Redesigning roles around human–agent collaboration

“The agent drafts; the human reviews” is not a role design. It names two activities while leaving workload, evidence, authority, skill and accountability unspecified.

When an agent changes a workflow, it also changes what people notice, practise, decide and learn. Removing routine tasks can create capacity for better work. It can also strip away the experience through which people learn to recognise exceptional cases. A reviewer may inherit more decisions per hour while receiving less firsthand context.

Design collaboration at task and decision level. Then rebuild roles around a coherent set of responsibilities, not around the list of activities the agent appears able to perform.

Map work beyond the formal procedure

Sample real cases and identify:

• information gathering and validation;
• sense-making and diagnosis;
• rule application and judgement;
• coordination and relationship work;
• approval and commitment;
• physical or digital action;
• exception handling and recovery;
• explanation, care and redress; and
• learning and process improvement.

Note tacit work: knowing which source is stale, recognising a request that signals a larger problem, adapting language to a distressed customer or asking a colleague about a rare exception. If the map contains only system steps, the redesign will undervalue the context that makes the workflow function.

OECD case studies show that AI can displace tasks while creating new work such as interpreting outputs, prioritising responses and feeding decisions back into the system (OECD). Role design must account for the whole shift, not subtract an automated task from a job description.

Allocate by properties, not prestige

For each task, assess:

• variation and ambiguity;
• consequence of error;
• need for contextual or relational judgement;
• data availability and quality;
• repeatability and volume;
• time sensitivity;
• ability to verify the outcome;
• authority required; and
• learning value for the worker.

Agents may be strong at searching broad material, generating alternatives and maintaining a consistent routine. People may be essential where legitimacy, empathy, contested values, novel exceptions or accountable commitment dominate. Deterministic software may be best for stable rules and transaction integrity.

Do not equate “human” with high-value and “machine” with routine. Repetitive checking can be a crucial control. Creative generation can be low consequence. Allocation should follow workflow properties.

Use a five-part collaboration contract

For each human–agent handoff, define:

1. Request: what the person or system asks the agent to do.
2. Evidence: what inputs the agent may use and must expose.
3. Proposal: what output, action or uncertainty the agent returns.
4. Decision: who accepts, changes, rejects or escalates it.
5. Learning: how corrections update cases, guidance or system design.

Consider contract review. The agent can identify clauses, compare them with a playbook and draft issues. The reviewer must see the clause, source playbook version, deviation and missing context. Legal or commercial authority remains with the eligible decision-maker. Corrections should distinguish extraction error, policy ambiguity and a genuinely novel negotiation point.

Without this contract, “review” becomes a catch-all transfer of responsibility to the person at the end.

Design authority separately from task allocation

An agent may perform most steps without owning the decision. A person may supervise the workflow without authority to approve the final commitment. Map five roles:

• initiator who supplies intent;
• operator who manages the case;
• reviewer who assesses evidence;
• approver who holds decision authority; and
• accountable owner who governs the service.

NIST emphasizes differentiating roles in human–AI configurations and notes that outcomes vary by configuration (Human–AI Interaction). Avoid the “moral crumple zone” in which a nominal human reviewer is blamed for a system they could not meaningfully understand or control.

Specify time, evidence, permission and alternatives for each role. If a reviewer has seconds to approve hundreds of cases, human oversight exists in the diagram but not in operation.

Preserve expertise and meaningful practice

Automation changes how expertise develops. Junior staff often learn by handling ordinary cases, seeing feedback and gradually encountering exceptions. If the agent takes all routine work and people receive only the hardest anomalies, development can become unsafe and discouraging.

Design learning routes:

• sample and review normal cases, not only failures;
• rotate people through assisted and unassisted work;
• use explanations as prompts for investigation, not answers to memorize;
• preserve simulations and case libraries;
• involve practitioners in evaluation and incident review;
• track whether people can still perform critical fallback tasks; and
• allocate time for improving rules, prompts, data and process.

The goal is not to keep obsolete work for tradition’s sake. It is to ensure that the future workflow can still produce the expertise it depends on.

Watch for deskilling signals after rollout: reviewers taking longer on unaided cases, declining ability to explain policy, repeated acceptance of familiar agent errors, or dependence on a small group of experts for every exception. These signals should trigger changes to allocation, coaching or interface design. They may also justify retaining a controlled manual exercise or requiring reviewers to form an initial view before seeing the agent’s proposal. The right intervention depends on the role; the point is to measure expertise as an operating dependency instead of assuming it will persist unchanged.

Measure workload redistribution honestly

Agent introduction can remove drafting while increasing review, exception, coordination and correction. Build a before-and-after workload map:

Work	Change to observe
Routine processing	volume and active time
Verification	evidence checks and review time
Exceptions	frequency, complexity and queue age
Coordination	handoffs and interruptions
Correction	downstream rework and appeals
Learning	coaching, analysis and improvement time
Emotional demand	difficult interactions and responsibility

Avoid claiming capacity release from model completion time alone. Measure end-to-end labor and outcome. If the agent makes first drafts rapidly but reviewers spend longer verifying them, the workflow may have shifted effort rather than reduced it.

ILO research on algorithmic management reports both process benefits and risks to job quality and intrusive surveillance in studied workplaces (ILO). Monitor autonomy, pace, contestability and worker experience alongside output.

Involve workers in design and correction

Invite practitioners and affected groups to shape task boundaries, test cases, interfaces, exception routes and measures. Make participation consequential: record issues, decisions and why suggestions were accepted or declined.

OECD research associates consultation with better perceived usability, acceptance and risk mitigation (OECD). ILO’s global case studies document social dialogue influencing workplace AI decisions (ILO). Participation does not remove management decisions; it improves the evidence on which they are made.

Provide routes to contest both agent output and the design of the workflow. People should be able to report systematic problems without being measured as resistant adopters.

Rewrite roles around outcomes and decisions

A revised role description should name:

• outcomes the role owns;
• decisions and authority limits;
• agent-supported tasks and required checks;
• exception and escalation responsibilities;
• collaboration with other roles;
• evidence and records to maintain;
• proficiency and fallback capability;
• improvement contribution; and
• how performance will be judged.

Do not measure a reviewer by agent acceptance rate. That rewards compliance. Measure decision quality, appropriate challenge, resolved exceptions, outcome and contribution to learning.

Human–agent collaboration works when the design makes it possible for both sides to do what they are assigned. The agent needs bounded tools and observable state. People need context, time, authority, recourse and a role worth performing.