When Should a Decision Be Delegated, Not Just Automated?

AA manager who hands a hiring decision to a trusted deputy and a manager who writes a fixed scoring rule for a hiring algorithm have both stopped making the decision personally. Only one of them has delegated it. Automating a decision means specifying, in advance, exactly what should happen in every case the rule-writer can anticipate. Delegating a decision means handing it to someone — or something — whose judgement you're trusting to handle the cases you didn't anticipate, which is a fundamentally different bet, and conflating the two is how an organisation ends up surprised by an automated system behaving exactly as specified in a situation nobody specified for.

Trust as Willingness to Be Vulnerable FoundationalKnowledge that endures for decades — core principles

RogerRoger Mayer, James Davis, and F. David Schoorman's integrative model of organisational trust gives this bet a precise shape. Trust, on their account, is a willingness to be vulnerable to another party's actions, based on the expectation that the other party will perform a specific action important to the trustor, independent of the trustor's ability to monitor or control that party¹. That last clause is the one doing the work: trust only becomes a meaningful question once monitoring and control stop being complete, which is exactly the condition that separates delegation from automation. Mayer, Davis, and Schoorman break the trustworthiness being bet on into three components — ability (competence in the specific domain), benevolence (whether the trustee is motivated to act in the trustor's interest, beyond what's contractually enforced), and integrity (adherence to a set of principles the trustor finds acceptable). A decision can be automated with only the first component verified. Delegating it responsibly needs some confidence in all three.

Delegation Needs a Model of the Delegate, Not Just a Rule FoundationalKnowledge that endures for decades — core principles

ThisThis cluster's page on shared mental models already describes what makes two experienced teammates need fewer words: an accurate model of how the other person is likely to act. Delegation draws on the same resource for a different purpose — not predicting what a peer will do next, but deciding how much latitude to hand someone whose next move you won't be watching closely. A deputy who has handled difficult hiring calls before, in ways the manager has seen and judged sound, earns wider latitude than one who hasn't, and the earning happens exactly the way a shared mental model accumulates: through repeated, observed instances, not through a single conversation about values. Automation never needs this model at all, because automation isn't betting on anyone's judgement in the first place — only on whether the rule was written correctly for the cases it covers.

What Changes When the Delegate Is an Agentic AI System Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

TheThe ability/benevolence/integrity split survives the move from a human deputy to an AI agent, but each leg changes shape. Ability becomes a capability evaluation — this site's page on designing auditable, robust agentic systems covers what that evaluation actually needs to check before an organisation hands a system real latitude. Integrity becomes alignment with a stated set of constraints the system can be audited against, which is more tractable to verify than a human's sincerity, precisely because the constraints can be written down and tested against directly. Benevolence is the leg that doesn't translate cleanly at all: an AI system doesn't want anything in the sense Mayer, Davis, and Schoorman's model assumes a human trustee does, so "benevolence" for an agentic delegate collapses into whatever values were specified by whoever built or trained it — the system is executing imported goals rather than holding its own, the same point this site's Philosophy of AI series raises about a system with capability but no independent ethical standpoint of its own. Delegating to an agentic system is, in this specific sense, always delegating to whoever set its objectives, one step removed.

Strongest Objection: Real Delegation Is Exactly Where Misuse Is Hardest to Catch FoundationalKnowledge that endures for decades — core principles

TheThe vulnerability that makes delegation meaningful — handing over a decision you can no longer fully monitor — is also the condition under which misuse, drift, or deception is hardest to catch, for a human deputy and for an agentic system alike. This site's page on when agents fail covers the brittleness and misaligned-incentive failure modes this risk takes for AI systems specifically. The honest version of delegation accepts this risk as the actual cost of the latitude being granted, rather than pretending auditability can be preserved in full while still calling the arrangement delegation — a decision an organisation can fully monitor and override at every step was never really delegated in the first place; it was automation with a human, or an agent, standing in as the executor of a rule.

Provisional Conclusion FoundationalKnowledge that endures for decades — core principles

AA decision is delegated, rather than merely automated, exactly when the delegator accepts genuine vulnerability to the delegate's judgement in cases a rule couldn't have anticipated — and that acceptance should track a real history of demonstrated ability, verifiable integrity, and, for a human delegate, benevolence earned over time, not the mere fact that handing the decision off is convenient. An agentic AI system can meet the first two conditions in principle and structurally can't meet the third in the way a human can, which doesn't make delegating to one impossible, but does mean the organisation is always, in that specific sense, trusting a chain of people one step removed from the system actually making the call.

Questions for Further Thought

• Can an organisation meaningfully delegate to an agentic system without delegating, at the same remove, to whoever set that system's objectives — and does naming that remove explicitly change how the delegation should be governed?
• How much demonstrated ability should it take before a decision moves from automation (a rule, checked carefully) to genuine delegation (latitude, granted on trust) — and is that threshold the same for a human deputy and an AI agent?
• Is there a way to preserve some of delegation's benefit — not needing to specify every case in advance — while keeping more of automation's auditability, or are the two goods genuinely in tension?

Related Topics

• Shared Mental Models: Why Experienced Teams Need Fewer Words — the accumulated model of a delegate's likely behaviour this page's trust judgement depends on.
• Designing Auditable, Robust Agentic Systems — the ability leg of this page's trust model, worked out in technical detail for an AI delegate.
• When Agents Fail: Brittleness, Misaligned Incentives, and Deception — what goes wrong inside the vulnerability this page argues delegation can't avoid accepting.
• What the Preceding Pages Make Me Think About: Have We Created Digital Gods? — the same imported-values point about an agentic system's benevolence, argued at greater length from a different angle.
• How Much Memory Does a Temporary Team Need to Become a Community? — what happens to trust and delegation when the team itself is too short-lived to build the history this page's judgement relies on.

Further Reading

• Mayer, R. C., Davis, J. H., & Schoorman, F. D. (1995). An integrative model of organizational trust. Academy of Management Review, 20(3), 709–734.

References