Where AI sits in this work

Organising client material and interview records; building lists of facts, assumptions, gaps and contradictions; assisting retrieval and comparison, and producing research drafts.

Filling material into the dimensions, routes and conditions the advisor has set; drafting reports, tables, checklists and knowledge assets; writing scripts and light tools to the advisor's design.

Retaining inputs, rules, outputs and revision history so a judgement can be replayed. That last one is the precondition for everything above it: output that cannot be replayed cannot serve as evidence in advisory work.

Setting method and structure: which dimensions the problem splits into, which route it follows, which data is used, how expert labelling is defined, and how output is debugged into something usable. These are advisor judgements, not model output.

Confirming the problem has been understood correctly. The problem as stated and the problem as occurring are frequently not the same one.

Judging whether a source holds and whether it is current; separating what is fact from what is inference from what remains unproven.

Handling value conflicts, internal organisational reality and client context; deciding whether a recommendation fits this particular company; explaining the limits externally and carrying the advisory responsibility.

Data. Client confidential material, trade secrets and sensitive documents have a stated scope of use, and the convenience of a tool does not extend it.

Evidence. Model output is a candidate signal. Anything without a source, or that cannot be cross-checked, does not become a stated fact.

Decision. AI does not make the significant operating, legal, investment, personnel or market-entry decision for a client.

Capability. A prototype or an internal tool is not a stable product. Only capability that has passed real acceptance becomes an external commitment.

That AI produces the correct strategy on its own. That a report generated quickly is advisory quality. That a technical demonstration substitutes for client value and real cases.

These read as self-imposed limits. They are closer to protection: let any one of them slip, and a client loses the ability to tell which part of a conclusion was judged and which part was merely generated.