Succession and AI Governance Share the Same Test

In one room, a board is deciding how much authority a successor can carry before the CEO leaves. In another, an operating team is deciding whether an AI agent may act without approval. The participants, technology, law, and human consequences differ. Yet both rooms eventually reach the same practical question: when are we willing to let this new decision-maker act without the incumbent?

A résumé does not answer it. Neither does a benchmark score. Historical agreement can be reassuring while hiding dependence on familiar cases. What the organization needs is evidence that the new decision-maker can recognize the relevant situation, act within a boundary, and seek help when the case crosses it.

That structural resemblance is useful only if it remains bounded. Succession is not an AI-alignment problem, and an AI deployment is not leadership development. The comparison earns its value by improving the questions each side asks, not by collapsing them into one science.

A decision benchmark is a body of representative cases, variation, evidence standards, failure criteria, and escalation expectations used to judge whether a decision-maker is ready for defined authority. It is more demanding than a knowledge test because consequential work rarely repeats in exactly the same form.

Critical Decision Method research offers one lineage for building the cases. By reconstructing real nonroutine incidents and probing cues, discriminations, expectations, and alternatives, it makes parts of expert work inspectable. Applied Cognitive Task Analysis offers related ways to identify cognitive demands and translate them into training or design artifacts.

These methods do not recover a complete internal algorithm, and neither was validated as a universal readiness system for CEOs or AI agents. Their contribution is a disciplined starting point: study the conditions that changed the call, then create variation around those conditions.

A successor can memorize the incumbent’s preferred response. An AI system can reproduce patterns from examples. Both may look strong when the evaluation resembles the material they have already seen. The harder question arrives when one fact changes, two familiar conditions conflict, or the old answer no longer fits the organization’s strategy.

Representative testing therefore needs altered cases and holdouts. Change the stakeholder, available time, authority, reversibility, or evidence quality. Include cases where escalation is the correct response and others where unnecessary escalation would be costly. Examine the reasoning a successor can articulate and the observable behavior of the full AI system.

Even then, the evidence remains bounded. Scenario performance does not predict every future act by a leader, and an evaluation set cannot prove general AI reliability. It tells the organization more than historical familiarity alone—and shows where uncertainty should limit authority.

A successor develops socially. They absorb culture, build relationships, revise their understanding through experience, and can explain, contest, or take responsibility for a decision in ways shaped by human institutions. Their judgment may appropriately differ from the predecessor’s as strategy and context change.

An AI system has different transparency limits and failure modes. Its behavior depends on models, prompts, tools, data, permissions, interfaces, and the people around it. Responsibility remains distributed among the organizations and people who design, deploy, authorize, and oversee the system; the model does not become a fiduciary or accountable executive.

The shared test therefore branches. A successor may need evidence of stakeholder judgment, learning, and progressively independent authority. An AI-enabled workflow needs technical evaluation, access constraints, monitoring, recourse, security, and explicit human accountability. Agreement on a set of cases is only one piece on either side.

The strongest decision-maker is not the one who answers every question. It is the one—or the system—that recognizes when available evidence, authority, or consequence makes independent action inappropriate. Escalation is part of competent performance, not an exception to it.

For a successor, the board can observe whether uncertainty is surfaced early, the right stakeholder is engaged, and advice is sought without surrendering legitimate authority. For an AI agent, the design must detect defined conditions, pass sufficient context to a qualified person, and permit that person to disregard, override, or stop the system.

EU AI Act Article 14 makes those human-oversight capabilities concrete for high-risk systems within its scope. The International AI Safety Report 2026 adds the caution that human oversight can help while automation bias and real-world evaluation gaps remain. On the succession side, widespread board concern about ready internal candidates shows the practical pressure, though survey concern does not validate any one readiness method.

A transition is not complete because the successor or system performed well. The organization should state what still depends on the incumbent, expert, model provider, reviewer, data pipeline, or surrounding team. Dependency may be reasonable; invisible dependency is the problem.

For a successor, residual dependency can include relationship access, rare-case exposure, or authority still held by the outgoing leader. For an AI workflow, it can include one expert who writes all evaluation cases, a vendor interface that conceals uncertainty, or a manual recovery path no one has practiced.

Naming those dependencies changes governance. It tells the board what evidence is missing, what redundancy needs to be built, and which conditions should prevent further transfer of authority. It also prevents an assisted success from being mistaken for independent capability.

Skagway’s current work sits on the human side of this bridge. The Passage uses captured decision cases, supervised practice, progressively independent authority, and readiness evidence to support a named successor. It does not certify a leader, predict all future performance, or replace the board’s fiduciary judgment.

The AI side remains research and practitioner point of view. Skagway does not currently validate AI agents or provide broad AI-governance consulting. Technical assurance, cybersecurity, regulatory compliance, and model evaluation require specialists and evidence beyond a succession methodology.

The longer-term question is worth studying precisely because neither domain has solved it generically: what evidence should justify trusting a different decision-maker with a consequential call under variation? A careful body of decision benchmarks could improve both conversations while preserving the differences that matter.