The manual tasks worth handing to an agent first

The temptation, when you first get serious about automation, is to automate everything at once. Don't. The fastest route to disappointment is a half-built robot trying to do twelve jobs badly.

The shortlist test

We score a candidate process on three questions:

1. Is it frequent? Something done daily pays back faster than something done quarterly.
2. Is it rule-based? Clear inputs and predictable steps automate cleanly. Judgement-heavy work doesn't, yet.
3. Is it low-stakes if it's wrong? Start where a mistake is cheap and reversible, not where it's catastrophic.

A task that's frequent, rule-based and low-stakes is your first win. Sending receipts, syncing records between two systems, formatting and filing reports, chasing unpaid invoices.

What to leave alone (for now)

• Anything requiring real human judgement or empathy
• One-off tasks where building the automation costs more than just doing it
• High-stakes decisions with no easy undo

Build, measure, widen

We automate one process, measure the hours it gives back, and only then move to the next. Each win funds the confidence for the one after it.

Automation compounds. The trick is choosing a first domino that knocks the others over.

Done in this order, automation stops being a scary leap and becomes a series of small, obvious wins.