What AI has not changed in a Magic Circle workflow

Every few weeks another elite firm announces a platform-wide AI rollout. Slaughter and May adopted Harvey in April. The Magic Circle has put serious money into legal AI over the past two years. The announcements are real, the spend is real, and the press coverage writes itself.

So here is a question with a more awkward answer: what has actually changed in the way the work gets done?

One number from our own data

We code every deal announcement we publish into DealDatabase, including every lawyer named on the deal team. Across 3,034 announcements from two comparable windows, July to December 2024 and June to September 2026, the average number of lawyers named per deal went up, from 9.3 to 10.5.

If AI were compressing deal teams, the naming convention is where you would expect to see it first, because firms have every incentive to name everyone who touched the file. It has not happened. Announcements also got more detailed in general, with disclosed values rising from 55.6 to 63.1 per cent of releases, so some of this is a PR trend rather than a staffing one. It is not proof that teams grew. It is reasonable evidence that the visible team has not shrunk.

What has not changed

Who carries the risk. A partner still reads the clause that matters and still signs the advice. No tool has moved that line, and none of the rollouts claims to.

The billable hour. The BARBRI research published in August found it is still the most stubborn barrier to adoption, and that training alone does not fix it. A tool that saves four hours on a document review is a tool that removes four billable hours, and the person asked to use it is usually the person whose hours are counted.

The second read. Every firm that has rolled out a drafting assistant has also added a verification step. The work moved, it did not disappear: from writing the first version to checking a version somebody else’s model wrote.

The training pipeline. The same research found that no firm has yet built the AI competency framework its associate pipeline needs, and that firms grade their own rollouts a C. The junior work that used to teach judgement is the work most exposed to automation, and nobody has replaced the teaching.

The timetable. Signing dates are set by regulators, lenders, counterparties and boards. None of those has got faster.

What has changed

First drafts, research triage, document review at volume, and the speed of finding the one precedent you half remember. That is a real gain and it compounds. It is also, so far, a gain inside a process whose shape is unchanged.

The honest summary: the tools have changed the first hour of a task. They have not yet changed who is accountable for the last hour, or how the firm charges for either.

What to watch instead of the announcements

• Leverage. If AI is really changing delivery, associate to partner ratios move. Watch that, not tool licences.
• Fee structure. Watch for fixed fees appearing on exactly the work AI touches. That is the first sign a firm believes its own efficiency claim.
• Trainee intake. A sustained cut, not a one-year dip, would say more than any platform announcement.
• Who gets named. If deal teams start getting smaller in announcements, something real has shifted.

None of those four has moved yet in a way you can see from the outside.

Sources: BARBRI research on law firm AI rollout and measurement, August 2026; Artificial Lawyer on Slaughter and May’s adoption of Harvey, April 2026; DealDatabase, 3,034 coded deal announcements. The team-size figure counts lawyers named in published announcements, which is a reporting convention, not a staffing record.