The manager’s AI advantage is knowing when not to delegate

A manager’s advantage with AI comes from knowing which parts of a decision can be delegated and which require an independent test, rather than expecting the same assistant to judge every answer it produces.

Why does a capable assistant still need a task boundary?

The HBS research summary of an experiment with 758 BCG consultants reports gains on tasks within the model’s capability frontier, alongside lower correctness on a task outside it. The result came from a 2023 GPT-4 setting, not today’s agents. The enduring lesson is the uneven boundary between tasks, not a permanent ranking of models or a universal accuracy figure. HBS23

For a strategy manager, this means that brainstorming a plausible market entry, reconciling a disputed demand figure and authorizing the investment are different jobs. The quality of the first output does not establish competence at the second. The third remains a commitment owned by the manager and the organization.

Ask what must be checked independently

Start with the failure that would change the decision. If the recommendation depends on a customer segment’s willingness to pay, fluent synthesis of interviews cannot replace a comparable demand test. If it depends on a margin estimate, the finance owner must inspect the definitions and arithmetic. Choose the independent check before reading the polished recommendation.

A second model can help find omissions, but it does not automatically provide independence. Two systems can repeat the same source error or persuasive framing. Independent evidence means a different route to the relevant fact: a source record, a reproducible calculation, a contrary customer observation or an appropriately qualified reviewer.

Use disagreement to improve the question

AI is useful as a structured challenger when its role is to surface alternatives, reveal an assumption or explain what would falsify a position. Ask it to identify the strongest competing explanation and the evidence that distinguishes it. Then assign the actual verification to someone who can access and evaluate that evidence.

In a revenue review, the model can propose conversion, customer mix and revenue timing as rival explanations. The sales-operations owner checks whether the periods and cohorts are comparable. The team chooses an investigation or a pilot only after acknowledging what is known. Asking the model to sound more confident would remove the very uncertainty the manager needs to manage.

Do not turn review into a second full production line

Exhaustive review of every low-stakes sentence can consume more effort than drafting the document manually. The alternative is a risk-based evidence standard: check consequential numbers and commitments directly, sample routine transformations, and keep uncertain recommendations separate from authorized action. Set the intensity according to the cost of being wrong and the chance of detecting it before harm.

This is a design choice to test, not a claim that human review always improves accuracy. Reviewers can be rushed, lack source access or defer to polished output. The manager must provide the time, expertise and authority to challenge the work; an approval checkbox is not a substitute.

Build a local delegation map next week

The strategy or operations lead can choose a recurring decision and separate its preparation, evidence checks and commitment. Collect representative tasks plus edge cases, agree the acceptance criteria with the decision owner, and compare the proposed AI-assisted route with the current one. Record errors that would have survived ordinary review as well as the time required to catch them.

Separate the evidence check from the authority to accept the result. The person who prepared the recommendation should not be its only consequential reviewer. Name a decision owner who can resolve disagreements between the author and reviewer, and keep unresolved differences visible rather than treating model agreement or specialist silence as acceptance.

One review cycle is one complete instance of preparing the selected brief, independently checking it and recording the owner’s acceptance decision. The decision owner schedules a review after two such instances, counting rejected and unfinished briefs too. That checkpoint never delays the immediate suspension rule.

Suspend the affected delegation as soon as a consequential error survives review or a reliable check is missing. Use the two-cycle review to assess the remaining route, not as permission to keep using a failed boundary. Extend the boundary only when the owner can show repeatable quality on the relevant task set, including exceptions. A model update, a new data source or a new use case starts a new evaluation question; last month’s success is not blanket permission.

Questions managers ask

When should a manager override an AI recommendation?

When a load-bearing claim fails its source or calculation check, the recommendation exceeds its authorization, or new evidence contradicts its assumptions. Record the reason so the next review can learn from it.

Can another AI verify the first model?

It can challenge the draft, but agreement does not establish independent verification. Check the relevant fact through a source, calculation or competent reviewer with a different evidence path.

Does an old AI experiment predict current model performance?

No. Historical experiments show how outcomes can differ across tasks. Use them to design a local evaluation, then test the current model, data and workflow.

Sources and scope

HBS23: Humans vs. Machines: Untangling the Tasks AI Can (and Cannot) Handle — Harvard Business School Working Knowledge; 2023-11-22. Institutional primary research summary, read in full public HTML.

MS-EVAL: Morgan Stanley uses AI evals to shape the future of financial services — OpenAI; Undated; accessed 2026-10-27. Primary vendor/customer account; commercial interest disclosed.

PG26: The Cybernetic Teammate: A Field Experiment on Generative AI and Teamwork — Organization Science; 2026-06-12. Primary peer-reviewed field experiment; public full text.

Research checked 2026-10-27. Next review 2026-11-27. Suggested pilots are editorial proposals, not observed Upskill Pro customer results.