Design the work around people.
The quality of human work and wellbeing are core outcomes of AR management — not side effects to consider later.
Assistance should not become alienation.
AI can remove drudgery, improve access to expertise and help people do work they value. It can also turn skilled roles into repetitive verification, reduce autonomy or make performance pressure more intense.
Both possibilities belong in the original design decision.
“Is the organisation becoming more capable — and are its people better able to do good work?”
Do not infer employee benefit from productivity alone. Ask workers directly and test the lived reality of the new workflow.
Better human–AI work.
Preserve judgement
Make clear where professional discretion, empathy and accountability belong to people.
Make review meaningful
Give reviewers sufficient context, time and authority to challenge AI output.
Protect learning
Maintain opportunities for junior staff to practise skills AI might otherwise absorb.
Measure the experience
Monitor workload, stress, confidence, autonomy and perceived fairness alongside speed.
Make change transparent
Explain what the system does, where it is used and what decisions remain human.
Include workers early
Co-design workflows with those who know their real exceptions, dependencies and quality criteria.
Four models of collaboration
| Human role | AI contribution | Potential failure | Design response |
|---|---|---|---|
| Creator | Research, drafting and options. | Human creativity becomes limited to accepting suggestions. | Protect original thinking and time to question the output. |
| Reviewer | Generates work for human approval. | Rubber-stamping and alert fatigue. | Use risk-based sampling, evidence and permission to reject. |
| Supervisor | Performs bounded autonomous tasks. | Unclear visibility and weak intervention controls. | Define dashboards, escalation, pause and rollback. |
| Specialist | Expands analytical and knowledge capacity. | Skills erosion or misplaced confidence. | Keep expertise development and periodic independent checks. |
Where value can be created
- Give staff more scope to solve difficult problems and improve services.
- Improve accessibility and provide timely assistance for less experienced workers.
- Build new roles in supervision, evaluation, integration and organisational learning.
What needs attention
- Increased output expectations remove any benefit from time saved.
- Monitoring or ranking workers using AI damages trust or fairness.
- Routine reliance on AI erodes domain knowledge and professional confidence.
Include wellbeing in the scorecard.
Consider worker-reported workload, sense of autonomy, confidence in the system, opportunity to learn, quality of interpersonal work and the burden of exception handling.
These are not interchangeable with productivity or output measures, and some may require careful qualitative assessment.
Explore the human-impact measure ↗Questions to ask workers
- Which tasks have improved, and which have become harder?
- Can you explain and challenge the system's output?
- Has work become more satisfying or simply faster?
- Are skills and professional judgement still developing?
- Do you have a safe route to report AI-related concerns?
Build confidence without losing judgement.
Human oversight only works when it is designed as a real control rather than an approval box.