Most legal teams have several processes that could be redesigned with AI: a weekly report, the same questions from the business each month, a contract type that arrives every few days. The first one you choose decides how much time you get back and what your colleagues see you deliver.
Our view is that the first choice should be a process someone senior already wants fixed, with a result you can count. We tested that advice against the research on who is given the work that does little for promotion and on how the use of AI is credited, and we say below where that research stops short. The case for redesigning the process itself is in a separate article: Why AI pays off only when the work is redesigned.
Who is asked to do the tasks that do little for promotion?
Women are asked to take on more of this work, and they say yes more often. A 2017 study in the American Economic Review looked at what its authors call tasks with low promotability: tasks everyone would prefer someone else to do, such as writing a report or serving on a committee. It found that women, more than men, volunteer for these tasks, are asked to volunteer, and accept when asked. Beliefs that women say yes more often appeared to be an important driver of the difference [1].
Lawyers report a similar pattern. In a survey of more than 2,800 US lawyers, published in 2018 by the American Bar Association's Commission on Women in the Profession and the Minority Corporate Counsel Association, the level at which white women reported doing more administrative tasks than their colleagues, such as taking notes, was 21 percentage points higher than for white men. For women of colour it was 18 points higher than for white men [2].
Who is given the visible work also affects promotion. A 2024 study of one international financial institution, using data on over 10,000 investment projects, found an early-career promotion gap for women, and that assignments to lead project teams were crucial in explaining it. Women also received slightly less credit for team performance. Once lead-role assignments and credit were taken into account, the remaining gap was too small to be statistically significant [3].
In a legal team, this is the gap between keeping the matter list and running the matter.
How is the use of AI seen and credited?
Similar gaps appear in the use of AI at work and in how it is credited. In a Lean In survey of 1,015 US adults published in June 2026, 18% of women who had used AI at work had been praised for it, against 23% of men. Asked about their managers, 30% of women said they were encouraged to use AI, against 37% of men [4]. A review of 76 sources covering over 300,000 people found that, weighting each source by its sample size, 39.3% of women used generative AI against 47.8% of men [5]. In an experiment with 1,026 engineers who rated identical code, reviewers rated an engineer's competence 9% lower on average when told AI had helped, and the penalty was more severe for women [6].
Exhibit 1
Women use AI at work less often, and are praised and encouraged less when they do
Women and men, percent. Praise and encouragement: 1,015 US adults, 2026. Use: weighted average across 76 sources.
Praise is among those who had used AI at work. Sources: Lean In [4]; HBS AI Institute, on Cranney, Delecourt and Koning [5].
From these findings we infer that a woman lawyer who uses AI privately may save time nobody sees, and that colleagues who do see her AI use may judge her work more harshly. A measured change to a process a senior person owns gives colleagues a result they can check. We found no study that tests this inference for lawyers.
Redesigning a process carries its own risk: the redesign can become one of the tasks nobody asked for. In Microsoft's 2026 survey of 20,000 people who use AI at work, only 13% said they are rewarded for reinventing work with AI even if results are not met [7]. So most of those surveyed did not report a reward for a redesign that falls short, which is why the fourth test below requires an owner who wants the result.
13%
Rewarded for trying
AI users rewarded for reinventing work with AI even if results are not met. Microsoft, 20,000 people [7]
28%
Track time to resolution
In-house legal leaders and legal operations professionals. ACC and Everlaw, 284 respondents [8]
12%
Track return on technology
Same survey: those who tracked the return on investment from their technology. ACC and Everlaw [8]
Five tests for a candidate process
- Hours per year. Multiply how often the process runs by how long it takes each time.
- Whether you can measure it now. You need a starting figure: hours spent, time from request to reply, or errors found. In an ACC and Everlaw survey of 284 in-house legal leaders and legal operations professionals, 28% tracked time to resolution and 12% tracked the return on investment from their technology [8]. Figures from before and after the change are evidence that most teams in that survey lacked.
- Risk and confidentiality. A candidate stays in only if you can build and test the redesign with fictional data, and a lawyer can check each output before it leaves the legal team. A score of 1 on this test rules the candidate out whatever its total. Singapore's Ministry of Law guide on generative AI (March 2026) states that "Users must safeguard client data and confidentiality" [9].
- An owner who wants the result. Name the person who receives the output or answers for the process, ideally someone senior to you, and confirm they want it improved. The redesign then becomes work someone asked for, and that person can confirm the result.
- Whose repetitive work it removes. Prefer a process where the time saved comes from copying, chasing, reformatting or answering the same question again, and where that work falls to the same one or two people each week. The studies above found that this kind of work goes more often to women, so removing it can free their time for work that counts towards promotion.
| Test | Scores 1 | Scores 2 | Scores 3 |
|---|---|---|---|
| Hours per year | Under 50 | 50 to 150 | Over 150 |
| Measurable now | Hard to define a measure | A record could start this week | Figures already exist |
| Risk and confidentiality (a 1 rules the candidate out) | Needs real client data, or errors would be hard to catch | Fictional data works; outputs take effort to check | Fictional data works; outputs are quick to check |
| Owner | Nobody has asked | An owner exists but has not raised it | A named person wants it fixed |
| Repetitive work removed | Little | Some | Mostly copying, chasing or repeating |
Add up each candidate's five scores. If two candidates are within two points of each other, prefer the one that scores higher on the owner and repetitive work tests. Lawyers already triage this way when deciding which matter gets attention first.
How do the tests sort three candidates?
Take a fictional four-lawyer team at a regional logistics company. Every figure here is invented. Ana, one of its lawyers, lists three candidates.
A. The weekly contract status report. Each Friday Ana collects the status of open contracts from emails and sends a table to the finance director: about 3 hours a week over 48 working weeks, or 144 hours a year. The finance director has said the report often arrives late.
B. Sales questions about templates. The sales team asks about 50 questions a month on which contract template to use, at about 20 minutes each, or about 200 hours a year. Nobody records the questions, and the head of sales has not raised the issue.
C. The regulatory change paper for the board. Ana writes it twice a year, at about 20 hours each, so 40 hours a year. It needs legal judgement throughout, and an error in AI-drafted text could be hard to catch before it reached the board.
| Test | A: status report | B: template questions | C: board paper |
|---|---|---|---|
| Hours per year | 2 (144 hours) | 3 (200 hours) | 1 (40 hours) |
| Measurable now | 3 | 2 | 1 |
| Risk and confidentiality | 3 | 3 | 1 (ruled out) |
| Owner | 3 | 1 | 3 |
| Repetitive work removed | 3 | 3 | 1 |
| Total | 14 | 12 | 7 |
The status report wins on the tests, although another candidate saves more hours. Candidate B, the template questions, would save the most hours, but it has no owner and no record, so a good result would have nobody waiting for it. Candidate C, the board paper, scores 1 on risk, so it is out. Ana chooses candidate A, the status report: the finance director wants it fixed, its time is easy to record, the build can use fictional contracts, and the work it removes is mostly copying. She starts logging the sales questions now, so they have a starting figure if they become her second project. The board paper needs her legal judgement and the board reads it, so she keeps writing it herself and puts the hours saved on the status report into it.
How do you make the result visible?
- Record the starting figures before changing anything. For the status report, record the hours spent each week and the time the report is sent, for four weeks.
- Agree the measure, a review date and the upkeep with the owner. For example: the report is compiled in under 1 hour and sent by 12 noon every Friday, reviewed after eight weeks. Agree who keeps the new process running, so its upkeep does not become a standing task for you.
- Decide where the saved time goes. In BCG's 2026 survey of 11,749 workers, 66% of frontline employees who use AI regularly said they get limited or no guidance on what to do with the time it saves, and more than half do not move that time into strategic work [10]. Unassigned time can be used up by more of the same requests, so agree with your manager that it goes to named work, such as the board paper.
- Report the result in the terms the owner uses. The finance director cares when the report arrives and whether it is right, so report those figures under your name, beside the starting figures, and describe the tools after them.
This is the same discipline as a good file note: what was done, by whom, and with what result.
Where the skeptics are right
The practical advice holds without the gender findings: a process with many hours, a clear measure, low risk and an owner is a sensible first choice for anyone. A reader who doubts the gender research loses little by following the five tests. The research gives women lawyers a reason to weight the owner and repetitive work tests more heavily, as the grid's tie-break does.
The evidence also has limits we want to state plainly. None of the studies cited tests whether choosing a measurable redesign improves how a woman lawyer's work is recognised. That conclusion is our inference from separate findings. The ABA and MCCA figures were published in 2018, cover the US, and come from a news report of the survey. The financial institution study covers one employer, the engineer experiment involved no lawyers, and the Lean In, Microsoft and BCG surveys cover workers generally. All the surveys are self-reported, and Microsoft and Everlaw sell technology.
What we'd ask
If you lead a legal team:
- Notice who in your team does the copying, chasing, note-taking and repeat questions each week, and favour first redesigns that remove that work.
- When someone brings you a redesign, agree the measure, a review date and who keeps the new process running, so its upkeep does not become a standing task for them.
- Agree where the saved time goes, and name the work it moves to.
If you are choosing your first process:
- Score each candidate on the five tests, rule out any that scores 1 on risk, and use the tie-break on the owner and repetitive work tests.
- Record the starting figures for four weeks before you change anything.
- Report the result under your name, in the terms the owner uses, beside the starting figures.
Why we built the course
The first process a lawyer redesigns shapes how her colleagues see the work, and the choice is easier with a structure and a group doing the same. We built AI Women in Law so that more women lawyers choose work that counts and show the result. Over five weeks, participants redesign one piece of their team's work and build a first version of a tool for it. See the course.
References
- Linda Babcock, Maria P. Recalde, Lise Vesterlund and Laurie Weingart, Gender Differences in Accepting and Receiving Requests for Tasks with Low Promotability, American Economic Review 107(3): 714 to 47, March 2017.
- Law Gazette (John Hyde), Poll reveals 'endemic' racial and gender bias in US legal profession, reporting the ABA Commission on Women in the Profession and Minority Corporate Counsel Association survey "You Can't Change What You Can't See", 25 September 2018.
- Çağatay Bircan, Guido Friebel and Tristan Stahl, Gender Promotion Gaps in Knowledge Work: The Role of Task Assignment in Teams, EBRD Working Paper No. 291, April 2024.
- Lean In, New research: Women use AI less often at work and get less credit, 5 June 2026. 1,015 US adults surveyed in March 2026.
- HBS AI Institute, The Gender Divide in Generative AI: A Global Challenge [Update], 6 August 2026, summarising Katelyn Cranney, Solène Delecourt and Rembrand Koning, "Global Evidence on Gender Gaps and Generative AI Over Time", HBS Working Paper 25-023, May 2026.
- Oguz A. Acar, Phyliss Jia Gai, Yanping Tu and Jiayi Hou, Research: The Hidden Penalty of Using AI at Work, HKU Centre for AI, Management and Organization Working Paper 2025-02, 1 August 2025.
- Microsoft, 2026 Work Trend Index, Agents, human agency, and the opportunity for every organization, 5 May 2026. 20,000 AI users in 10 markets.
- ACC and Everlaw, The Role of Generative AI in Proving Corporate Law Department Value, press release, 17 November 2025. 284 respondents.
- Singapore Ministry of Law, Launch of Guide for Using Generative Artificial Intelligence in the Legal Sector, 6 March 2026.
- BCG, AI Is Reshaping Jobs Faster Than Companies Are Reshaping Work (AI at Work 2026), press release, 3 June 2026. 11,749 workers in 14 markets.