Give a board recommendation on an enterprise AI investment.

Question

Fictional board case. Exhibit A: A six-week internal policy-search pilot involved 24 volunteers and 240 sampled tasks; reviewers accepted 204 answers, but 12 answers cited superseded policies. The exhibits do not say whether those 12 overlap with the 204 accepted answers. Exhibit B: Median drafting time fell from ten to six minutes; review time was not measured, and the tasks were not matched to a control group. Exhibit C: Access tests passed for the pilot group, but permissions for two additional departments have not been tested. Exhibit D: The board can approve a maximum of $60,000 for an eight-week controlled expansion; a company-wide commitment would cost $400,000. A policy owner and service owner can be assigned, and support staff can cover at most 50 trial users. Exhibit E: The board requires appropriate access, supported answers, and a credible account of total task effort before wider release. No numeric quality threshold has been agreed. Recommend approve, defer, narrow, or stop; reasonable alternatives need an explicit evidence-based rationale.

Accepted Answer

Example Answer

I recommend authorizing an eight-week controlled evaluation, capped at $60,000 and 50 users, with user expansion conditional on corrected sources, agreed release criteria and permission tests for the affected departments. I am not seeking the $400,000 company-wide commitment.

The current evidence is promising but insufficient for wider release: reported acceptance is 85%, yet we do not know how the 12 obsolete-policy citations overlap that count. Drafting became faster, but total effort and a causal improvement are unproved. Access results apply only to the tested pilot group.

The memo records what we would learn, who is accountable and when we stop. At eight weeks I would return with a recommendation against criteria approved before new results arrive. If source ownership or access prerequisites cannot be met, I would defer expansion and return uncommitted funding rather than convert this authorization into an automatic rollout.

Board decision memo

Decision requested: Authorize up to $60,000 over eight weeks for a controlled evaluation serving no more than 50 users. Approve no company-wide spending. The ceiling is authority to spend on the agreed plan, not a requirement to use the full amount.

Evidence and limitations: Of 240 sampled tasks, 204 were accepted: 85%. Twelve answers cited superseded policies, but their overlap with the accepted set is unknown; reconcile labels and re-review affected answers before drawing a quality conclusion. Median drafting time changed from ten to six minutes. Missing review time and unmatched tasks prevent a total-effort or causal productivity claim. Passed access tests cover only the pilot group. Volunteers do not establish performance for ordinary employees.

Required action	Accountable role	Decision consequence
Correct superseded sources and reconcile acceptance labels	Assigned policy owner	No affected policy content enters expanded use without authoritative review and ongoing source maintenance.
Define supported tasks, supported-answer criteria and release thresholds	Service owner with policy, security and board delegates	Approve measurable criteria before collecting the next results; do not choose a pass mark after seeing performance.
Test additional-department permissions, including denied access	Security owner	Exclude either department until its required tests pass; existing pilot results cannot authorize it.
Measure representative end-to-end task effort and quality	Service owner and evaluation lead	Include review, corrections, exceptions and support; use comparable task assignments and report uncertainty and sample limits.
Operate within approved users, cost and incident limits	Service owner	Retain the non-AI fallback, review consequential answers, monitor spend and pause affected use on a serious access or unsupported-answer incident.
Present decision at eight weeks	Executive sponsor	Return evidence against agreed criteria, costs and incidents; recommend a separately costed expansion, explicit deferral or closure.

Release decision: Wider release requires supported answers, appropriate access and credible evidence of total task effort, plus a viable operating plan. Passing this evaluation does not itself authorize $400,000. Inconclusive results require an explicit decision on whether additional evidence is worth buying; they do not renew funding automatically.

Defensible alternative: Defer the expansion budget now, resolve source ownership and permission-test planning first, and resubmit a scoped request. That approach is reasonable if the board cannot authorize conditional spending or the prerequisites cannot credibly fit within eight weeks.

What the interviewer is assessing

Essential: Makes a specific funding and scope recommendation within $60,000 and 50 users, reports 204/240 as 85% without treating it as an approved quality standard, and keeps $400,000 as a separate decision.

Stronger: Detects unknown overlap in the 12 obsolete citations, missing review time and unmatched tasks, and the limited scope of access tests. Deferring new authorization until these prerequisites are resolved is also defensible if the candidate identifies the evidence and owner needed to return to the board.

Red flags: Claims 40% total productivity improvement, subtracts all 12 obsolete citations from 204 without knowing overlap, treats 85% as sufficient because it looks high, or commits company-wide spending from volunteer evidence.

Follow-up

What would make you recommend stopping instead of another evaluation cycle?

Example response: “I would recommend stopping if corrected authoritative sources still fail the agreed supported-answer standard, if acceptable access cannot be enforced, or if representative end-to-end work shows insufficient value at a credible operating cost. I would distinguish a remediable evidence gap from a failed business case, state the remaining remediation cost and avoid extending simply because money has already been spent.”

Pressure test

Changed constraint: The policy owner cannot participate during the eight-week period.

Adaptive response: “I would defer user expansion unless a qualified replacement with authority over the sources can be assigned. Permitted offline evaluation may continue only if it has useful objectives within the approved plan. If no owner can be secured, I would postpone or close the evaluation and return uncommitted funds; the service owner cannot simply declare obsolete policies correct.”