The situation
Professional-services firms adopt AI under conditions most businesses do not face. The product is expert judgement, signed off by a partner. The most valuable raw material is client information, held under confidentiality obligations, information barriers and increasingly explicit client terms about AI. And the commercial model often rewards time spent, which AI is designed to reduce.
Inside most firms the result is uneven. One practice group pilots a drafting tool; another bans everything. Business services quietly use public chat tools. A major client sends a questionnaire asking how AI is used on its matters, and nobody can answer for the whole firm. The managing partner is asked to approve spending without a view of where the value or the exposure lies.
The task is sequencing: deciding what goes first, under which rules, sponsored by whom, so the firm learns safely and partners can see evidence before client work is affected.
What the work involves
The use-case portfolio. I interview practice and business-services leaders and collect candidate uses: research and first drafts, document review preparation, proposal and pitch assembly, knowledge retrieval over precedents, time narrative drafting, meeting notes, internal reporting. Each is scored on four things that matter in a firm: value to the practice, client-data exposure, review burden on partners and senior staff, and effect on recorded time.
The client boundary model. With your risk partner, I classify client information (public, firm-generated, client-confidential, specially restricted by contract or regulation) and map each class to the tools and environments it may enter, the permissions those tools must respect, and the client consent or contract term each use depends on. Questions the model cannot answer, such as whether a particular client’s terms permit any AI processing, go to the risk partner as named decisions.
The governed pilot sequence. Stages run from internal, non-client work, through client work in approved environments with existing matter permissions, to client-facing outputs where the client has agreed. Every stage has a partner sponsor, a review rule (who checks what before it leaves the firm) and an exit criterion.
My approach to review rules follows published work on approval boundaries: the system should make the required review unavoidable, not rely on people remembering it. My book on everyday AI automation covers the research, drafting and reporting workflows that usually make up the first stage.
The signature deliverable, illustrated
Illustrative extract from a pilot sequence, not taken from a client:
| Stage | Scope | Client data allowed | Review rule | Partner sponsor | Exit criterion |
|---|---|---|---|---|---|
| 1 | Proposals and knowledge retrieval from firm material | None | Associate checks citations; partner approves proposal | Head of business development | Reviewers report fewer corrections over the pilot |
| 2 | First-draft research memos on live matters | Client-confidential, in approved environment only | Supervising partner reviews every memo | Practice group head | No boundary breaches; review time recorded |
| 3 | Client-facing summaries | Only where client terms permit | Partner sign-off; client informed | Managing partner | Client feedback and risk partner approval |
How acceptance is judged
The management committee accepts the portfolio, boundary model and sequence as a decision record, and each pilot is accepted against its own exit criterion. The risk partner signs off the boundary model before any stage involving client information starts. Pilot evidence covers time, reviewer corrections and any boundary incidents.
Ownership and handover
The COO owns the sequence; each stage belongs to its partner sponsor; the risk partner owns the boundary model and keeps it current as client terms change. I hand over the scoring method so the firm can add new use cases without restarting.
When to choose something else
To enable one client-delivery team inside an approved stage, see AI enablement for client-service teams. For the firm’s own finance function, see AI enablement for finance teams. If tools are bought but unused, start with stalled rollout rescue. Other sectors are on the industries overview.