How to Tell Which Parts of Your Job Can Be Automated

Map work at the right level

Write each task as an input, action, output and recipient. 'Prepare the weekly variance commentary from approved ledger data for the finance lead' can be tested. 'Do finance' cannot.

Include hidden work such as chasing information, resolving exceptions, explaining results and taking responsibility. Demonstrations often omit these costly edges.

Screen for fit before choosing a tool

Prioritise high-volume, stable work with clear acceptance criteria. Pause when source data is fragmented, policy changes frequently, permissions are uncertain or quality relies on tacit context.

• Frequency and time: enough repetition exists to repay design, control and maintenance.
• Input quality: the tool can lawfully access complete, current and well-structured information.
• Rule stability: common cases follow a documented process and exceptions can be routed.
• Verification: a reviewer can detect an error without repeating the whole task.
• Consequence: failures can be reversed before they affect a customer, right, safety outcome or material decision.

Design a controlled pilot

Run the current and proposed methods on a representative sample. Predefine acceptable quality, the person who can stop the pilot and the cases that must escalate. Keep an audit trail of inputs, versions, outputs and corrections.

Measure elapsed time, human attention, error severity, rework and downstream satisfaction. An output that appears quickly but requires anxious checking may not improve the work.

Decide to automate, assist or keep human

Full automation is appropriate only when the process and controls can handle normal variation. Assistance is usually the better first state for drafting, classification and analysis because a named person retains the decision.

Keep work human when legitimacy, empathy, negotiation, professional duty or responsibility is central. Revisit the choice as tools and the process change.