AI Doesn't Save Time, It Creates Organisational Choices.

Most of the discussion about legal AI eventually reduces to a single measurement. Agentic systems, assistants, workflow automation and better prompting all promise the same underlying thing, which is time returned to the lawyer. Vendors quantify it, firms build business cases around it, and the debate that follows is mostly about whether the 20% number is real.

I want to accept the number rather than argue with it. Suppose the optimistic case holds and within a few years, AI genuinely gives every lawyer in a firm something like a fifth of their week back. The drafting is quicker, the first-pass review is quicker, the research is quicker and the technology does more or less what was promised.

The question I find more useful here is not when that time appears, but what happens to it once it does. That turns out to be an organisational question rather than a technological one and it's answered way less often than it's asked. The honest default answer, the one the history of professional work keeps offering... is that the time goes nowhere in particular.

Where the time has gone before

The profession has already lived through several productivity 'revolutions' and none of them left much visible slack behind. Email is the clearest case. It collapsed the cost of correspondence to almost nothing, and the effect was not calmer afternoons but a client expectation that a reply arrives within hours rather than days. The speed became the norm, and the norm became the obligation. Nobody now thinks of email as the tool that gave lawyers their time back, though by any narrow measure of minutes per letter, it did.

Online research followed a similar path. When authority lived in printed reports and physical libraries, the depth of an analysis was bounded by what could reasonably be found in an afternoon and perhaps what you could carry to the desk. Once everything became searchable, that fell away and the expected thoroughness of the work expanded to occupy the space that had opened up. Advice that once cited a handful of cases now surveys the field, partly because it can and partly because the firm on the other side will.

Then document automation raised the volume of documents rather than lowering the hours behind them. Practice technology in general made firms faster at nearly everything while leaving them no less busy than they were two decades ago. Economists gave this pattern a name long before software existed. Jevons noticed in the 1860s that more efficient coal engines increased total coal use rather than reducing it, because efficiency made the underlying activity cheaper and demand rose to meet it. Efficiency tends not to produce spare capacity, just ends up producing more appetite.

If AI returns a fifth of the week and the firm responds by expecting a fifth more output, everyone settles back at full utilisation and the accurate description of what changed is: the same work, produced faster.

Constraints move, they do not disappear

There is a second reason the time-saving frame can mislead, and it has less to do with economics than with how systems behave. Any process, whether a production line or a piece of litigation, moves at the pace of its tightest constraint. Speeding up a step that is not the constraint does not speed up the whole. It just moves work more quickly towards wherever the real bottleneck sits, and lets it pile up there instead.

Now I think this is worth keeping in mind when imagining faster drafting. If AI compresses the time to a solid first draft, drafting stops being the constraint, and the constraint reappears somewhere further along the line. Partner review becomes the place work queues. Client decision-making, never fast, starts to look like the slow part. Governance steps, sign-offs, matter administration and knowledge management all become newly visible as friction, not because any of them got worse, but because everything around them got quicker and left them exposed.

Manufacturing worked through this decades ago and arrived at an uncomfortable conclusion, which is that optimising individual steps is mostly wasted effort until you know where the system's true constraint lies. Much of the current legal AI effort is step-level optimisation carried out with real enthusiasm and comparatively little attention to the system those steps belong to. The likely result is not a firm that is uniformly faster, but a firm that discovers, one bottleneck at a time, where its delivery was actually fragile. A firm that has only ever measured lawyer productivity will find it has few instruments pointed at the places that suddenly matter.

What software engineering did with its productivity

Software engineering is the nearest thing to a natural experiment here, because it went through this transition a decade or so ahead of law. Engineering output per person rose dramatically over that period through better languages, shared frameworks, cloud infrastructure and then more recently: AI assistance. What is interesting is what the stronger engineering organisations chose to do with the gain.

They did not, on the whole, simply ask each developer to ship proportionately more. A good share of the new capacity was turned back into the system that produces the work. Teams were formed whose job was to make other engineers more effective: internal tooling, automated testing, deployment pipelines, monitoring, documentation. Some of the most capable engineers stopped shipping customer facing features almost entirely, on the reasoning that a person who improves the platform raises the output of everyone who builds on it, while a person who ships a feature raises only their own.

That was a deliberate move of capacity away from production and towards the capability to produce, and is precisely the move a billable hour firm has rarely had the room to make.

The legal similarity is not hard to see once you look for it. It resembles legal engineering treated as a discipline rather than a curiosity and looks like workflow design done on purpose, governance built by people who understand the matters the system is touching, evaluation that checks whether an AI output is genuinely reliable for a specific task rather than merely plausible and knowledge capture that turns each matter's hard-won learning into something the next matter can reuse. None of that work generates a bill in the month it's done, but the entire value is that it compounds.

Capacity to produce, or capacity to design

This points towards a different way of describing what AI hands a firm. It may not be creating legal capacity so much as organisational design capacity, which is a scarcer and more awkward resource to hold.

Firms have long lived with a structural bind that is easy to describe and has proved very hard to escape. The people who best understand how the work actually behaves, where matters tend to go wrong and where the real judgement sits, are the same people whose time is most expensive and most fully booked. Asking them to redesign a workflow or build a knowledge system has meant asking them to do it in the margins, after hours, or not at all. The redesign then tends to fall to people further from the work, who produce something plausible rather than something accurate, and the gap between the two is exactly where these efforts usually fail.

AI may loosen that bind, not by doing the design itself but by creating enough room for the people who ought to be doing it. A senior lawyer with a genuine slice of the week back is, perhaps for the first time, positioned to reshape how their practice delivers work while still being close enough to that work to reshape it well. Expertise and available attention rarely coincide in the same person at the same time in a law firm.

As agentic systems take on more of the execution, the workflow itself, as in the encoded judgement about how a matter should move, what gets checked and where a human steps in becomes the thing of lasting value. Capacity to do that design work is not a luxury in that world, but more it's capacity to build the asset that ends up mattering most.

Consuming capacity, compounding capability

A simple distinction makes the choice easier to reason about, with any hour of AI-created, the capacity can be either consumed or compounded.

Consumed capacity flows straight back into production: more matters, higher utilisation, faster turnaround, more billable output, finance teams happy. This is the path of least resistance, because every existing incentive points to it. It improves this quarter's figures, it fits the current model without translation and it requires no decision from anyone. Consumption is just what happens when nobody makes any choices.

Compounded capacity goes into the firm's ability to deliver rather than into delivery itself: better workflows, governance that reflects how the practice actually runs, evaluation datasets that grow more useful with every matter, knowledge assets that make the next similar instruction cheaper, training that lifts the whole team, a closer read on what clients are really trying to buy. None of this appears in the utilisation figures, but all of it shapes what the firm can do in three years.

The honest issue here is that the billable hour measures the first kind of value precisely and the second kind... not at all. A firm that consumes everything becomes a faster version of its present self. A firm that compounds even a portion becomes slowly (but surely), a different and more capable organisation, because the instruments only register consumption, compounding has to be chosen deliberately or it does not happen.

Now obviously I'm wary of recommending a ratio, some tidy figure for how much capacity ought to be reinvested, because the right split depends on the firm, its clients and what it's ultimately trying to become. What does seem safe to say is that the split should be a decision rather than just happening, and that a few questions tend to help towards it.

The first is simply where AI created capacity in a given firm will actually go, and who has the standing to direct it. Then the second is what becomes the constraint once drafting, review and research all speed up, and whether the firm is ready for that constraint to sit in partner time or client responsiveness or governance rather than in the work itself. A third, and the one I keep thinking about is related to protection: capacity that is not deliberately held aside is reabsorbed, so which of the most experienced people should have part of their week ringfenced for improving how the work is done rather than doing more of it.

If lawyers return to full utilisation after each AI advance, then I'd say it's worth asking what has actually changed...