Imagine the managing partner of a 25-person consulting firm reading the FT over a quiet coffee. The piece is about Allen and Overy putting Harvey, a legal AI platform, in front of 3,500 lawyers across 43 jurisdictions. She gets to the end and the obvious question lands. None of this is her firm. The numbers are not her numbers. The vendor is not her vendor. Is there anything in the story that is actually useful at her scale?
There is, and it sits underneath both the tool and the scaffolding around the tool. The portable piece is the discipline.
Four global law firms have gone public with their generative AI rollouts in detail. Allen and Overy first, from November 2022, with Harvey across the network. Slaughter and May firmwide, governance and training caught up before the wider release. Hengeler Mueller firmwide across Europe, all practice areas. Eversheds Sutherland setting up a dedicated AI innovation function, with around 40 per cent of US lawyers using the platform. Four different firms, four different markets, four different starting points. The same four-part shape under the public stories.
This post is about reading those four-part rollouts honestly. The principles port to firms a hundredth the size. The scaffolding does not. Telling the two apart is the work.
What does the law firm AI rollout pattern actually look like?
The published version is messy and the press loves the big numbers, but the operating shape is consistent. Each firm picks a small set of bounded internal use cases first, mostly first-draft generation and due-diligence triage. Each firm builds the governance layer before widening access. Each firm frames the tool as augmentation. Each firm treats the rollout as a multi-year project, not a quarter.
The headline at Allen and Overy was the 3,500-lawyer figure across 43 jurisdictions. The underlying choice was the use cases the firm did not pick first. The technology touched drafting and review well before it touched anything client-facing or advisory. That sequencing is the lesson, not the headcount. Slaughter and May made the same call. So did Hengeler Mueller. Eversheds Sutherland built a dedicated innovation function around it, with around 40 per cent of US lawyers using the platform, and made governance the prerequisite for access rather than an afterthought.
Why does it matter for an owner-managed firm?
The same four-part discipline ports cleanly to a 20-person practice. The shape survives translation because it is about how a professional services firm absorbs a new tool without breaking its supervision model, not about how a magic-circle firm scales an enterprise contract. An accountancy practice rolling out an AI drafting tool for management reports faces the same four questions a global law firm does.
Which bounded job first? What supervision rule sits around it? How is the team told it works? And how long are you giving this to mature? The British Chambers of Commerce reports 54 per cent of UK firms now using AI in some form, with 95 per cent of those firms saying it has had no impact on workforce size in the past year. The pattern is augmentation in practice, not just rhetoric. That matches the legal pattern more than the press narrative does.
Where does the law firm pattern actually meet your business?
It meets you at four specific decision points. The first is the use case shortlist, pick one or two bounded internal jobs where the worst-case error is a wasted hour, not a misled client. The second is the supervision rule, one sentence per use case that names who checks the output. The third is the augmentation script the team hears from you, said out loud, in the room, more than once.
The fourth is the calendar you give yourself. Three years to maturity is the legal market’s number. Two to three at your scale is realistic. Three months is the wrong unit, and a common source of disappointment.
A working example. Take a single recurring document a senior person in your firm spends six hours a week on. Pick one. Pilot one tool with one team member for six weeks. Write the supervision rule before you start. Widen from there if it holds, and only if it holds.
When to copy the law firm move and when to ignore it
The principles travel. The scaffolding does not. Top-tier firms have CPD departments, dedicated innovation functions, professional-responsibility officers, and procurement teams that can run a six-figure vendor selection. None of that exists at owner-managed scale, and importing the structure wholesale is a fast way to spend money on the wrong things. Harvey is not the tool for a 20-person practice. The named-vendor selection is the scaffolding, not the principle.
The lighter version is honest about what is left after the scaffolding is stripped away. Bounded use case, written supervision rule, augmentation framing, multi-year horizon. Four things, all of them free apart from your time, none of them dependent on a six-figure platform. Orgvue’s research that 78 per cent of AI projects stall or fail points at the same gap, almost always missing one of those four.
Related concepts to read alongside this post
The four-part rollout discipline sits alongside three other ideas worth holding in view as you plan. The first is professional supervision under AI, which the Law Society’s libguide on AI lays out in concrete terms a 20-person accountancy or consulting firm can adapt. The second is the bounded use case as a unit of work, well illustrated by Freshfields’ contract review case study.
The third is the failure side of the same coin. The hallucination cases (Mata v. Avianca and the rest) are what bounded use cases and supervision rules exist to prevent. Read them alongside the rollout stories, not separately. The big-firm rollouts and the courtroom horror stories are two sides of the same coin, and the four-part discipline is what keeps you out of the second pile.
The British Chambers of Commerce SME adoption survey is the reality check on what is actually happening in firms closer to your size, 54 per cent using AI, 95 per cent reporting no workforce impact. The Harvey case notes on Slaughter and May, Hengeler Mueller and Allen and Overy are existence proofs that this can work at scale, less useful as blueprints.
The work, at any scale, is to take the principles and leave the scaffolding. The win for an owner-managed firm is to apply the four-part discipline at the scale you actually run, in the language your team actually uses, against jobs your firm actually has on the floor.
If you want a peer pair of eyes on which bounded job to pick first and what the supervision rule should look like, Book a conversation.
