A field-service company gives coordinators an AI tool that proposes daily schedules. The tool accounts for appointments, travel estimates, and available staff. A coordinator notices that one proposed visit requires specialist equipment that the assigned technician does not carry. They can edit the schedule, but the interface records the change as a deviation and requires a manager’s approval before dispatch.
The coordinator formally retains a role in the decision. Whether that role provides meaningful discretion depends on what they can inspect and change, how quickly approval is available, and how the organisation evaluates departures from the recommendation. The same scheduling capability could also be designed to help coordinators explore alternatives and identify missing constraints.
Worker autonomy concerns the scope people have to make meaningful choices about their work within legitimate organisational requirements. AI can expand or restrict that scope. The outcome depends on task design, decision rights, interfaces, incentives, and the operational circumstances in which people use the tool.
Identify Which Decisions Are Moving
Autonomy is more specific than a general promise that people remain in control. A coordinator may choose appointment timing while having no discretion over contractual service commitments. A technician may choose a diagnostic approach while following mandatory safety requirements. Those boundaries can be appropriate and should be explicit.
Map the decisions before and after adoption. Who identifies the problem, selects the approach, applies the relevant constraints, and authorises action? Which choices become recommendations, defaults, automatic steps, or matters requiring approval? This reveals where authority changes rather than assuming that all assisted work has the same effect.
For the scheduling tool, the change may be modest if coordinators retain access to the inputs and can amend assignments within agreed rules. It is more substantial if they receive only a completed schedule and are expected to approve it without examining trade-offs. Both arrangements use AI, but they allocate judgement differently.
The OECD’s analysis of AI and job quality reports potential benefits alongside risks to autonomy and work intensity. Its evidence does not support an inevitable decline in discretion. Findings depend on the setting, and reported associations between implementation practices and outcomes should not be treated as universal causal rules.
An Override Needs More Than an Available Button
A person can have nominal permission to reject a recommendation while facing practical barriers to doing so. They may lack source information, time, access to an alternative path, or an approver who can respond within the operational window.
Some documentation for overrides is useful. The scheduling company needs to know why an assignment changed, particularly where contractual or safety requirements matter. The question is whether recording the reason supports the work or creates unnecessary friction that discourages justified changes.
For the missing equipment, an authorised coordinator could correct the assignment, record the constraint, and send the issue to the team maintaining the scheduling data. Requiring a separate approval may still be appropriate for a change outside their spending or staffing authority. It should not follow automatically from the fact that the proposal came from a model.
Distinguish the reasons for intervention. Correcting an incorrect input, choosing among acceptable alternatives, and overriding a mandatory restriction are different actions. Grouping them all under a deviation count makes it harder to assess professional judgement and can make appropriate correction look like noncompliance.
Defaults and Performance Measures Shape What People Actually Do
An interface can make one path much easier than another. If accepting a schedule takes one click while changing it requires re-entering several appointments, the design favours acceptance even without an explicit instruction. That may improve speed in routine cases while making necessary adjustments harder.
Acceptance rates alone do not show good use. A high rate might reflect strong recommendations, a narrow set of cases, missing information, or pressure to keep work moving. A low rate might indicate poor recommendations or a workflow that intentionally uses AI to generate options. Inspect outcomes and reasons before drawing conclusions about either staff or the model.
The company should compare dispatch timeliness with assignment suitability, repeat visits, corrections, and relevant service outcomes. If coordinators improve the result by editing recommendations, that work should be visible as a contribution. Otherwise the performance system may penalise the activity needed to keep the service reliable.
This is a specific instance of the relationship between metrics and judgement. The useful measure is tied to an outcome and interpreted in context. A tool-usage target should not silently become a requirement to accept the tool’s substantive decisions.
Make Review a Substantive Task
Reviewing a proposal can require considerable expertise. A coordinator who compares routes, checks constraints, and weighs service priorities is making decisions, even if a tool prepared the initial schedule. Treating all review as passive validation understates the work and obscures what the reviewer needs.
The role becomes weaker when the person can check only superficial features or is held responsible for matters they cannot assess. For example, a coordinator cannot reliably verify technician availability if the interface hides the source calendar and there is no route to resolve a conflict. Adding their approval to the record does not fix that information gap.
Set review expectations around decisions the person can actually influence:
| Part of the workflow | Meaningful discretion requires |
|---|---|
| Defining the task | A way to identify missing needs or an unsuitable objective |
| Assessing the proposal | Relevant evidence, understandable constraints, and sufficient time |
| Choosing an action | Authority to amend, reject, or escalate within clear boundaries |
| Improving the process | A route to report recurring defects and obtain a response |
These conditions should be tested with people doing the work. A demonstration may show that an override is technically possible without revealing whether it can be completed during a busy dispatch period. Observe representative use and ask which constraints the trial failed to capture.
Preserve Participation in How the Workflow Changes
Workers can contribute information that a central implementation team lacks: recurring exceptions, misleading data, the sequence in which evidence becomes available, or tasks that need a different approval route. Consultation should reach those design decisions while there is still scope to change them.
For the scheduling team, involve coordinators and technicians in defining unavailable equipment, travel uncertainty, and appointment constraints. Ask how disagreements are resolved and how a recurring problem reaches someone who can change the configuration. Participation is more useful when the organisation explains what it changed in response and what constraints remain.
Different tasks may justify different levels of automation. A well-defined scheduling adjustment may proceed automatically within agreed limits, while an assignment affecting a specialist commitment needs review. Increasing discretion in one area does not require removing every control in another.
Revisit these boundaries as the system expands. Permission settings and performance expectations can change after a successful pilot, leaving the actual job different from the one staff initially agreed to test. Communicate material changes and evaluate their effects on the work.
The design should account for time as well as formal permission. A coordinator who may correct an assignment only after an approver responds can lose the practical ability to resolve a problem before dispatch. Define which changes are within the coordinator’s delegated authority, which require another decision, and what temporary action is permitted while that decision is pending. The appropriate boundary follows the consequences of the change.
Also distinguish necessary standardisation from unnecessary restriction. A common format for recording appointments may make coordination easier without prescribing every scheduling choice. An enforced equipment requirement may protect a legitimate service constraint. By contrast, preventing staff from proposing an alternative route merely because the model ranked it lower needs a clearer operational justification. Discussing these differences makes autonomy concrete enough to design and review. It allows the company to preserve useful consistency while retaining the local knowledge and discretion needed to handle cases the standard proposal does not adequately address.
Keep Capability and Authority Connected
Autonomy and expertise are related but distinct. A worker may have authority without enough practice to use it confidently. They may also retain deep expertise while lacking permission to act. The guide to AI and tacit knowledge examines the learning opportunities that sustain capability.
Bainbridge’s Ironies of Automation highlights difficulties when operators must take over abnormal situations while automation changes their ordinary work. For the scheduling company, a relevant question is whether staff can manage the agreed fallback and understand the state left by the system. The answer should come from appropriate practice and exercises, not an assumption that human presence supplies unlimited backup.
The responsibility assigned to coordinators should reflect their information and decision rights. Problems with model configuration, staffing, or source systems may need owners elsewhere. Escalation should connect those responsibilities rather than leave the frontline worker accountable for every consequence of a schedule they could not meaningfully change.
AI can reduce repetitive work and provide useful alternatives while preserving professional judgement. It can also make a job more tightly prescribed if proposals become compulsory defaults and correction becomes difficult. The organisation can assess the difference by following actual decisions: what people can see, what they can change, how their contributions are measured, and whether the process responds when they identify a better course.
