Does AI Actually Save You Time?
Yes for some tasks, by much less than the vendor numbers suggest, and the saved time rarely becomes free time. The verification-value paradox, why the hours do not show up, and the four-step test to run before you believe any number - including ours.
- Doba čtení: 12 min
- Končí kvízem
Co tato kapitola pokrývá
- The Verification-Value Paradox
- Why the Saved Hours Do Not Show Up
- What Separates Those Who Get Time Back
- Three Conditions
- Measure It, or You Are Guessing
Kapitoly kurzu jsou napsány anglicky. Zbytek Academy je přeložen.
The short answer: yes for some tasks, by much less than the vendor numbers suggest, and the time saved rarely turns into free time. A course that sells legal AI owes you that sentence before any of the rest.
The verification-value paradox
The sharpest version of the case against comes from the academic Joshua Yuvaraj, who calls it the verification-value paradox arXiv · The Verification-Value Paradox, Oct 2025 (opens in a new tab). Because a lawyer answers for every word they put their name to, the value of a tool is never the time it took to produce the draft. It is that time minus the time to verify it. And on tasks where verifying properly means doing the traditional work anyway, that subtraction lands on zero or below.
The counter-argument is real too: lawyers already verify everything (junior memos, counsel opinions, expert reports) and nobody calls that a paradox. Both sides are right about different tasks, which is why "does AI save time" has no general answer and the only useful question is which of your tasks.
Research in an unfamiliar area is where the paradox bites hardest, and it is worth being blunt: removing a fabricated citation does not leave a competent brief behind. Nor is skipping the check an option. The ground truth there is Correctness Over Confidence, and the errors that actually survive review are not invented cases. They are real cases cited for a holding they do not support.
Why the saved hours do not show up
Suppose the tool works. Most lawyers still do not get a shorter week, for three reasons that have nothing to do with the model. Faster work raises what is expected of it. Scope expands to fill the time. And because prompting feels like conversation rather than labour, the work slips into the evening without registering as work.
In a law firm there is a fourth reason, and it is structural rather than behavioural: under the billable hour a lawyer cannot bill time they did not spend, so efficiency arrives as a pricing problem rather than as a gift of time. That is the argument of The Death of the Billable Hour, and it is why in-house teams and fixed-fee practices see the gain first.
In-house counsel face a cost that firms mostly do not: AI output arriving from everybody else. Executives send contracts drafted in a chatbot. Counterparties send AI-generated terms. The hours you saved on your own drafting go into correcting someone else's.
What separates the people who do get hours back
It is almost never the tool. It is whether the method behind the task was written down. The mechanism is the one chapters 15 and 16 already established from the other end: ask a general-purpose model to simply do a legal task and it will often fail, because you have given it nothing to be right about; give it your definitions, your positions and your format first and the same model often does well. The catch is the one from the previous chapter: for a task you rarely do, writing those instructions takes longer than doing the work.
Which gives the shape of the thing. The first run of a written-down method costs more than doing it by hand. The crossover is somewhere in the first few runs, and everything after that is the return.
Written once, then reused
Where the time goes, run by run, for one task of about an hour
| Phase of the work | 1st run | 5thbreaks even | 50th |
|---|---|---|---|
| By handFull price, every run, forever. | 60 min | 60 min | 60 min |
| Write it downThe checklist, your fallbacks, the shape of the answer. Once. | 200 min | none | none |
| Run and reviewPoint it at the matter, then check what came back. | 20 min | 20 min | 20 min |
| Saved so farBy hand, less the method, counted from the first run. | 2h 40m behind | level | 30h ahead |
A task you run twice never reaches the column that pays. A task you run weekly is past it inside six weeks, and 30h up by the end of the year.
Modelled, not measured: an hour by hand, most of a morning to write the method down, and twenty minutes a run to use it and check it. The crossover depends entirely on the task. A longer task or a shorter write up moves it earlier; a method still changing every month never gets there at all, because you pay the write up again each time it changes.
Three conditions
The task repeats
A task you do once a year never earns back the cost of setting it up. The gain is in the second, tenth and fiftieth run, which is why the unit of value is a method rather than a prompt.
The method is written down
An undocumented method produces a different answer every run, so you cannot tell improvement from noise. Writing it down is what makes the output comparable, and comparable is what makes it improvable.
Checking is faster than producing
This is the one that decides it. You can confirm a bad clause on sight; you cannot confirm an unfamiliar line of authority without doing the research. Where checking means redoing, there is no gain to find.
Measure it, or you are guessing
Almost nobody measures whether AI works on their own tasks. The headline numbers in this market are overwhelmingly self-reported: vendor surveys of their own customers, or polls asking lawyers how they feel. That is the same problem Practitioner-Led Benchmarks describes at market scale, arriving inside your own team. Few teams time the review. Fewer check the output against a standard.
Three measures answer most of it:
- Time end to end, including your review. Not the time to a first draft, which is the number vendors quote because it is the flattering one.
- Accuracy against a reference: the output scored point by point against work your team has already approved. Without a reference, "pretty good" is the most precise thing anyone can say.
- Errors by type: which checks get missed, which rules get misread, which sources go unused. This is the one that tells you what to change.
Once you have a reference, verification stops being a feeling and becomes a score, and a score is optimisable. Change the method, run it again on the same cases, keep the change only if the number goes up. That loop is the entire difference between a team that improves and a team with opinions.
Two honest limits, because this is the chapter for them. Automated checks catch mechanical errors: a quotation absent from the source, a citation that does not resolve, a step the agent skipped. They do not catch errors of judgment, and a real case cited for the wrong proposition is an error of judgment. Automation shortens the review. It does not move the responsibility, which is the same conclusion the chapter on delegation reached from the other direction. And on open-ended research, where most of the work is judgment, we think the skeptics are mostly right.
In practice
Run the test on one task before you believe any number, including ours.
The four-step test
- Pick one task you do every week
- Time it by hand, three times
- Time it three times with AI and a written procedure, review included
- Note every error you had to catch along the way
How to read the result
- Gap survives the review time → you found real time
- Gap disappears in review → you found where your judgment is the work
- Errors cluster by type → you found the gap in the method, not the model
If a vendor quotes you a time saving that excludes the lawyer's review, you have learned something about the vendor rather than the tool.
Chapter Quiz
Test your understanding of the concepts covered in this chapter.
Question 1 of 6
One task is checking a clause against the firm's standard positions. Another is researching an unfamiliar line of authority. Why does the chapter expect a saving on the first and not on the second?