Evaluation of internal AI usage across a professional practice

Context

A practice working on sensitive matter types had AI use growing informally across teams. Partners suspected real productivity lifts; they also suspected quiet quality erosion. Nobody had a shared view of what was in use, where value was real, and where the risk surface was expanding.

Policy had been drafted twice and shelved twice. Practice had moved faster than policy; the gap between them was widening.

Problem

Informal AI adoption is a slow-burning governance gap. Policy without practice is unacceptable; policy written to match current practice is already too late. The team needed a current-state picture they could act on without slowing delivery.

The risk was not a single incident. The risk was a compounding quality drift nobody could evidence at the moment it started to matter.

Constraint

Nothing could be seen to slow the team down. No compliance-theatre. The review had to produce a guide the team would actually read and use.

The written output had to survive a partner meeting on a Monday morning without a defender in the room.

Intervention

Structured interviews with a cross-section of operators. A usage map categorised by matter type and by risk class. An evaluation rubric per live use case, focused on failure modes rather than capability.

The deliverable was a minimum-viable internal guide: a short list of patterns explicitly approved, a tiered set of guardrails for each, and an off-limits list with reasoning. Language was the team’s own, not a borrowed compliance template.

Outcome

A short number of high-leverage uses formally adopted and documented. An explicit off-limits list — with reasoning — written by partners and signed off by operations. A working review cadence owned inside the team.

The engagement continues on a quarterly refresh rhythm. The guide now has a named owner inside the practice; R-Labs is no longer in the loop for day-to-day decisions.