Before approving an AI pilot, ask its sponsor to name the work that should disappear and the person who will still own the result. “Faster drafting” leaves the operating case unfinished. What happens to checking, handoffs, exceptions and approval?
For AI workflow automation in professional services, make a bounded workflow the unit of investment. Start with repetitive preparation around a professional decision, then test whether reducing that effort improves delivery after review and rework are counted. Use the framework below to select a candidate, not to infer a universal ranking of jobs or a promised return.
Start where work accumulates
Choose a delivery workflow and trace it from intake to approval. Separate active effort from waiting, and first-pass work from corrections. Locate the skilled time spent finding, reorganising or checking information before anyone applies professional judgment.
Identify required inputs, permissions and missing information.
Locate source material and structure it for review.
Record corrections, queues and acceptance requirements.
Name the decision owner and route unresolved exceptions.
A proposed mapping sequence, not a ranking of automation opportunities.
Trace the delivery sequence, recording effort, waiting and responsibility at each stage.Define the candidate by its start and end points. A broad research remit needs permitted inputs, a specific preparation task and a condition for reviewer acceptance before it is ready to test. If approval waiting time is the suspected bottleneck, require the proposal to explain how faster preparation would affect that queue.
Separate preparation from responsibility
Treat retrieval, classification, structuring, drafting and first-pass flagging as functions to assess. Keep professional judgment, client accountability, exception resolution and approval with named people. This is a recommended operating boundary, not evidence that every preparation task is suitable for AI.
Make human review specific: what must the reviewer verify, which supporting material will they need, and when must they reject or escalate an output? “Human in the loop” is not a sufficient definition of decision rights. Where an output cannot be checked reliably, narrow the task or defer it.
Set confidentiality and access boundaries before testing client material. Specify what may enter the tool, who may retrieve it and where outputs may go. Route missing, conflicting or unauthorised inputs to an explicit exception path rather than treating preparation as complete.
Prioritise bottleneck relief and controllability
Compare team proposals against the same criteria to decide which are ready to test, which need repair and which should wait. Treat confidentiality and accountability as conditions of entry. Do not let a high-volume task compensate for unresolved controls in a weighted score.
Otherwise, establish the workload and task boundary first.
Confirm availability, usable structure and permitted access.
Assess exception frequency, error consequences and escalation ownership.
Include correction effort and reviewer time in the assessment.
Specify retained manual steps and restrictions on client data.
Choose an observable cycle-time, throughput or effort measure.
Editorial selection criteria, not validated scores or an evidence-backed priority ranking.
Use these criteria to choose between a bounded test, further workflow design and deferral.Prefer a candidate with a visible source of friction, clear inputs, an accountable reviewer and a workable fallback. Include integration effort in the investment decision, including manual copying, access checks and transfers that would remain. If the team cannot describe how the output enters the next step, keep that workflow design inside the pilot scope.
Assess internal-document work on task and tool fit
BCG’s 2025 analysis provides a bounded example: 80% of respondents said they used specialised GenAI often or very often for consulting internal documents, compared with 35% using general GenAI for that purpose. Here, “consulting” describes the activity of consulting documents, not an occupational group. The finding measures reported use, not time saved, quality improvement or end-to-end automation.
BCG also reports that 49% of general-tool users described significant or extensive rework, versus 29% of specialised-tool users. The proportions reporting no rework were 14% and 38%, respectively. The supplied passage does not identify the tasks behind these responses. Do not attribute those rework figures specifically to internal-document work.
Neither comparison establishes causality or independently audited quality. The available excerpts also do not establish sample size, geography, occupational mix or detailed survey methodology. Together, the findings justify investigating task and tool fit, but do not establish which workflow to automate first.
For an internal-document candidate, compare specialised and general-purpose tools using the same permitted material and acceptance requirements. Assess retrieval, source traceability, required corrections and reviewer effort. Make generation speed one measurement within that test, rather than the purchasing criterion.
Connect task speed to an operating change
The UK Government’s Professional and Business Services AI Adoption Plan, published on 8 June 2026, describes AI helping individuals complete the same work faster while gains remain localised without corresponding changes to workflows, decision-making and organisational design. This is the plan’s characterisation of the UK sector, not a controlled estimate of firm-level productivity.
Require the investment case to specify the surrounding operating change. Identify which handoff would change, which duplicate check could be retired after validation and who can authorise the revised process. Decide how any released time would be used: additional delivery, shorter queues or other work.
Measure through to approval. If preparation becomes faster but reviewer effort rises, count both. If approval remains the constraint, do not declare a throughput gain from preparation speed alone. Reserve claims of reduced cost or increased capacity until the firm has measured those outcomes.
Make the first automation measurable and reversible
In Thomson Reuters’ 2026 AI in Professional Services Report, 18% of professionals said their organisations tracked AI ROI; another 40% did not know whether it was measured. These responses concern self-reported knowledge of organisational measurement, not actual returns or successful workflows. The available excerpts do not provide sample size, geography or survey methodology.
For your pilot, agree the baseline and scale criteria before testing. Record the intended improvement and the quality or control conditions that must not deteriorate. Compare sufficiently similar work, document differences in complexity and count setup, review, correction and ongoing operating effort.
- — Ownership: name the workflow owner, approval authority and person authorised to stop the automation.
- — Measurement: define cycle time, throughput, rework or review effort consistently before and during the test.
- — Auditability: retain permitted source references, output versions, review decisions and exceptions needed to reconstruct the work, within applicable retention rules.
- — Fallback: specify how work returns to the existing process when inputs, outputs or the tool fail acceptance checks.
- — Scale criteria: require demonstrated improvement within agreed quality and control limits. Otherwise revise, narrow or stop the pilot.
At the scale decision, require a plain answer: what work was removed, who still owns the result, and what improved after the full review burden was counted? If the evidence shows only faster preparation, keep the claim and the investment case at that level.
Written by the studio, from work in delivery
