Research
A consulting firm should draw its own AI task boundary
Separate tasks the tool handles well from adjacent tasks that require stronger human judgment.
Zocri · 2 min read · Sources checked
Similar-looking tasks can behave differently
The consulting experiment described in Navigating the Jagged Technological Frontier finds task-dependent effects across its tested capability boundary. The result concerns particular tasks, consultants and technology. It does not supply a permanent list of everything AI can or cannot do for a small consulting practice.
A proposal outline and a recommendation based on conflicting client evidence may both look like writing tasks. Their failure costs and checking requirements are different. Treating them as one category can lead the firm to expand from useful drafting assistance into analysis it has not actually evaluated.
Break a deliverable into decisions
List the components of a typical engagement: gathering facts, organizing evidence, proposing alternatives, choosing a recommendation and explaining it. Identify which components can be checked directly against client materials and which require judgment about incomplete or contradictory information.
Give each component its own trial. A model that organizes notes accurately has demonstrated that capability, not authority to make the final recommendation. This separation makes the evaluation more useful and lets the firm retain assistance where it works without accepting the entire generated deliverable as one indivisible product.
Test a plausible wrong answer
Include a case where the obvious narrative is misleading. Ask a reviewer to evaluate the output against the original evidence rather than its polish. Record whether the assistant acknowledges uncertainty, invents a missing connection or gives disproportionate weight to an attractive but weak explanation.
Have another consultant challenge the proposed conclusion. The purpose is to identify where verification is expensive enough to erase the drafting advantage. If a generated recommendation takes longer to audit than developing one directly, narrowing the tool's responsibility may be the more economical choice.
Maintain a living task policy
Record which tasks are approved for assistance, what evidence a reviewer needs and when a partner must intervene. Revisit the policy as the firm's work or tools change. Keep it understandable enough that a new consultant can apply it without guessing the intended level of autonomy.
This task boundary is an operating recommendation derived from the experiment's central distinction. It should become more precise through the firm's own results. The useful objective is consistent client work with a visible basis, rather than a broad declaration that a consultancy is either fully automated or unwilling to use AI.