Do not begin with a catalogue of tools
A new model or automation platform can make almost any process look like an opportunity. That is exactly why the starting point should be the work: what happens, who owns it, where time is lost and which mistakes matter.
Tool-first projects often create a polished demonstration without solving the hand-off, decision or data problem underneath it.
Use a five-part workflow filter
1. Frequency
How often does the task happen? A small saving repeated every day may matter more than a dramatic saving once a quarter.
2. Friction
Where do people wait, re-enter information, search for context or correct predictable mistakes? Map the actual path rather than the ideal process described in a document.
3. Variability
Some work follows stable rules; some requires judgement because the inputs or consequences vary. Automation should reflect that difference instead of treating every case as identical.
4. Consequence
What happens if the system is wrong, late or unavailable? Higher-consequence work needs stronger review, fallback and monitoring.
5. Ownership
Who can approve the process, provide access, judge output quality and maintain the system after launch? If ownership is unclear, the technology will not fix it.
A strong first automation is valuable enough to matter, contained enough to understand and observable enough to improve.
Keep a person at the right control points
Human oversight should be designed around decisions, not added as a vague reassurance. Specify where approval is required, what information the reviewer sees and how corrections feed back into the process.
For an enquiry workflow, the system might collect context, classify the request and prepare a response. A person may still approve the message when commercial judgement, sensitive information or a commitment is involved.
Define success before building
Useful measures might include handling time, delay between steps, correction rate, completion rate or the amount of manual preparation. The measure should connect to the original problem, not simply count how many AI calls were made.
Write down one repeated process in six to ten steps. Mark each wait, re-entry, judgement and failure point. The best opportunity often becomes clearer before a tool is selected.
Build a controlled first version
Use representative examples, test difficult cases and agree what happens when information is missing. Launch to a limited group, watch the workflow and make the next decision from evidence.
The aim is not to remove people from a process at any cost. It is to reduce avoidable work while keeping responsibility visible.