Every law firm, accountancy practice, consultancy and agency of any size runs on the same arithmetic, whether or not anyone says it aloud. A partner sells work. Junior people do most of it. The firm bills those juniors at several times what it pays them, and that spread is the profit.
The technical term is leverage. In this setting, leverage means the firm makes money by selling junior hours at a multiple of their cost. Once that arithmetic is in place, the rest of the firm takes its shape from it.
That shape includes:
- the pyramid of headcount
- the training model
- the promotion ladder
- pricing
It has been a remarkably durable structure. It survived word processors, email, offshoring and the internet. The current moment is different because AI attacks the specific layer the arithmetic depends on. The issue is bigger than work getting faster. The junior layer that used to create the spread is the layer that gets compressed first.
The junior layer carries the economics of the firm
Look at what the junior tier actually does. The work is familiar across law firms, accountancy practices, consultancies and agencies.
- First-pass document review.
- Research and summarisation.
- Assembling a schedule from a pile of contracts.
- Producing the first draft of something a senior person will heavily mark up.
- Reconciling two datasets that should agree.
- Populating a model from a set of source documents.
Every one of those tasks has two properties in common. They require judgement, but bounded judgement, of a kind that can be checked. They also consume a large amount of time compared with their difficulty, which is exactly why firms could bill them in volume.
That combination is the sweet spot for current models. That is no coincidence. The tasks were already defined, already had a quality standard, and already had a review step built into the process. AI fits into work like that before it fits into work where the rules are unclear or the risk sits with one senior professional.
When hours fall, the old spread falls with them
If a task that took a junior twelve hours now takes them ninety minutes with a model and a review, the firm faces a difficult set of options. None of them preserve the old economics.
- Bill the ninety minutes, in which case revenue on that matter falls by roughly eighty percent and your cost base has not moved.
- Bill the twelve hours anyway, which is fraud.
- Keep the hours and fill them with other work, which requires demand to expand at the same rate your delivery capacity does, and it will not, at least not immediately.
- Stop selling hours entirely and sell the outcome, which is the only option that survives contact with an informed client, and it requires you to rebuild your commercial model from the ground up.
Most firms are currently in the second or third position and describing it as the fourth.
Clients can already see the time change
This is the part that makes the timeline shorter than firms expect. In-house counsel, finance directors and procurement teams are using these tools themselves. They know approximately how long a document review now takes, because they have tried it.
A bill for two hundred hours of first-pass review lands very differently on somebody who spent last weekend getting a reasonable answer in an afternoon. The client does not need to be right about the quality of what they produced. They only need enough confidence to ask the question, and they have it.
Once that question is asked, the conversation about hours has changed. The firm is now defending a price rather than describing a value.
The training model loses its bottom rungs
The hard part is training, and we do not have a tidy answer to it. Junior work was also how firms taught people. A lawyer who has reviewed four thousand contracts has a kind of pattern recognition that cannot be taught in a seminar. They acquired it by doing work that a model now does in minutes.
If you remove the bottom rungs of the ladder, you have a pipeline problem in about eight years, and the fix is not obvious. Reviewing a model’s output is a different cognitive activity from producing the work. It is faster, it is more comfortable, and it produces a shallower kind of expertise, because you are checking against a pattern you did not have to construct.
The firms thinking seriously about this are deliberately having juniors do some work the slow way, as training rather than as delivery, and absorbing that cost as an investment. That is expensive and it is honest. The firms leaving the problem alone will discover the gap when the current senior cohort begins to retire.
Fees are defended by judgement, accountability and risk
Strip out leverage and a smaller set of things remains. We should be precise about those things, because they are where the business goes.
- Judgement under genuine uncertainty. Deciding what to do when the rules conflict or are silent, rather than applying a known rule.
- Accountability. Somebody carries professional liability and can be sued. A model cannot be, and clients in regulated sectors need a name on the advice.
- Relationship and context. Knowing this client, this board, this history, and what they will actually do rather than what they say.
- The consequences of being wrong. Work where an error is expensive commands a premium precisely because someone is standing behind it.
None of those scale with headcount, which is the whole problem. They scale with the number of genuinely senior people you have, and that number cannot be increased quickly at any price.
Confidentiality has to be built into the system
There is a practical obstacle that firms hit immediately, and it is commercial only after it is technical. Client data cannot leak between matters, and it certainly cannot end up in a shared model. You do not solve that with a written policy alone. You solve it with a build decision made once, at the start.
Practically, it means matter-level isolation in whatever retrieval system you build, meaning the system that finds the source material given to the model. It also means tenancy boundaries, where each client’s or matter’s data is separated, and those boundaries have to be enforced by infrastructure rather than by a filter in application code.
The firm also needs a clear, documented answer to what happens to the data at the end of an engagement. It needs an audit trail of what the model was given and what it produced, because ‘we do not know what was in the context window’ is not an answer anybody can give a regulator. We wrote about the general shape of this in our professional services work.
Client work needs a traceable record
The second constraint is attribution. If a model contributed to something that goes to a client, somebody has to be able to say which parts, based on what sources, and reviewed by whom. Firms that skip this discover the problem the first time a piece of advice is challenged and nobody can reconstruct how a particular sentence came to be there.
This is manageable to build if it is designed in. Retrieval systems can cite. Generation can be logged with its inputs. Review can be recorded as an event with an author. Adding it afterwards fails, because by then the interesting material has already gone out of the door with no provenance attached.
Tools fail when they add steps
The most common failure we see in this sector has nothing to do with model quality. Professionals abandon tools that add steps. If the new workflow is not faster than the current habit on day one, it will not be used in week two, and no amount of mandate from a managing partner changes that.
The practical implication is that these tools must live inside the software people already use rather than beside it. A separate application with its own login, however good, is a tool people visit occasionally and then stop visiting. The successful deployments we have built are the ones where nobody had to change where they work.
We face the same economics ourselves
It would be dishonest to write this without acknowledging that we are a professional services firm describing the collapse of professional services economics. The same arithmetic applies to us. Software development has its own pyramid, and the same tasks are the ones getting faster.
We changed our model rather than waiting: fixed scope at a fixed price against a written definition of done, with the conventional cost stated beside it. It cost us revenue per engagement, materially, and we wrote about exactly what that cost on our engagement models page.
Volume and margin both rose, but that outcome was not guaranteed when we made the decision, and anyone telling you the transition is painless has not made it.
The pressure hits the middle tier first
Most commentary on this focuses on the graduate intake, which is the visible part. In the firms we have worked with, the pressure arrives one level higher, on the people three to six years in.
That cohort is expensive relative to a graduate, still below the level where they originate work, and their value has historically been that they can run a piece of delivery with light supervision and produce a competent draft without being told how.
That is precisely the profile the tooling substitutes for. A partner with a good retrieval system and a capable graduate can now do what previously required a manager in between, and the manager is the most expensive part of that chain to carry.
It is a genuinely difficult position for those individuals, because the skill they have spent six years acquiring is real, and it is being repriced through no fault of their own.
The firms handling this decently are being explicit about it rather than managing people out quietly. They are moving that tier toward client ownership earlier than the traditional ladder allowed, which is uncomfortable and is at least a direction. The ones handling it badly are holding utilisation targets constant and wondering why their best people are leaving.
Contracts are already asking sharper questions
The commercial conversation is moving faster than the technical one, and the questions arriving in engagement letters are already sharper than most firms are prepared for.
Clients are asking:
- Whether a model was used at all.
- Which parts of the deliverable it touched.
- Whether client material left a controlled environment.
- Whether the fee reflects the tooling, given that the client is paying for the firm’s efficiency gain rather than sharing in it.
That last one is the awkward one and it will not go away. If a firm invests in tooling and keeps the entire benefit, the client eventually notices and negotiates it back, usually less pleasantly than if it had been offered. If the firm passes all of it on, the investment has no return.
The stable answer is somewhere between, stated openly, and firms that put a number on it early will have an easier time than firms that wait to be asked.
We took the blunt version of this: state the conventional cost next to ours in every proposal, and decline the work if we cannot beat it meaningfully. It removes the negotiation entirely, at the price of removing our ability to be vague.
Start with the junior work that is easiest to check
If we were advising a firm on where to start, and we are occasionally asked to, the sequence would be this.
- Find the three tasks that consume the most junior hours and have the clearest definition of correct.
- Build an evaluation set for one of them, of a few hundred examples a person has already done properly, before building anything else.
- Only then build the tool, and measure it against that set on every change.
- Route anything below a confidence threshold to a person rather than guessing.
- Publish the accuracy internally, including when it is disappointing, because the alternative is a tool nobody trusts and everybody quietly works around.
The timing is directional, not tied to a specific year
We would not bet on a specific year, and anyone who does is selling something. What seems safe is directional: the spread on routine work compresses, the premium on genuine judgement widens, the pyramid flattens, and the firms that move first do it while they still have the profit to fund the transition rather than after the profit has gone.
The harder version is that this is a question about the business model, not only technology adoption. A business whose profit comes from selling time has to survive the time requirement falling by an order of magnitude. The honest answer is that it can, but not in the same shape.
More on the pricing side of this in how we price work, and on the sector itself in professional services.
