Written by: Wiktor Stefański, Head of People & Operations, Digital Colliers
The SRA published its AI guidance in November 2023. Two years on, walk into most UK firms and ask which matters had AI involvement in the drafting, and you'll get a shrug or a promise to check with IT. That's the gap. The regulator asked a reasonable question, and the operational plumbing to answer it doesn't exist in most practices.
This isn't a compliance panic piece. It's about the boring middle layer that separates firms who can answer the question from firms who can't.
What the SRA actually asked for
The November 2023 guidance didn't ban anything. It laid out expectations around competence, confidentiality, accountability, and client communication. The subtext, though, is that if a regulator or client asks whether AI touched a specific piece of work, you should be able to answer.
Most firms treat AI usage as a tooling question. It's a records question. If you can't reconstruct what happened on a matter, you can't supervise it, can't bill it honestly, and can't defend it later. The ABA's Formal Opinion 512 in 2024 made this explicit on billing: you cannot bill hours the AI actually saved. That only works if you know where the AI was used.
And the reputational tail is real. The Stanford tracker of AI-fabricated citations in court filings went from 87 cases to over 1,300 in eleven months during 2024. That's the visible failure mode. The invisible one is a partner who cannot tell a client what tools touched their confidential documents.
What a matter-level AI usage log actually contains
When people hear "AI log" they picture a spreadsheet of ChatGPT prompts. That's not it. A useful log ties to the matter number and captures enough to reconstruct what happened without recording every keystroke.
A practical shape looks like this:
- Matter ID and the fee earner responsible
- Tool used, with version if the vendor exposes it
- Task category: research, first draft, summarisation, review, translation, redlining
- Inputs classification: was client-confidential material sent to the tool, and under which vendor agreement
- Output disposition: kept, discarded, materially edited, used verbatim
- Human reviewer and timestamp of sign-off
- Time saved estimate, for the billing conversation
You do not need every prompt. You need enough to answer three questions later: what did the AI do on this matter, who checked it, and did any confidential data leave the perimeter. If your log answers those, you're ahead of most of the market.
Where the integration actually breaks
The reason firms don't have this today is not laziness. It's that the AI tools lawyers actually use, Copilot in Word, Harvey, Legora, ChatGPT on personal accounts, don't push structured events into the matter management system. So the log has to be manually maintained, which means it isn't.
The pattern I keep seeing in firms that do get this working:
- They pick a small set of sanctioned tools and block the rest at the network layer.
- Each sanctioned tool either has an API or a SSO-based audit export.
- A thin service reads those exports on a schedule and posts events into iManage, NetDocuments, or the practice management system, tagged by matter.
- Fee earners confirm or edit the auto-generated log entry when they close their time. That's the only manual step.
That last step matters. If the log requires a separate workflow, it dies. If it's a two-click confirmation inside the time entry the lawyer is already doing, it survives.
Why this gets harder, not easier, from here
The EU AI Act's Article 50 transparency obligations apply from 2 August 2026, and the high-risk provisions from 2 December 2027. Legal services touch several categories where the high-risk framing is being actively debated. Whatever the final scope, the direction is more disclosure, not less. GDPR fines already reach up to €20M or 4% of global turnover, and the AI Act adds its own regime on top.
Firms that already have matter-level AI logs will treat the 2026 and 2027 deadlines as configuration changes. Firms that don't will treat them as projects, and projects at law firms tend to run late.
The SRA gave everyone a two-year head start in November 2023. The firms that used it are quietly building the plumbing now. The ones that didn't will be answering the question under time pressure, which is the worst way to answer any regulatory question.