI, and should we discount for it?
Disclose, and decide the discount deliberately rather than by default. Of 300 in-house respondents surveyed in October 2026, 62% expect disclosure and 60% expect lower fees or fewer billed hours. Most engagement letters require neither yet, so you still choose your split. That will not hold for long.
Can we use utilization rate as the headline number?
Only with hours worked held constant, because utilization also rises when people stay later, which is the opposite of the result you wanted. Report it beside total hours worked, or a partner will find the loophole before you do.
What does a program like this cost, and when does it pay back?
Hands-on sessions run about $2,000 to $15,000 in the market by group size and scope, with custom automation work in the $25,000 to $150,000 range. At a realized rate of $170 to $300 an hour, a mid-range session clears its cost on a few recovered hours per participant. Set that hour threshold as the break-even figure at the start rather than arguing it afterward.
Does this work the same way for consultancies and agencies as for law firms?
The method is identical and the labels differ. Swap matter for engagement, realization for recovery, and the fixed-fee point gets stronger, because retainer and project work is already most of an agency book. That is the cleanest version of this experiment.
Key takeaways
- An hour saved has a sign. Invoiced hours removed are revenue gone until something refills them; non-billable hours removed are margin on contact.
- Start where the sign is unambiguous. Average utilization of 38% leaves roughly five and a half hours of an eight-hour day already unsold.
- Answer the conversion question before building a model: can you sell another forty hours this quarter, or is there fixed-fee work to bank the margin on? If neither, book the gain as cost.
- Effective hourly rate on flat-fee work is the cleanest signal in the building, and most firms with a fixed-fee book never run it.
- Two hours a week per user is the measured figure, five is a forecast, 15 to 25 is marketing. Keep the three apart.
- Write the stop conditions down first: no change in hours after a second pass, or hours down with revenue down and no fixed-fee book to catch it.
If you want this run against your own matter types, time data and fee mix, book a free AI readiness call. We will tell you which part of your book gives the cleanest measurement, what gain is realistic for the work your people do, and whether your constraint is fluency or pricing.
Arthur Wandzel
