Automate a controlled step, not an unknown process
Automation can remove repetitive handling and create time for more valuable work, but it can also repeat an error quickly, conceal a decision or expose data. Start with a stable, understood step and preserve human accountability.
Choose a repetitive, rules-based step with known inputs and exceptions. Map the current process, define the quality threshold and owner, minimise data, test normal and failure cases, monitor results, keep a manual fallback and review whether the time saved improves the intended outcome.
Use a proportionate automation gate
| Question | Proceed when | Pause when |
|---|---|---|
| Process stability | The step, inputs, rules and exception route are understood. | The process changes often or different people perform it differently. |
| Consequence | An error is low consequence or will be reviewed before it matters. | The output affects rights, safety, money, sensitive records or a critical decision. |
| Data | Only necessary data is used with appropriate access and retention. | Data purpose, privacy, ownership or supplier terms are uncertain. |
| Control | There is an owner, monitoring, correction route and usable manual fallback. | No one can explain, stop, repair or reverse the automated result. |
Test more than the happy path
- Normal input and representative volume
- Incomplete, duplicate, late or malformed input
- Unusual but legitimate case
- Unavailable dependency or supplier delay
- Incorrect output, duplicate action or unintended recipient
- Access, audit trail and alert route
- Stop, correction and manual fallback procedure
AI needs extra caution
Where automation uses AI, do not treat fluent output as verified work. NIST frames AI risk management around govern, map, measure and manage; the functions are not a substitute for case-specific judgement. Organisations subject to UK law should also consider the ICO guidance on AI and data protection.
Safe automation FAQs
What work is safest to automate first?
A stable, repetitive, low-consequence step with clear inputs, a known exception route and an easy human check. Start with a bounded pilot instead of automating an entire workflow.
Will automation always save time?
No. Setup, maintenance, exception handling, review, data work and training can outweigh savings. Measure whole-life effort and the quality of the outcome, not only the duration of the automated step.
Who is accountable for an automated outcome?
Name an accountable person or role before launch. Supplier responsibility or tool capability does not remove the organisation responsibility for the decision, monitoring, correction and fallback.
Continue your productivity decision
Use the next guide that fits the bottleneck, measurement, work design, communication, attention, automation, knowledge or review question still in front of you.