Decide what AI should do.
Good AR strategy starts with the work and its consequences, not the latest model or agent.
Ask the harder question first.
Before introducing an artificial resource, be explicit about the task, the result, and the human responsibility it is intended to support.
The aim is not to automate every possible activity. It is to choose an operating model that delivers the best combined outcome for customers, workers and the organisation.
Work first.
Technology second.
Define success before selecting a provider. Include quality, rework, supervision and workforce effects in the business case.
Four practical choices
Most tasks are not simply “human or AI”. The operating model can be designed along a spectrum of authority and oversight.
| Operating choice | Suitable characteristics | Required management attention |
|---|---|---|
| Human-led | Highly ambiguous, sensitive, relationship-centred or difficult-to-reverse decisions. | Support the person with information and specialist tools without weakening accountability. |
| AI-assisted | Research, drafting, analysis or triage where a person can judge output quality. | Measure review effort, overreliance, skill retention and acceptance criteria. |
| Human-supervised automation | Repeatable activity with meaningful exceptions and a reliable escalation pathway. | Define boundaries, sample checks, interruption rules and exception ownership. |
| Bounded automation | Low-consequence, well-specified, reversible actions with proven monitoring. | Test unusual conditions, monitor drift and maintain a practical rollback route. |
Measure net benefit,
not simply time saved.
An attractive demo is not a business case. Compare against the genuine human-led baseline and include ongoing resource use and operational consequences.
How much usable work is produced after verification, correction and rework?
Models, tokens, compute, integration, security, review and incident response.
Job design, learning, trust, workload, autonomy and the quality of retained roles.
Consequences of error, uncertainty, vendor lock-in and ability to switch off.
Where value can be created
- Deploy scarce specialist capacity where it genuinely improves results.
- Use incremental pilots to learn before committing large budgets.
- Redesign work so people can concentrate on judgement, relationships and improvement.
What needs attention
- Automate an ill-defined process and scale its weaknesses.
- Mistake reduced task time for better end-to-end service.
- Underestimate integration, model drift and hidden human checking.
A one-page case for using AI
Before investment, capture the intended purpose, alternatives considered, authority boundaries, business outcome, human impact, principal risks, evidence required and review date.
Get the decision templates ↗Five questions for the sponsor
- What improvement are we buying, and how will we know?
- Who remains accountable for the outcome?
- What happens when the system is wrong or unavailable?
- How will working life change for the people involved?
- What evidence would make us stop or change course?
Turn a decision into an operating model.
Move from the business case to the lifecycle of the models, agents, workflows and compute you will manage.