A client’s e-billing system rejects an invoice because one time entry is block-billed and another carries the wrong task code. Someone at the firm opens it, splits the entry, fixes the code, and resubmits. Nobody can bill for the correction, and payment on the whole invoice waits until it’s done.
That rejection is a record. It has a date, a reason code, and a dollar amount attached, and a firm of any size has a great many of them. The usual way of finding work worth automating never looks at them, because the usual way is to ask people.
Asking is sound advice in most industries. The people doing the work know where it drags. In a law firm the method runs into the revenue model. The firm sells hours, so a faster workflow can mean less revenue for the partner whose cooperation you need. I wrote in August about what that does to adoption numbers. It does the same thing to the search for use cases.
What an interview can’t collect
Partner compensation is tied to originations and billed hours. Changing an established legal process also carries professional risk, and that risk lands on the lawyer whose name is on the file. Ask that lawyer which of their workflows should take half the time, and the sensible answer is none of the ones that get billed.
That answer is a correct reading of the compensation plan. Training doesn’t change what the plan says.
The interviews still produce a list, and usually a long one. It ranks ideas by how interesting they sounded in the room and whether a model can do the task at all. Neither tells you who gains when the work gets faster.
I sent my own team a version of this argument recently, as we began our own discovery work. The point I most wanted to land was that a long list of AI use cases is the wrong thing to come back with.
The records say more than the interviews
A firm already produces evidence of where time goes without producing value. Billing data, realization reports, intake queues, rejected invoices, and document histories all record what happened. An interview records what someone believes should happen.
Each record points somewhere specific. Repeated write-downs on the same kind of work mean a partner has already decided, invoice after invoice, that the client won’t pay what the work costs. Poor realization on fixed-fee and alternative-fee matters marks the places where efficiency is worth money directly. Non-billable queues show the work that piles up before, between, and after the billable tasks. Recurring waits expose dependencies the formal process map leaves out.
Rework shows up the same way. Rejected invoices, engagement letters sent back by the risk team, formatting corrections, and documents with far too many revisions are all effort the firm performs and can’t charge for.
One kind of record deserves more attention than the rest: the personal spreadsheet, the template, the macro, the checklist, the private prompt library. Each one means somebody already tried to make a repetitive task easier, unasked, on their own time. That is a nomination, and it cost the person something to make.
Clients add their own evidence. Outside counsel guidelines, panel reviews, fixed-fee demands, and limits on which tasks can be billed all give a firm an external reason to redesign a workflow, and that reason arrives with the client’s authority behind it.
Where efficiency already pays
The most promising workflows are the ones where the incentive to get faster already exists. On a fixed-fee matter, fewer hours means more margin. In conflicts and matter intake, less delay means billable work starts sooner. Preventing an e-billing rejection brings the payment in earlier. Getting an engagement letter right before it reaches the risk team removes a correction loop nobody can bill for.
None of these asks anyone to give up hours that are worth money to them.
That puts the early opportunities around legal judgment and leaves its centre alone for now. Conflicts pre-screening, engagement letter assembly, RFP preparation, e-billing compliance, deposition summaries, privilege logs, closing binders, lease abstraction, client status reports, and filing-quality checks share a profile. The inputs are observable, the steps repeat, and the output can be checked against a reliable source.
“Draft the contract” makes a better demonstration and a worse first project. It sits inside the billable model, professional identity, privilege obligations, and legal judgment all at once. It will be in bounds eventually, and it’s a poor place to begin.
What the records can’t tell you
The records have a limit, and it sends you back to the people you stopped asking.
A write-down shows that the firm absorbed a cost. It doesn’t say why. The process may have been wasteful. The partner may have discounted to protect a relationship. A junior may have taken three times as long as the task deserved because that was the first time they had done it.
So the records choose where to ask, and then somebody still has to ask. The question has changed, though. “This kind of work was written down again and again last year. What happened?” is about one thing that already occurred, and a partner can answer it without volunteering any of their own hours. “What should we automate?” never offered that.
The third explanation is the uncomfortable one. Work that is routinely written down and regularly named as a source of junior burnout looks like the easiest call on the list. It is often also how juniors learn the underlying skill, a problem I spent part of an earlier essay on. A firm that removes that work has taken on a training problem along with the saving, and should say so before the build starts.
The test
When a candidate workflow turns up, whether a model can do the task is only the first question. Four others decide it: whether efficiency benefits the person who controls the workflow, whether the problem shows up in operational evidence, whether it happens often enough to matter, and whether the output can be verified.
Then come two governance questions: how heavy the burden of running it responsibly will be, and what happens on the day the system and the human reviewer both miss the same error.
All of that compresses into one test. Where is the firm already absorbing the cost of delay, rework, write-downs, or repetitive non-billable work, and where could a better workflow recover it without competing with legal judgment?
The answer is a short list, and it’s supposed to be. Every item on it has an owner who gains and an output somebody can check. Those are also the conditions for learning to operate one of these systems in production.
Drafting the contract can wait. When that conversation comes, it will go better for a team that already has a workflow in production, a record of how it was reviewed, and at least one partner who came out ahead the last time.
The August essay on why adoption numbers mislead under hourly billing is the companion to this one. Subscribe to AndrewLewisWasHere if you’d like the next piece.