Chapters
The question in the room
Alejandro opens with a conversation he had with a partner at a thousand-attorney firm. A class of contracts that used to take two hours to review now takes 20 minutes. Multiply the hour and a half saved across every contract and every attorney who does that work and you are into tens of thousands of hours a year. So is the firm seeing millions more in revenue, or in profit? It is not. Clients are pushing for more AI, and the firm is not yet seeing the benefit.
The argument of the session is that the value is real. It is just not landing on the last line of the P&L, and the rest of the hour is about why, and how to move it there.
What the session covers
The timesheet was a capacity tool
The timesheet started with a Boston lawyer, Reginald Heber Smith, as a way to understand how many clients a firm could take on and how many people it needed. Only later did the ABA push lawyers toward selling their time in units, and the billable hour became the business model. Firms kept measuring the practice of law to the tenth of an hour and stopped measuring the rest of the firm. The full history is in the history of the billable hour.
Two economies inside every firm
The practice of law is billable and watched closely. The business of law (intake, conflicts, matter setup, time entry, billing, collections) is barely measured, and it hits both sides of the ledger. On revenue, firms leak billable work through time that was never captured and invoices that break a client's guidelines. On cost, 59 of every 100 net new hires in the industry are not lawyers. The hiring data is on the Data Hub.
Spending money to earn less
Most AI spend so far has gone to billable work, which under an hourly model means the firm pays to bill less. Fixed fees and subscriptions are the obvious fix, and Alejandro expects them to come, but clients and attorneys both resist them today. The near-term lever is the other side: take hours out of non-billable work and redeploy that capacity to billable work, where the change shows up immediately.
You cannot automate what you have not measured
Ask an operations leader which bottleneck they would remove first and there is rarely data behind the answer: how many hours a process takes, what it costs, and what it delays downstream. The end state is every back-office process with its own cost, so the ROI of automating it is known before anyone buys software. Part of why that is hard: a process like conflicts sounds like five steps on the happy path and turns out to have forty exceptions once someone shares their screen and walks through it.
Why copilots stall in production
Handing each team a general assistant feels magical in the demo, then accuracy drops once it meets every variation of the real process. And a copilot still leaves the person as the pilot. Attorneys should stay in the loop on legal work. Operational work, once its edge cases are captured, should run end to end, with a person notified only when something needs them. Otherwise the efficiency gain stays too small to reach the P&L.
Where to start: three questions
- Is it bounded? Is the process standardized in any way, or is every run different?
- Does it recur? Does it happen daily, weekly, or monthly, or is it a one-off request?
- Does it depend on legal judgment? Opening a new matter is routine. Deciding how much of a research effort to bill is not.
Once you know how often a process runs and how long it takes, the baseline cost is simple arithmetic, before counting any of the secondary effects. The method is worked through in proving AI ROI at a law firm.
What the live polls said
Two polls ran during the Q&A. Asked whether their firm is seeing a measurable return from AI, no one in the room said yes, and most said they are not measuring it at all. Asked where they would apply AI in the back office first, time entry came out on top. Alejandro's caution on that answer: fixing one step just moves the bottleneck. Speed up intake and conflicts become the constraint, speed up conflicts and time entry gets worse, automate time entry badly and invoices get rejected.
“You don't buy AI because it's AI. You should actually find the right use case and the impact that it will have in your firm, and then do a basic ROI analysis.”
Alejandro Castellano, Co-founder and CEO, Caddi
Next: see it built, live
The Master Class makes the argument. On Wednesday, October 7 at 1pm ET, the ILTA Product Briefing turns it into software: we build a working back-office agent live, on Box, from a workflow a registrant submitted. Register on LinkedIn. If you would rather read the argument, the longer version is in AI and the business of law and in Alejandro's Artificial Lawyer piece.