Automate the handoff, not the judgment

Many teams do not need a dramatic AI transformation. They need fewer hours copying values between systems, checking whether a file arrived, reminding someone about an approval, renaming attachments, and rebuilding the same weekly report. The work is small, but it returns every day and grows with the business.

Google SRE describes toil as work that is manual, repetitive, automatable, tactical, and that scales with service growth. The same lens is useful outside infrastructure. If a process requires another person each time volume doubles, it is worth examining before hiring around it.

Write the current process as events and decisions: what starts it, which data is required, who can approve, what can fail, and what proves completion. This often reveals that the painful part is not the decision itself. It is collecting inputs, routing the request, updating status, and notifying the next person.

Those handoffs are strong automation candidates because the correct result is visible. A system can validate required fields, create a record, request approval, and send a receipt without pretending to decide whether a high-value exception should be accepted.

Choose a process with enough volume to matter, a clear owner, and an output that can be checked quickly. Automate the normal path first and make exceptions obvious. A smaller loop reaches production sooner, generates real evidence, and limits the damage when an assumption is wrong.

Avoid rebuilding every legacy rule in the first version. Some steps exist only because the old tools could not share state. Preserve the business control, not necessarily the workaround that used to implement it.

Actions involving payment, deletion, permissions, or customer communication should be idempotent where possible, logged, and reversible. Use explicit approval for unusual amounts or low-confidence matches. A person should see what the system plans to do and the evidence behind it, not a vague “approve automation” button.

Recovery is part of the experience. Define who receives an alert, what context they need, and how the process resumes after correction. Silent failure turns saved minutes into hours of investigation.

Track cycle time, manual touches, exception rate, failed runs, and time spent maintaining the workflow. A process that saves five minutes but creates a weekly debugging ritual is not finished. The best automation becomes boring: it handles the repeatable path and gives people a clean queue for everything else.

Revisit the workflow as the business changes. Rules that were safe at fifty transactions may be fragile at five thousand. Automation creates leverage, and leverage deserves the same ownership and review as any other production product.