What to Automate First With AI (and What to Leave Alone)

Start with what is countable

The tasks worth automating are boring and measurable: someone retypes supplier invoices into the accounting system, someone reads support email and decides who it goes to, someone rebuilds the same three reports every Monday morning. Each is a countable number of hours, which is what makes the business case checkable rather than aspirational — and what we baseline before anything is built.

What usually gets automated

• Document automation — invoices, delivery notes and forms read and posted into your system
• Data-entry automation between two systems that do not talk to each other
• Support automation — classification, routing, suggested replies, with escalation
• Reporting automation — the recurring pack built and delivered on a schedule
• Approval workflows routed by amount, type and history, with exceptions raised rather than auto-approved

The human-in-the-loop rule

Every automation is classified before it is built, and the classification decides the interface: automatic for reversible, low-value, high-volume work; suggested where a person confirms with one click; assisted where the consequence is financial or legal; exception-only where the system handles the ordinary case and escalates anything unusual.

Does it always need AI?

No, and a good deal of what gets automated does not use it. If the rule can be written down, ordinary code runs it faster, cheaper and more predictably than a model. AI is for the step where the input is unstructured or the judgement is genuinely fuzzy — using it anywhere else is a cost with no return.

If the process is not written down anywhere

Then that is the first phase, not the automation. Documenting what actually happens — including the exceptions — usually takes a week or two with the people who do the work, and it is common for that alone to remove a chunk of the original problem.