The adoption race in legal is effectively over. In ILTA’s 2026 technology survey, 94% of member firms said they were using or exploring generative AI, up from 80% a year earlier.1 Microsoft 365 Copilot alone is in use at three-quarters of them.1 If AI tools were the differentiator, nearly every firm would now be differentiated.
They are not. Deployment is shallow: in ILTA’s 2025 survey, most Copilot and Harvey users were still piloting.7 Thomson Reuters found that 64% of law-firm professionals whose organisation has a clear AI strategy say AI meets or exceeds expectations; where there is no strategy, that falls to 29%.2 Clients are watching the gap. In the same research, almost a third of corporate legal professionals say they are reconsidering outside firms that cannot show AI-enabled value within a year.2
So the question for a firm’s leadership has changed. It used to be “are we using AI?” It is now “is our firm built to turn AI into faster, safer answers?”
The return has moved downstream
Firms with their own development and innovation teams run into what the broader software industry is already learning. Google’s DORA research describes AI as an amplifier of the system it runs inside, and finds that AI adoption has a negligible effect on organisational performance where platform quality is low.3 When AI makes writing code faster, the work does not disappear. It queues up wherever the system is slowest. One vendor’s telemetry across more than ten thousand developers found high-AI teams merging nearly twice as many changes, while the time those changes spent waiting for review rose 91%.4
In a law firm, four places are worth looking at first.
The front door. Requests arrive as a conversation, an email or a chat message. The developer builds to an interpretation, is corrected, and rebuilds.
Somewhere to build. Development environments can be copies of production, carrying production’s controls. Ask whether the vendor systems your software depends on have test instances at all. Trying an idea safely can take weeks.
Two queues. Operational work and innovation work reach the same team by different routes, and nothing ranks them against each other.
Review. AI writes the code in hours. A person still has to read it.
A useful test: map where AI sits in each of your delivery stages, from intake to release. If it sits only in coding — which was already the fastest stage — the other stages are where the return is waiting.
Five signs your firm has the pattern
- “Use AI” is the strategy. There is a mandate, but no written view of where AI applies, where it does not, and in what order.
- Partners’ prototypes set the clock. A weekend demo becomes the expectation for a production system that has to meet client audit standards.
- “Security won’t allow it” is said, but never tested. Ask your security team directly. They may have been asking for the same fix for longer than you have.
- Nobody can say how long a change takes to reach production, because nobody measures it.
- Your best AI practitioners’ workflows live on their own laptops, with no shared standard, so quality depends on who picks up the work.
What to do first
Start with the moves that have no dependencies and need no platform decision.
- Write down what “ready” and “done” mean, so requests arrive with enough to build from.
- Put a shared AI instruction and context layer into your repositories, so the best practitioner’s workflow becomes the team’s default.
- Use AI as a first-pass reviewer ahead of the person, aimed straight at the new bottleneck.
- Take a baseline of delivery now, so improvement can be shown rather than claimed.
Then make the decisions that unlock the rest: a two-tier development platform, where push-button playgrounds hold nothing sensitive and so need no security review to enter, and a secured tier takes the work that graduates; AI at the front door, turning a request described however the requester naturally would into a specification it then interrogates for gaps; and a sanctioned path for developers to try new AI tools.
The culture is not the obstacle
These moves can be designed so that nobody has to say no to a partner. That matters, because a recommendation that depends on partners changing how they ask for work is unlikely to survive its first busy week. The aim is narrow: the only change a requester notices is a faster answer. Client-data obligations stay a fixed floor — the work is to automate and evidence the controls, not to loosen them. None of this is about replacing the people doing the work; it is about removing what makes them wait.
Regulators point the same way. ABA Formal Opinion 512 maps generative AI onto lawyers’ existing duties of competence, confidentiality, supervision and fees.5 In the UK, the SRA’s 2026 warning notice expects firms to show checks, training and audit trails.6 A firm that can show how AI-assisted work is built, reviewed and released is in a better position with its clients, its regulator and its own partners.
Where to begin
Ask one question of your team this week: if a partner brings us an idea on Monday, where does it wait? The answer will show you where AI’s return is being lost.
Sources. Several of these are published by vendors active in this market, and figures are attributed accordingly.
- ILTA 2026 Technology Survey, 500+ member firms, as reported by LawSites, 14 September 2026.
- Thomson Reuters Institute, Future of Professionals 2026, law-firm paper, 21 July 2026 (the research arm of a vendor).
- DORA, State of AI-assisted Software Development 2025, approximately 5,000 respondents.
- Faros AI, telemetry across 10,000+ developers and 1,255 teams, 23 July 2025 (vendor).
- ABA Formal Opinion 512, 29 July 2024.
- Solicitors Regulation Authority, AI warning notice, 17 August 2026.
- ILTA 2025 Technology Survey, as reported by Law360, September 2025.
