Imagine a manager arriving at an executive review with an uncomfortable update. The AI pilot has produced the promised analysis, the presentation is polished, the deadline has been met…but the team isn’t convinced that the underlying work is sound. What happens next depends on the executive. One might ask what remains unverified and what it will take to resolve it; another might ask why the team is still finding problems when the rollout has already been announced.
That hypothetical exchange captures a leadership choice worth examining. An organization can ask employees to exercise judgment while making it difficult for them to deliver the judgments leadership does not want to hear. My argument is that leaders need to make an evidence-backed “not ready” a legitimate professional conclusion. That requires more than telling people to speak up; it requires executives to respond constructively when the evidence challenges a deadline, a success story, or their own decision.
What the new research actually shows
Two preprints posted in mid-October 2026 offer a useful reason to examine this issue (although neither studies executive behavior).
In Into the Unknown, submitted October 13, researchers evaluated nine frontier AI agents on 15 unpublished scientific research directions developed with scientists at U.S. Department of Energy national laboratories. None of the 135 resulting submissions met every condition. The researchers reported that 114 failed their integrity check, most often because a paper claimed work that had not been performed or reported numbers inconsistent with the recorded evidence. These findings come from a small, demanding research benchmark, so they don't establish a general failure rate for AI-assisted work (arXiv: Into the Unknown).
BizBench, submitted October 14, examines business execution using 50 tasks grounded in procedures from a payments company. Its checks distinguish reaching a goal from respecting constraints along the way. For one tested model, goal attainment was 42.0%, while success under the stricter criteria was 32.7%. This is experimental evidence from a particular environment; it doesn't measure ordinary enterprise deployments (arXiv: BizBench).
The studies examine different kinds of work. Together, they give us a reason to question what an apparently successful deliverable establishes. They don’t show that AI is useless, that every output needs exhaustive manual inspection, or that leadership culture caused the observed failures. The leadership implication is my interpretation: if polished work and proven work can diverge, executives should examine whether their own management practices help people expose the difference.
Your response becomes part of the standard
Consider a hypothetical service operation testing an agent that resolves customer disputes. Its dashboard shows faster closure. A team member reviewing a sample discovers that some resolutions bypassed an approval requirement. The immediate technical question is how to prevent that behavior. The leadership question is what happens to the person who raises it. Does the executive treat the finding as useful work? Is the rollout decision reopened? Does the team have time to investigate? Or does the discussion turn immediately to protecting the promised launch date?
I would want leaders to examine those questions before issuing another instruction to “use good judgment.” Employees need to know what exercising that judgment permits them to do when it creates an inconvenience.
This doesn't mean we have to celebrate delays, rather, it means distinguishing a supported concern from an unexplained objection and dealing seriously with the former. Someone who identifies a significant gap should be expected to describe it clearly and leadership should be expected to make a decision about it.
Separate progress from permission to proceed
One practical improvement is to stop treating a single status label as a complete account of the work. An analysis can be drafted while its assumptions remain untested, a workflow can function while an important exception remains unresolved, and a pilot can demonstrate potential without being ready for broader use. Ask the team to state what’s been achieved, what the evidence supports, and what decision is justified. These descriptions don’t have to become a new reporting system; a few explicit sentences may be enough. For example: “The agent resolved the test cases faster. We’ve checked the approval history for this sample and two exceptions remain unexplained. We recommend continuing the bounded test while holding the wider rollout.”
That’s a more useful basis for leadership than a green indicator whose meaning nobody has agreed on. It also preserves genuine progress, because finding a limitation does not erase everything the team has learned. At Whole Enterprise, this is viewed as part of Intelligence Stewardship™: taking responsibility for how intelligence informs high-stakes choices. The executive’s contribution includes setting a standard the team can apply honestly, even when the answer changes the executive’s plans.
Make room for challenge without creating paralysis
A fair objection: If leaders encourage people to withhold approval whenever there’s any level of uncertainty, work could stall indefinitely. Absolute certainty is unrealistic and ongoing review consumes time that could otherwise create value. The answer is to define sufficient evidence for the decision at hand. A reversible internal experiment should not carry the same burden as an action that changes a customer’s finances or exposes confidential information.
Before work starts, agree on the important constraints, the checks required, and the person authorized to accept the remaining uncertainty. Where possible, make the next step smaller and reversible. This gives a concern somewhere productive to go besides a permanent stop.
AI can help perform checks, compare records, and identify inconsistencies. The question is whether those checks provide meaningful evidence. Asking the same system to confidently restate its conclusion is not, by itself, independent verification. For work with real consequences, use records or tests that can establish what occurred, with human expertise where the decision warrants it.
Nor should “not ready” become a shield against accountability. Require a specific gap, its consequence, the next useful test, and a date for reconsideration. If a rule is outdated, assign someone the authority to change it instead of quietly allowing exceptions to accumulate.
Start with your next executive review
Choose one AI-enabled initiative with a visible sponsor and a meaningful commitment. Before its next review, ask the sponsor and delivery lead to agree on three things: 1) what would justify proceeding, 2) what finding would require a pause, and 3) who can make that decision. Then ask the team to bring one unresolved issue alongside its achievements. The purpose is to understand the limits of the evidence; nobody should feel they have to manufacture doubt or perform skepticism.
In the meeting, examine your own response. If the finding threatens a commitment you made, acknowledge that conflict directly. Explain which tradeoff you are choosing and why; don’t leave the delivery team to reconcile a public promise with a private instruction to be careful. Make the resulting decision visible. If you narrow the rollout, name the condition for expanding it. If you proceed, record the uncertainty you are accepting and who will monitor it. If more work is necessary, identify the time and support available to do it. Finally, recognize the contribution accurately. A team that discovers an important limitation before expansion has supplied information leadership needed. It may not have proven the pilot successful, and it can still count as valuable progress.
The leadership test arrives when the work looks finished and someone has the evidence, and the courage, to say otherwise. What you do in that moment tells people whether your standards can survive contact with your ambitions.
