Home/ Journal/ Memo

AI for mid-market law firms, and the question of where the saved hour goes.

In a law firm the billable hour is the accounting unit, so an hour saved is revenue removed unless something else absorbs it. That single fact sorts AI for the profession better than any feature comparison, and it explains the outcome firms report most often: a pilot that worked, staff who preferred it, and financial statements that look the way they would have looked anyway. There are three places a saved hour can land, and only two of them pay. The first is recovered time, meaning work that was performed and never captured, or captured and written down before it reached an invoice. That pays immediately and needs no change to the fee model, because the firm is not billing less, it is billing what it already earned. The second is work whose fee is already set: flat fee, capped fee, contingency, subscription, or fixed-scope corporate work. There the revenue is fixed, so a lower cost of production falls straight to margin. The third is an ordinary hourly matter with no backlog waiting behind it, and that one does not pay. It converts realized revenue into idle capacity, and idle capacity becomes money only if demand is queued to fill it. A firm that cannot name the demand it will pour into the freed hour is buying a tool that makes its own invoices smaller. The buyer test follows: before signing anything, name the matter type and say which of the three destinations the hour lands in. If the answer is the third and no waiting demand exists, the right move is to fix intake and demand first, which makes the purchase premature rather than wrong. Two further facts decide most outcomes. The metric is realization and collection rather than hours saved, because hours saved is a vendor number and realization is one the firm already tracks. And partner compensation is the quiet blocker, since where origination and personal billable hours drive the draw, an individual partner can be personally worse off from a firm-level gain.

A managing partner weighing an AI purchase is being sold time. The pitch names a task, estimates the hours it consumes across the firm in a year, multiplies by a rate, and presents the product of that arithmetic as savings. The arithmetic is usually correct and the conclusion usually is not, because in a practice that bills by the hour those hours are not a cost avoided. They are the invoice. Everywhere else in the economy, taking time out of a process is the entire point of the exercise. Here it is the first half of a transaction whose second half nobody has specified, and the second half is what decides whether the firm ends the year with more money or with a well-liked system and flat financials.

MemoJuly 2026
Read time11 minutes
AudienceManaging Partners, Practice Group Leaders, Firm Administrators

The accounting unit decides the answer, and in a law firm the unit is an hour.

Most businesses can treat saved time as unambiguous progress. A manufacturer that takes an hour out of a process has lowered the cost of a unit it still sells at the same price. Revenue attaches to the thing sold, time attaches to the cost of making it, and only one of the two numbers moves.

A law firm billing hourly has no such separation. The hour is the cost and the hour is also the product. It is what the timekeeper is paid for and what the client is invoiced for, and the invoice is assembled out of the same units the efficiency project is removing. Take an hour out of the work and, absent some other change, you have taken it out of the bill.

That is the fact behind the outcome we hear described more than any other in this profession. The pilot worked, the associates preferred the new process, and the vendor's hours-saved figure was defensible. A year later the financial statements look the way they would have looked without it. Nobody misrepresented anything. The hours were genuinely saved; they landed somewhere that does not convert into money.

There are exactly three places a saved hour can land, and the difference between them matters more to a buyer than any comparison of models or platforms. It can recover work that was performed and never billed, fall inside a matter whose fee was fixed before the work started, or come out of an ordinary hourly matter with nothing waiting behind it. The first two pay. The third does not, and it is where most legal AI spending goes.

Everything below assumes the prior questions are settled: where the data runs, how outputs get reviewed, what the governance policy says, how to judge a vendor. Those belong on the law firm industry brief and are not repeated here. This memo takes the narrower question, the one no vendor can answer on the firm's behalf. When the hour comes back, whose revenue does it become?

Destination one: the hours that were worked and never billed.

The first destination pays soonest and requires nothing to change about the fee model, the compensation formula, or the firm's demand. The firm is not billing less; it is billing work it already performed and failed to collect on, which is the rare improvement inside a partnership nobody has a reason to argue against.

The mechanism is realization, and it leaks in two distinct places. Time is lost before it ever reaches a bill: the call taken in the car, the ten minutes on a document between two meetings, the Friday afternoon reconstruction of a week that a partner does from memory and rounds down out of conservatism. Time is also lost after it reaches a bill, at the review, where an entry whose narrative does not describe the work convincingly gets written down rather than sent out and defended.

Both leaks suit a system unusually well, for a reason that has little to do with how capable the models are. The evidence of the work already exists inside software the firm runs: calendars know who met whom, and document management knows which file was opened, edited, and sent, by which user, against which matter. The task is reconstruction and description rather than judgment, and the person best placed to check the draft is the person who did the work, who can do it in seconds because they were there.

The specific shape of such a build is written out in what we would commission first at a $30M law firm. What matters here is the accounting rather than the architecture: recovered time is new revenue against work already done, so it does not wait on a new client, a change in rates, or a partner accepting a different measure of their contribution.

The limit deserves stating plainly. Recovery is bounded by the size of the leak, and firms differ enormously in how large theirs is. A practice with disciplined contemporaneous entry and a billing partner who rarely writes down has less available here than one where entries get reconstructed weekly from memory. Which of those a firm is, its own reports already answer.

Destination two: the hours inside a fee that was set in advance.

Wherever the fee is fixed before the work is done, the relationship between time and revenue inverts. The revenue is decided. Time becomes what it is everywhere else in the economy, a cost of production, and lowering it moves straight to margin without altering the invoice the client receives.

More of a mid-market firm's book sits in this category than partners generally estimate. Flat-fee transactional work. Capped fees the firm agreed to in order to win an engagement. Contingency matters, where the fee is a share of a result and every hour spent reaching it is pure cost. Subscription and outside-general-counsel retainers. Fixed-scope corporate work. Panel and insurance defense rates negotiated down to a number the firm has to live inside. Any matter, in short, where somebody quoted a figure in order to be chosen.

The economics here are the cleanest of the three destinations. There is no demand to generate, no capacity to fill, and no conversation to have with the client. The engagement produces the same deliverable at the same price with fewer hours consumed making it, and the difference belongs to the firm. That is margin rather than volume, a less exciting story than growth and a more reliable one.

A second-order move opens up here, and it is where the strategic argument actually lives. A firm that can produce a defined matter type for materially less can quote a fixed fee on matter types it previously would only take hourly, which is a competitive position rather than a cost saving. Clients have been asking for that pricing for years, and what usually blocks it is the firm's own uncertainty about what a matter costs to deliver.

The complication is bookkeeping. Many firms cannot produce a list of which matters are effectively fixed-fee, because the arrangement lives in an engagement letter while the financial reporting treats everything as time and billing. Building that list is a records exercise, not a technology project, and it reshapes the decision more than any demonstration will. The same asymmetry between fixed and hourly pricing applies to how the build itself gets bought, and what each pricing model does to the incentives on both sides is worth reading first.

Destination three: the hour with nothing queued behind it.

The third destination is where most legal AI purchases land, and the market never describes it accurately, because describing it accurately would end the conversation.

The mechanics run in your head without a spreadsheet. An associate saves six hours a week on research inside ordinary hourly matters. The firm bills six fewer hours. The client notices the smaller invoice and is pleased. The associate's salary has not changed, so the cost side is identical. What the firm bought is a reduction in its own revenue, financed by a subscription, and the pilot report will still say the tool works, because it does.

The saved hour has become idle capacity, and capacity is worth something only when demand is queued to fill it. That is a question about the firm's pipeline rather than about the tool. Being busy is not the same as having work waiting: a firm can be exhausted and still have nothing queued, because matters arrive at roughly the rate they are completed.

There is a version of this that is a decision, not an accident. A firm may hand the efficiency to a client as a fee concession, to hold a relationship or win a panel seat, which is legitimate and sometimes correct. The difference is whether the partnership decided it or discovered it in the year-end numbers.

So the buyer test is one sentence long. Before signing anything, name the matter type the system will touch and say which of the three destinations the recovered hour lands in: unbilled time recovered, work already on a fixed fee, or hourly work with a backlog behind it. If the honest answer is the third and no backlog exists, the firm's constraint is demand rather than capacity, and capacity is the wrong thing to buy. Fix intake and demand first. That makes the purchase premature rather than wrong, which is the more useful conclusion, because it comes with a next step attached.

This is the least popular finding in a diagnosis conversation, and the one partners tend to still repeat a year later, because it is the part of the analysis that would have changed the outcome.

The number that settles it, and the formula that quietly overrules it.

Hours saved is a vendor's metric. No report the firm produces contains it and no comparison to last year is possible with it. Realization and collection are the firm's own numbers, tracked monthly by matter and timekeeper, and sitting in the practice management system before anyone signs a proposal. The baseline is therefore free, which is worth exploiting, because the most common reason an AI project cannot be defended afterward is that nobody wrote down the year before it. The general form of that problem is set out in how to measure the return on a mid-market AI engagement.

Read the numbers by matter type rather than firmwide. A firmwide figure blends the three destinations together and moves too slowly to attribute anything to. Realization on flat-fee corporate work and realization on hourly litigation answer different questions, and averaging them destroys the signal.

Then there is the blocker that never appears in a proposal. In most mid-market partnerships, origination and personal billable hours drive the draw. A firm-level efficiency gain that moves hours out of a partner's personal book improves the firm's economics and reduces the number that determines what that partner takes home. They are not resisting change. They are doing arithmetic, correctly, and reaching an answer the rollout plan did not account for.

That is the honest reason a good many legal AI rollouts stall a few months after a successful pilot. The pilot ran on curiosity and goodwill, both of which are free. The rollout runs against the compensation formula, which is not, and the formula wins every contest it is entered into. From outside it looks like adoption fatigue and it is nothing of the sort. The broader pattern of rollouts stalling around month four turns up across industries; inside a law firm the cause has a known address in the partnership agreement.

The fix is a compensation design question rather than a technology question, and it belongs to the partnership. Firms approach it differently, by crediting recovered time to the originating timekeeper, or by holding credited hours harmless while a system beds in. We are not the right people to redesign a comp formula. We are the right people to say that a build which leaves an individual partner worse off will not survive the year, whatever the pilot showed.

One category inverts the whole calculation. Every firm has work it declines or refers out because it cannot be done profitably at the rate a client will pay: the small matter not worth the intake cost, the document-heavy engagement that would consume an associate for a month, the client segment the firm has quietly stopped serving. There a saved hour cannibalizes nothing, because the revenue does not exist yet. It is the one case where every part of the arithmetic runs the same direction, and the least examined, because a declined matter leaves no record in the firm's systems.

The case for buying nothing this year, and what a managing partner can do this week.

We commission custom systems for a living, so read the following with the appropriate suspicion. It is also the section we would most want read, because the firms that buy at the wrong moment decide the whole category is a waste of money.

A genuinely backlogged hourly firm does convert saved time into revenue. The third destination is not always a trap. If matters are actually waiting on capacity, freed hours flow directly into billable work and the cannibalization argument does not apply. The question is whether the backlog is real or aspirational, and the test is whether you can name it. A real backlog looks like signed engagements not yet started, clients waiting past a date they were given, or matters declined last quarter with a client name attached. An aspirational backlog is a belief that the firm could get more work if it had more time. The first is a queue; the second is a marketing plan, and freed capacity will sit against it unused.

A firm mid-migration should do nothing this year. If the practice management platform or the document management system is being replaced, anything built against the current one is built against a moving target and the firm pays for it twice. Waiting a year costs nothing and removes the largest source of rework. Scoping against the systems of record rather than the workflow in the abstract is most of what the first phase of how we work does, and its honest output is sometimes a recommendation to wait.

For a single narrow task, check the stack you already pay for. Practice management platforms, document management systems, and the major research providers have shipped AI features into tiers many firms already own. If the need is generic to the profession, buying the tier is correct, and a custom version of a shipped feature is a maintenance liability. Run one calculation before defaulting to a subscription: what it costs across the full timekeeper count at renewal, since per-seat pricing at scale is where firms find they chose the expensive option without meaning to. The build versus buy decision turns on how specific the workflow is to your firm rather than on either sticker price.

Three things a managing partner can do this week, with no vendor involved, sorted by destination.

For destination one, pull realization and collection by matter type for the last four quarters rather than the firmwide average, and ask the billing partner where the write-downs concentrate and why. Thin narratives and reconstructed time point at capture rather than pricing.

For destination two, list every matter type on a flat fee, a cap, a contingency, or a fixed scope, with last year's revenue against each line. Most firms find the list longer than the partners' impression of it, and every line is a place where a lower cost of production is worth money immediately.

For destination three, take the hourly work and ask whether demand is queued behind it, then require names instead of a feeling. Signed and not started. Clients waiting past a date. Matters turned away in the last ninety days, and why.

Those three lists take an afternoon between a firm administrator and a billing partner, and they decide the question. If the first two carry weight, there is a build worth scoping and the constraint is capacity. If they do not, and the third comes back empty, the return this year sits in intake and pricing rather than software, and the AI decision keeps until next year at no cost. A firm that walks into a vendor conversation with the destination already named has supplied the one input no vendor can supply.

Field-note context

What we notice inside a billing practice.

The write-down happens after the work, which is why nobody argues about it.

Realization loss is rarely a decision anyone remembers making. It happens at bill review, quietly, when a partner reads an entry that does not describe the work convincingly and reduces it rather than sending it out and having a conversation with the client about it. The work was done. The client would most likely have paid for it. The narrative was thin, and thin narrative is cheaper to write off than to defend. Because the adjustment sits between the work and the invoice, it leaves almost no trace that anyone reviews: the timekeeper generally does not see it, the reports carry the reduced figure as though it were the whole story, and the firm's record of what it produced quietly disagrees with what it produced. Counting where those adjustments concentrate is a couple of weeks of work for a firm administrator, and it tends to reorder the priority list more than an assessment does.

An hour returned to an associate and an hour returned to a partner are not the same asset.

Savings models tend to run on a blended rate, which assumes one recovered hour is interchangeable with another. Inside a partnership they are not. An associate hour is close to fungible capacity: it can be pointed at whatever matter needs it, and its value is roughly the realized rate. A partner hour is usually claimed before it is freed, by origination, by management, by the client relationships only that partner can hold, and by supervision the associates need. Freeing it may produce business development rather than billable work, which can be worth more and shows up in a different period entirely. A firm that models the return at a blended rate and then staffs the pilot with partners has made two errors that point in opposite directions, and the net will be unreadable at the end of the year.

Nobody keeps a list of the work the firm turned away.

The one place where the economics run entirely in the firm's favor is work it currently declines, and it is also the place where the firm holds no data at all. A matter refused at intake leaves no record in the practice management system, no line in any report, and usually no memory beyond the person who took the call. So the opportunity that would justify a build most cleanly is invisible during exactly the conversation where it would count for something. The cheapest correction available is to start logging declines with one line of reason, beginning now and well before any purchase is contemplated. A quarter of that log tells a partner more about where the firm should be investing than most assessments do, and it costs one field on an intake form.

Extended questions

The questions partners ask once the hour is on the table.

Why do law firms report a successful AI pilot and no change in revenue?

Because in a firm that bills by the hour, saved time is removed revenue unless something else absorbs it. A pilot measures hours saved, which is a real number and the wrong one. Whether those hours become money depends entirely on where they land. If they recover work that was performed and never captured, or captured and written down before it reached an invoice, the firm bills more immediately, because it is billing what it already earned. If they fall inside a matter whose fee was fixed in advance, the revenue is unchanged and the cost of producing it falls, so the gain arrives as margin. If they come out of an ordinary hourly matter with nothing queued behind it, the firm bills fewer hours against the same cost base, and the pilot has reduced revenue while working exactly as advertised. Most pilots run on that third kind of matter, because it is the easiest place to demonstrate a tool, which is how a firm gets an honest success report and unchanged financial statements out of the same project.

Does AI reduce revenue at a firm that bills by the hour?

It can, and the mechanism is not subtle. The billable hour is the accounting unit, so an hour removed from a matter is an hour removed from the invoice unless something absorbs it. Three things can. Unbilled or written-down time being recovered, which raises revenue with no change to the fee model. Work whose fee is fixed in advance, such as flat fee, capped fee, contingency, or subscription arrangements, where the revenue is already set and a lower cost of production becomes margin. Or queued demand, where freed capacity is refilled immediately with matters that were waiting. Absent all three, the tool converts realized revenue into idle capacity and the firm is worse off in the year it buys. That is not an argument against buying. It is an argument for naming the destination first and choosing the matter type accordingly. Firms that get hurt here rarely chose it. They bought a general capability, deployed it wherever adoption was easiest, and never asked which side of the ledger the freed hour would land on.

Which work at a law firm gets the clearest return from AI?

Two categories, plus one most firms overlook. The clearest is work that was performed and never billed, because recovering it requires no change to fees, no new client, and no change to how partners are compensated. The evidence of that work already sits in the firm's calendar, document management, and communications systems, which is what makes reconstruction feasible and checkable by the person who did the work. The second is any matter type with a fee fixed in advance: flat fee, capped fee, contingency, subscription retainer, negotiated panel rate, fixed-scope corporate work. There the client's invoice does not change and the saving is margin. The overlooked category is work the firm currently declines or refers out because it cannot be done profitably at the rate a client will pay. Lowering the cost of production on work the firm turns away cannibalizes nothing, because that revenue does not exist yet, and it is the one case where every part of the arithmetic runs in the same direction.

How should a law firm measure the return on an AI system?

With realization and collection by matter type, not with hours saved. Hours saved is a vendor metric. It appears in no report the firm produces and cannot be compared against a prior year. Realization and collection are already tracked monthly by matter and timekeeper, which means the baseline exists before the engagement starts and does not have to be constructed afterward by the party with an interest in the answer. Read the numbers by matter type rather than firmwide, because a firmwide average blends work whose fee is fixed with work billed hourly and hides the only movement worth attributing to anything. Capture the before figures for the four quarters preceding any change, and agree in writing which matter types the system is expected to move and by when. A firm that skips this will spend the following year arguing about whether the system helped, and that argument gets settled by whoever is most invested in the answer rather than by evidence.

Why do legal AI rollouts stall a few months after a good pilot?

Usually because of the compensation formula rather than the technology. In most mid-market partnerships, origination and personal billable hours determine what a partner takes home. A firm-level efficiency gain that moves hours out of an individual partner's book improves the firm and reduces the number that partner is paid on. That partner is not resisting change. They are doing correct arithmetic about their own position, and no rollout survives that tension for long. The pilot did not surface it, because pilots run on curiosity and goodwill and both are free, while a rollout runs against the incentive structure written into the partnership agreement. The remedy is a compensation design decision that belongs to the partnership, such as crediting recovered time to the originating timekeeper, or holding credited hours harmless while a system beds in. It is worth settling before the build rather than discovering it in month four, by which point the firm has spent both the money and the enthusiasm.

How does the AI Maturity Index help a law firm sort this?

It runs the destination question without a call and without a vendor. The Index asks which process is worth investing in first, who touches it and how often, and where that process gets its inputs, which is most of what a firm needs in order to tell recovered time from fixed-fee margin from idle capacity. For a law firm the most useful output is usually the elimination rather than the recommendation. It surfaces the candidates that only free hourly capacity the firm has no demand queued behind, which are the ones to leave alone this year, and it sizes the ones that touch unbilled time or fixed-fee work. Ten minutes, no call, and the result is specific enough to bring to a partners' meeting alongside the realization report the firm already produces every month.

Not sure which of the three your firm's hours land in?

Start with the AI Maturity Index. Ten minutes, no call, and it names the one workflow worth investing in first, sizes who touches it and how often, and tells you whether the hour you are about to free becomes revenue or idle capacity before you talk to anyone.

Take the AI Maturity Index → Or book a diagnosis call →