A law firm that drafts with AI should have each junior lawyer do a fixed part of their core drafting without it, and have a senior lawyer grade that unaided work every month. In a three-month controlled trial of patent lawyers, those with fewer than seven years of practice drafted better while they had AI, then showed no average gain over juniors who never had it when they later worked without it.
The trial was released in September 2026 as a National Bureau of Economic Research working paper, a series the NBER says has not been peer reviewed. It randomly split 133 practicing patent lawyers at eleven U.S. intellectual property law firms two to one: 90 drafted with a custom AI assistant, 43 worked without it as the control group, and blinded expert patent attorneys scored all the work. Assisted drafts beat the control group's at 10 days and at 90 days, with larger gains for juniors, though the authors report that the junior-senior gap is not statistically significant.
After three months, the 91 lawyers still in the trial redlined an existing patent application without AI, which the authors call a core task of patent practice requiring expert judgment. Those who had used the assistant beat the control group, but this advantage was concentrated entirely among senior lawyers. Junior lawyers showed no average gain: fewer of them had mediocre scores, and more had poor or good ones. The authors conclude: "The largest gains from AI thus accrued to the lawyers who retained the least."
In the Thomson Reuters Institute's Future of Professionals survey, run in February and March 2025 and published that June, the 1,363 legal professionals among its 2,275 respondents expect to free up nearly 240 hours per year, up from 200 in 2024. An unaided share gives up part of that time saving on the drafts it covers. Where a client's billing terms will not cover the slower draft, the firm writes the extra hours off.
The one takeaway the authors call unambiguous is that assessing the effect of AI assistance on skill development requires evaluations that separate the contribution of the software from the expertise of its user. So set the unaided share and have a senior lawyer grade it against the firm's own standard, blind where the group is large enough. Grades go on a monthly sheet by lawyer, beside the grades on that lawyer's assisted work and the hours written off.
In a corporate group, the share might be one first markup of a purchase agreement a month per junior, graded by the senior associate or partner who would review it anyway. Three rounds give a first look at whose unaided work trails their assisted work most, and those juniors draft more without the assistant for the next three rounds.
For a PE-backed legal services business with lawyers in several offices, the same sheet in every office shows what the unaided rule costs in written-off hours and whether its future seniors can do the work without the tool.
The authors say their sample is not representative of professional services generally, and they state their conclusion as a possibility: foundational expertise may be a prerequisite for extracting durable skill from AI-assisted practice. Google paid the direct costs of the experiment. The assistant was an unreleased Google Labs product, all eleven firms have ongoing patent drafting relationships with Google, and five of the paper's seven authors are Google employees. A sixth discloses a paid contractor relationship, and the seventh acknowledges support from the Google Technology and Society Visiting Fellows Program among other funders.
