Measure the work that is ready to use
The time to produce a draft is only one part of a legal assignment. Source review, corrections, formatting, and lawyer approval can consume more time than generation. A useful comparison measures the same finished work to the same quality standard, with and without AI.
The calculation below is illustrative. Its times are invented to explain the method; they are not measurements from Jonathan Nessler’s practice, a client matter, or a model benchmark.
Define a comparable assignment
Consider an internal case brief prepared from one public opinion. Both versions must identify the procedural posture, holding, disposition, material qualifications, and source paragraphs. Both must pass the same lawyer review. If the assignment also requires subsequent-treatment research, include it on both sides and hold its scope constant.
Choose that standard before running the comparison. A polished draft that omits a limiting condition has not reached the same endpoint as a verified brief. The case-briefing walkthrough shows concrete omissions to look for.
Worked example: drafting is faster, review still takes time
These numbers are illustrative minutes of active human work. The AI column includes time spent directing the tool; unattended generation is tracked separately below.
| Active work, in minutes | Without AI | With AI |
|---|---|---|
| Read sources and establish the assignment | 25 | 25 |
| Prepare the first draft, including AI instructions | 30 | 5 |
| Verify facts, legal propositions, and source references | 15 | 20 |
| Correct and format the draft | 10 | 15 |
| Final review and approval | 10 | 10 |
| Total active human time | 90 | 75 |
In this example, first-draft work falls by 25 minutes, from 30 to 5. Additional verification and correction use 10 of those minutes. The net saving is 15 minutes: 90 minus 75. Dividing 15 by the 90-minute baseline gives a 16.7% reduction in active human time.
The result depends on completing all five stages. If the AI draft needs another 20 minutes of corrections, its total rises to 95 minutes. The same assignment would then take 5 minutes longer than the 90-minute baseline. A fast first draft can coexist with a slower completed assignment.
Track elapsed time separately
Suppose the illustrative AI run also spends 3 unattended minutes generating text. Active human work remains 75 minutes. If every stage happens in sequence, elapsed time is at least 78 minutes, before queueing or other delays. If the lawyer does other work while the model runs, record the overlap rather than counting the same minute twice.
An agent working on five matters at once may shorten waiting time without reducing review effort on any individual matter. Record both the time people spend and the time from assignment to approved delivery. Those numbers answer different staffing and service questions.
Use a small, repeatable time log
| Record | Why it belongs in the comparison |
|---|---|
| Assignment and acceptance criteria | Establish that both methods produce comparable work. |
| Source set and complexity | Avoid comparing a short, familiar opinion with an unfamiliar record. |
| Tool, model, and date | Make the observation specific to the tested setup. |
| Active minutes by stage and role | Include the lawyer or staff member who reviews and repairs the draft. |
| Unattended minutes and elapsed time | Separate waiting from human effort. |
| Defects and additional passes | Count omitted qualifications, unsupported citations, and rework. |
| Final quality and approval | Exclude unfinished drafts from claims about completed work. |
| Setup and software cost | Account for configuration, templates, training, and operating expense. |
Use several comparable assignments and keep failures in the record. Alternate the order of methods or use matched tasks where feasible: someone who has already read an opinion may finish a second attempt faster because they know the case. Record that limitation rather than presenting a second pass as a clean experiment.
Report the number of assignments, their scope, the median and range of completion times, and how many met the acceptance criteria. Do not average away an assignment that was abandoned or required a full rewrite. Its cost belongs in the evaluation.
Convert saved time into a business decision carefully
Saved minutes create capacity. Whether that capacity becomes revenue, shorter delivery times, lower costs, or less overtime depends on how the practice uses it. Subtract setup and operating costs before claiming a financial return. Use actual internal costs by role for a cost analysis; an hourly billing rate is not the same thing as the firm’s labor cost.
For a simple break-even calculation, divide one-time setup minutes by net active minutes saved per completed assignment. With an illustrative 180-minute setup and 15 minutes saved per comparable assignment, the time investment breaks even after 12 assignments, before ongoing costs. If net time saved is zero or negative, there is no positive time-based break-even under those assumptions.
A time study is not a billing rule. ABA Formal Opinion 512 addresses fees when lawyers use generative AI, including time actually spent under an hourly arrangement. Fee agreements, applicable rules, and client requirements govern billing; hypothetical manual hours should not be treated as hours actually worked.
A pilot that answers a useful question
Begin with public opinions or other material authorized for the tool. Define the deliverable, record the work through approval, and decide in advance which defects require a second review or abandonment of the AI draft. Compare the completed results, not just the drafting screen.
The practical question is whether this workflow produces acceptable work with a better use of people’s time. A record of the sources, defects, corrections, and approval makes that conclusion easier to assess and repeat.
