An operations manager begins the morning with twelve agent-generated summaries. Each is polished, prioritised and slightly inconsistent. The manager spends an hour reconciling them, then misses a hesitant comment from a team member about a new failure pattern. Reporting volume has doubled. No decision improved. The tools have automated the production of managerial signals without designing the service those signals were meant to support.

Digital agents can assemble status, route commitments, watch thresholds, prepare evidence and generate alternatives. They can lower administrative load. They cannot accept institutional accountability, build reciprocal trust, own a consequence or decide which legitimate interest should lose in a difficult trade-off. Those boundaries do not make every present managerial activity valuable. They make it possible to remove low-value administration without confusing its disappearance with the disappearance of management.

Observation-based research described managerial work as fragmented, relational and embedded in information and action rather than as detached planning alone.12 Kotter likewise emphasised the agendas and networks through which general managers act.3 An agentic organisation should therefore redesign management as a portfolio of services delivered to teams, decisions and boundaries, then allocate support role by role.

01 / Clarify the roleFrom administrative signals to management services

Administrative activity often grows around a real service. A status deck exists because someone needs to see commitments and intervene. A performance narrative exists because someone must develop, reward or redirect a person. A staffing meeting exists because capacity and priorities conflict. The right redesign starts with the consumer and decision, not the document. If no decision or outcome uses the activity, remove it rather than automate it.

If the service is real, ask what information and relationship it requires. A machine-readable queue may support routine allocation. A weak signal about fear, conflict or workarounds may require direct observation and conversation. A policy exception may need evidence integration and legitimacy-bearing trade-offs. The service determines the delivery mode.

02 / The strong rivalAccept bounded algorithmic management where it works

Algorithmic management is a credible rival when objectives, rules and data are explicit. A stable rota constraint can be solved automatically from availability, skill, rest and coverage rules. Workers should see the constraints, correct their data and appeal an infeasible result. A manager handles exceptions rather than approving every valid schedule. Consistency and speed are valuable here.

The boundary changes when measures are incomplete or action affects dignity, opportunity and livelihood. Performance data can omit invisible help, difficult case mix, access needs or system failures. Objectives can conflict. A worker may need to contest not only a factual input but the interpretation placed on it. Automated allocation becomes unsafe when the person affected cannot understand, challenge and receive a reasoned decision.

Research on algorithms at work describes a contested terrain of control rather than an inevitable replacement of managers.12 Levels of interaction with automation also range from information acquisition to action, so there is no single generic human role.10 The design task is to bound the automated contribution, preserve appeal and locate accountable authority.

03 / Nine servicesThe managerial work ledger

Outcome framing and priority trade-offs translate strategy into a choice among competing results. Agents can assemble constraints and scenarios. Managers accept which commitment takes precedence and explain the loss. Delegation and permission design assigns authority, evidence and stop conditions. Agents can implement and monitor permissions, but accountable roles decide their scope.

Evidence and context integration combines formal data, local history and weak signals into a problem representation. A clerk can assemble evidence; a sentinel can surface anomalies; a manager tests meaning. Exception and conflict resolution receives cases outside ordinary rules and resolves interests, authority or ambiguity.

Coaching, feedback and judgement development changes future capability through reciprocal understanding, practice and consequence. An agent may prepare examples or recall agreed goals, but personalised text is not a coaching relationship. Team climate, trust and psychological safety depends on repeated conduct, fair response and the ability to raise risk without punishment. Edmondson linked psychological safety with learning behaviour in teams.7

Cross-boundary negotiation and resource commitment integrates teams and functions whose incentives differ. Agents can model options and track promises; authorised people commit resources and relationships. Risk, incident and consequence ownership decides when to stop, remediate, communicate and resume. Reflection, learning and operating-model change asks not only how to correct an error but whether objectives, policies and assumptions must change.

Operate the ledger service by service

Begin with the consumer. A service may support a team member, another manager, a decision forum, a control function, a customer or the manager's own future judgement. If the consumer cannot be named, investigate whether the activity persists only because a calendar or template demands it. Removing purposeless activity creates more capacity than automating it.

Name the decision and consequence next. A queue review might decide which customer population receives scarce capacity. A one-to-one might develop judgement, repair trust or resolve a conflict. A risk review may accept a temporary exposure. The consequence determines what evidence, authority, relational context and record the service needs. It also distinguishes useful preparation from decision substitution.

Frequency and current load should include interruption and preparation. A service that occurs weekly may consume daily attention through data chasing and informal questions. An agent clerk can reduce that load by maintaining a trusted evidence packet. If it produces an additional report beside the old packet, reporting inflation rises. The manager should retire the replaced artefact and measure net attention, not generated output.

Information required is more than a field list. Record authoritative sources, freshness, uncertainty, local history, people who hold context and signals that are not yet structured. Decide which inputs an agent can assemble and which the manager must obtain through direct observation or conversation. The service fails when the manager sees only an agent's interpretation of evidence they are accountable for judging.

Relational and legitimacy content needs a concrete description. Ask whether the service changes opportunity, status, workload, identity, trust, safety or a commitment between people. A resource choice can be analytically sound and still require negotiation with those who bear its cost. Agent preparation can make alternatives visible, but a manager must conduct the interaction when acceptance depends on reciprocal explanation and fair treatment.

Human authority should match consequence. Record what the manager may decide, what another role controls, how an employee contests the input and which action the agent can take. A manager who remains accountable for an incident but cannot access raw logs or stop the workflow holds nominal responsibility without practical control. Reallocation should restore the evidence and permission boundary before adding more automation.

Failure signals belong in the design before launch. For a clerk, signals include missing provenance, duplicated reporting and rising correction. For a sentinel, they include false-alert load, missed weak signals and managers ceasing direct observation. For a challenger, they include repetitive alternatives, anchoring and option volume without decision improvement. Relational services fail when contest is suppressed, trust declines or feedback becomes generic.

Learning effect is measured at both person and system level. Does the manager form better independent frames, understand more of the work and improve future delegation? Do employees develop judgement and know how decisions are made? Does the organisation revise an objective or control after consequences appear? Efficiency that removes these learning channels may make the next unfamiliar decision harder.

Finally, write a reassignment trigger. A clerk contribution can expand when rules and evidence remain complete. It should contract when exceptions reveal interpretation. A sentinel can gain a bounded automated response only after action, reversal and appeal are tested. A manager-led service can later use more support if relational or legitimacy content becomes safely separable. The boundary is present evidence, not a permanent claim about humans and machines.

04 / Three bounded agent rolesClerk, sentinel and challenger

A clerk assembles, retrieves, transforms and routes under stable rules. It can prepare a decision packet, update commitment state or solve the constrained rota. The output needs provenance and an acceptance test. The clerk should not quietly interpret an ambiguous objective.

A sentinel watches a defined field and surfaces a threshold, anomaly or missing receipt. It does not decide what the signal means. Alert design must control volume and give managers access to raw evidence and the work itself. A sentinel that optimises for detection can detach managers from direct observation.

A challenger produces an alternative problem representation, counter-plan or evidence test before the manager decides. Its value lies in disciplined disagreement, not in generating more options. The manager should form an initial view, inspect the challenger evidence and retain responsibility for the trade-off.

05 / Direct observationDashboards lose weak signals by design

Structured telemetry is valuable because it compresses volume. Compression also removes texture. A hesitant explanation, a workaround, an unusual customer need or a change in team energy may not fit the schema. If managers replace direct observation with dashboards, weak signals disappear before they can become evidence.

Dashboard detachment is not repaired by adding more metrics. Managers need deliberate contact with work: case sampling, open questions, conversations, boundary reviews and time where employees can raise context that contradicts the system. The agent sentinel should invite investigation, not declare meaning.

Design direct observation alongside telemetry

Set a sampling practice for each consequential service. A manager might inspect complete cases, join an operating hand-off, listen to a customer interaction or walk through an incident with the people involved. Sampling should include apparently healthy work, because weak signals often appear before a threshold fires. It should also include cases the model classifies with high confidence so the manager can test whether confidence hides a systematic blind spot.

Create an unstructured route that is still accountable. Employees need a way to raise a concern that does not fit a category, explain a workaround or challenge the premise of a dashboard. The route should reach a person with time and authority, and the response should be visible enough to build confidence that speaking changes something. The organisation can later structure recurring signals without closing the original channel.

Compare telemetry with observation explicitly. Record where direct work contact confirms, qualifies or contradicts the agent's signal. The purpose is not to prove that human intuition is superior. Managers can overlook patterns and carry bias. The comparison reveals which information source is reliable for which condition and improves the sentinel contract.

Attention is finite, so alert volume must be governed. Give each alert an owner, evidence link, service condition and disposition. Suppress duplicates and retire alerts that never change action. Preserve a route for severe anomalies even when normal alert budgets are reached. An overloaded sentinel converts potentially useful observation into another reporting queue.

06 / Relation and legitimacyCoaching and performance decisions require contestability

Synthetic coaching can sound personal because it uses a person's history and preferred tone. It still lacks reciprocal understanding unless a person can interpret what is unsaid, accept challenge, negotiate a goal and remain accountable for how feedback affects the relationship. Agents can prepare evidence, examples and follow-up. The manager conducts the consequential conversation.

Performance narratives create a sharper boundary. Before a narrative affects workload, opportunity, reward or career, the worker must see relevant evidence, correct factual errors, add context and challenge interpretation. The manager needs access to raw evidence and authority to reject the agent's framing. Contestability after the decision is weaker than contestability within it.

Protect the reciprocal part of management

Coaching begins with a joint account of the work. The manager listens for the person's goals, constraints and interpretation, then tests observations against evidence. Advice has meaning because it occurs inside a relationship with memory, commitment and future follow-up. An agent can retrieve examples and help a person rehearse, but it cannot make the manager's promise to provide opportunity or accept responsibility for an unfair condition.

Automated feedback should therefore be labelled as preparation or practice. It should not be inserted into a personnel record as if a responsible manager had observed and endorsed it. The employee should know what data shaped it, what the system cannot know and whether the manager has reviewed it. Material conclusions require a conversation where correction can change the result.

Contestability also protects managers from false certainty. A worker may explain that a performance pattern reflects difficult case allocation, inaccessible tools, invisible mentoring or a safety-preserving slowdown. This does not mean every explanation is accepted. It means the decision uses a complete adversarial record rather than treating system output as neutral fact.

The appeal route should be proportionate and non-retaliatory. Routine input correction can be immediate. A disputed interpretation may need another manager, people function or independent review. Record the disposition and feed verified corrections into the evidence system. Track whether particular groups or case types contest more often and whether appeals reveal a measurement defect.

Trust is not measured by agreement with recommendations. Employees may comply because the system appears unchallengeable or because dissent creates work. Look for speaking-up, quality of disagreement, correction before harm and whether managers explain decisions that depart from agent advice. These behaviours show whether human authority is active rather than ceremonial.

The NIST AI RMF Core includes human-AI configuration, feedback and affected-party considerations, while the EU AI Act gives particular significance to employment and worker-management contexts and human oversight.1314 These sources do not turn every conversation into a formal hearing. They reinforce the need for access, competence, authority and a usable route to challenge.

07 / Managerial learningRecommendations can improve or atrophy judgement

A manager develops judgement by forming a problem representation, choosing, observing consequence and revising assumptions. If agents always present polished options first, anchoring can replace independent framing. The manager remains accountable but loses the practice needed to exercise that accountability during an outage or novel case.

A challenger should enter at a deliberate point. For high-consequence work, the manager can record an initial frame, then inspect the counter-plan and evidence. The decision receipt notes what changed. Periodic unaided exercises and after-action reviews test retention. Bainbridge's automation irony applies: the more routine work automation absorbs, the harder the remaining exceptions may be for a person kept outside ordinary practice.11

Maintain a managerial judgement cycle

Before using a challenger, the manager writes the objective, constraints, uncertainties and likely action in a short decision note. The challenger then tests assumptions, introduces evidence or constructs a materially different option. This ordering limits anchoring by polished machine output and makes the agent's contribution observable. It also reveals when the challenger merely rephrases the initial view.

After action, connect the decision with consequence. Review expected and actual outcomes, affected people, boundary effects and any repair. Ask whether the problem was framed correctly and whether the governing objective or permission needs revision. Argyris and Schön's organisational learning perspective makes this distinction central: learning can correct action or change the theory that produced it.9

Use unaided practice on representative cases, not artificial puzzles disconnected from work. Managers should be able to gather essential evidence, form a frame, choose within authority and activate fallback during an outage. The standard is safe service, not matching agent speed. If capability declines, narrow the agent contribution, restore practice or add expert support before the next tail event.

Homogenisation is another failure signal. If managers across units choose the same apparently optimal option despite different context, inspect whether agent recommendations are collapsing legitimate variation. Compare rejected alternatives and talk with those who experienced the outcome. Consistency is valuable for rules; it can be destructive when different local commitments require different trade-offs.

08 / Worked managerAn operations day with bounded support

An operations manager is accountable for queue health, staffing trade-offs, employee development, consequential exceptions and recovery. A clerk assembles verified queue evidence, open commitments and staffing constraints. It removes manual status collation. A sentinel surfaces an unusual pattern in repeat contacts and links the underlying cases. It does not label the cause.

Before a staffing decision, a challenger produces an alternative allocation and tests the manager's assumptions against service, cost and wellbeing constraints. The manager has already framed the decision, inspects both plans, talks with team leads and commits the trade-off. The receipt records expected outcomes and when to reverse.

The manager personally investigates the anomaly, coaches a new team lead, resolves a policy exception and negotiates temporary capacity with another function. Agent support prepares evidence but does not conduct synthetic coaching or make the resource commitment. At the end of the week, the manager reviews which recommendation changed judgement and whether direct observation contradicted telemetry.

The counter-case is the stable rota. A deterministic service applies transparent constraints, shows inputs and produces a schedule. Workers correct availability and appeal conflicts. The manager handles exceptions and changes rules when outcomes reveal a problem. Requiring manual approval of every valid rota would add management activity without improving a decision.

Measure the redesigned managerial week

The operations manager records time and outcome by service rather than by application. Clerical assembly should fall without a parallel rise in reconciliation. Sentinel alerts should lead to useful investigations at a manageable rate. Challenger use should change or strengthen a decision often enough to justify review. Direct observation, coaching, boundary negotiation, recovery and reflection should remain visible rather than being treated as residual calendar space.

Employees provide evidence too. Ask whether they can reach the manager, correct system interpretations, understand delegation and receive useful development. Compare what the dashboard says with where teams report friction or silence. If administrative time falls but access, trust or coaching declines, the reallocation has not improved the complete service portfolio.

Downstream outcomes close the ledger. Track whether decisions resolve the issue, whether resource promises hold, whether incidents recur and whether affected people understand the result. An agent contribution earns continued use when it helps the manager deliver these services with less avoidable load and no loss of authority, context or learning. Reporting volume, response speed and generated options are intermediate measures only.

The weekly review also identifies activity to remove. A summary that nobody uses, a meeting that duplicates a visible commitment or an approval that never changes a complete rule should not survive because automation made it cheap. Stop it, observe the effect and restore only a service with a real consumer, decision and acceptance condition. That distinction matters.

09 / Interactive allocationManagerial Work Reallocation Test

The test asks whether an activity produces a real decision or outcome, which agent role fits, what information and relation the service requires, and whether managers and employees retain evidence, authority, contestability and learning. It has no managerial maturity score.

How the controls route

Complete rules and structured data route clerical processing while a manager interprets exceptions. Threshold monitoring routes to a sentinel and investigation. A bounded alternative routes to a challenger before decision. Coaching, career impact, weak contestability, lost authority, dashboard-only evidence or judgement atrophy keep the service manager-led. Reporting that serves no decision is removed rather than automated.

10 / Management after administrationMeasure decisions, development and outcomes

The manager in an agentic organisation should not become the final approval node for a machine-generated stream. Nor should the role be protected by romantic claims that every judgement is ineffable. Routine status, evidence assembly, constraint solving and threshold observation should move when the rules and failure boundary are clear.

What remains is more demanding: frame competing outcomes, design authority, interpret context, resolve exceptions, develop judgement, maintain trust, negotiate boundaries, own incidents and change the operating model. Agents can contribute evidence, signals and disciplined alternatives to each service. They do not absorb the manager's accountability by appearing knowledgeable.

Build the ledger service by service. Remove activity with no consumer or decision. Type the clerk, sentinel or challenger contribution. Preserve direct observation, contestability and reciprocal conversation. Test unaided judgement and fallback. Give managers the authority and evidence that match the consequences they own. Then measure management by better decisions, stronger people, healthier boundaries and learning from outcomes, not by how many messages a dashboard helps one person process.

SourcesReferences

1. Mintzberg, H. (1971). “Managerial Work: Analysis from Observation.” Management Science, 18(2), B97–B110. DOI.
2. Mintzberg, H. (1973). The Nature of Managerial Work. Harper & Row. Publisher record.
3. Kotter, J. P. (1982/1999). “What Effective General Managers Really Do.” Harvard Business Review. Publisher record.
4. Simon, H. A. (1947/1997). Administrative Behavior. Simon & Schuster. Publisher record.
5. Garicano, L. (2000). “Hierarchies and the Organization of Knowledge in Production.” Journal of Political Economy, 108(5), 874–904. Verified publisher DOI.
6. Galbraith, J. R. (1974). “Organization Design: An Information Processing View.” Interfaces, 4(3), 28–36. DOI.
7. Edmondson, A. (1999). “Psychological Safety and Learning Behavior in Work Teams.” Administrative Science Quarterly, 44(2), 350–383. DOI.
8. Argyris, C. (1977). “Double Loop Learning in Organizations.” Harvard Business Review. Publisher record.
9. Argyris, C. and Schön, D. A. (1978). Organizational Learning: A Theory of Action Perspective. Addison-Wesley. Library record.
10. Parasuraman, R., Sheridan, T. B. and Wickens, C. D. (2000). “A Model for Types and Levels of Human Interaction with Automation.” IEEE Transactions on Systems, Man, and Cybernetics Part A, 30(3), 286–297. DOI.
11. Bainbridge, L. (1983). “Ironies of Automation.” Automatica, 19(6), 775–779. DOI.
12. Kellogg, K. C., Valentine, M. A. and Christin, A. (2020). “Algorithms at Work: The New Contested Terrain of Control.” Academy of Management Annals, 14(1), 366–410. DOI.
13. NIST. AI Risk Management Framework Core. Official resource.
14. European Union (2024). Regulation (EU) 2024/1689, Artificial Intelligence Act. Official Journal text.
15. OECD, BCG and INSEAD (2025). The Adoption of Artificial Intelligence in Firms. Official publication.