How to automate admin work: a triage method, not a tool list

To automate admin work, triage it first: log every recurring admin task for one week, then rank each by volume, by how statable its rules are, and by what happens if it is wrong. Automate the highest-volume, most rule-bound, lowest-consequence task first — one workflow, proven end to end — and only then move to the next. The tool is the last decision you make, not the first.

Most automation projects in Australian SMBs fail in the opposite order. Someone buys a platform, then goes hunting for something to automate with it. Six weeks later there are four half-built flows, nobody trusts any of them, and the admin pile is exactly where it was.

Step 1: log the week, don't estimate it

Estimates are wrong in a specific direction — the annoying tasks feel bigger than they are, and the constant small ones disappear. For five working days, write down every recurring admin task, who did it, and roughly how long it took. Nothing fancy; a shared note is enough.

What comes out of this is usually uncomfortable. The task that everyone complains about turns out to be forty minutes a week. The thing nobody mentions — re-typing enquiry details from the form into the CRM, chasing the same status update, sending the same three-line reply — turns out to be six hours.

Step 2: score each task on three tests

Every candidate gets the same three questions. They are the same three that decide whether a workflow is agent-shaped at all, covered in more depth in the operator's definition of an AI agent.

Three yeses and real volume means the task is a candidate. Two yeses means it can often be reshaped into one — usually by fixing the intake, not the automation. One yes means leave it alone and stop paying for a tool that pretends otherwise.

Step 3: rank by payback, not by irritation

Rank the surviving candidates by hours per month multiplied by how repeatable they are, and break ties with consequence-if-wrong: a task where a human can check the output before it goes anywhere is a much better first automation than one that touches money or a client relationship unsupervised.

The winner is almost never the most annoying task. It is usually the boring high-frequency one — the first reply to every enquiry, the extraction of every intake form, the same reminder to every booked client. Volume is what pays back, and volume is exactly what a person is worst at doing consistently.

Step 4: automate one thing, supervised, then move

Build the top-ranked workflow only. Run it with a human approving output until the exceptions stop surprising you, then let it run and check the log. This is the whole method in how we scope and build agents — scope one workflow, prove it under supervision, hand it over, then take the next one off the list.

Document-heavy businesses tend to find their first candidate in intake: migration agents automating visa document intake are a clean example, because the input is readable, the checklist is statable, and every action needs to be logged anyway.

What this method rules out

It rules out the tool-first purchase, the automate-everything project, and the pilot with no owner. It also rules out automating work that should simply be deleted — a report nobody reads is not an automation candidate, it is a decision nobody has made yet.

If the triage says your admin pile is varied and low-volume, the honest answer is that a person is still the right tool. That is the same crossover covered in the AI agent versus virtual assistant comparison. The method is worth running either way — you end the week knowing which it is, with a ranked list instead of a feeling. A Foundations Session is that triage done with you, in writing, with payback math attached.

Questions people actually ask