How to tell if a task is worth automating

Most automation gets sold on the wrong number. Someone says a task takes four hours a week, multiplies it by an hourly rate, and presents the annual saving. The maths is fine. The conclusion is often wrong.

Time spent is the easiest thing to measure and one of the weakest reasons to automate.

The question that actually predicts success

Ask this instead: **how often does this task get done the same way twice?**

A task that runs identically every time is a candidate. A task where a person makes a judgement call each round is not — or at least, not yet. Judgement is where automation quietly breaks. The agent handles the ninety cases you described and does something confidently wrong on the ten you did not.

So before we quote anything, we ask a client to describe the task twice: once as the rule, once as the exceptions. If the exception list is longer than the rule, we say so, and we usually say don't.

Four hours a week is not the bar

Some four-hour tasks are worth automating. Some twenty-minute ones are worth far more.

What matters is not duration but the shape of the time:

**Interrupting work beats bulk work.** A task that takes eight minutes but arrives eleven times a day is worse than a half-day job done every Friday. The half-day is scheduled. The eight minutes destroy a morning, because each one costs the person their place in something else. Automating the small interrupting task usually returns more real capacity than the number suggests.

**Delayed work costs more than slow work.** If applications sit for two days before anyone screens them, the cost is not the screening time. It is the good candidate who accepted another offer on day two. Automating the delay is worth more than automating the labour.

**Error-prone work compounds.** A task done wrong occasionally, discovered late, is expensive in a way no hourly rate captures. Not because of the mistake, but because of the reconciliation afterwards.

None of those show up in hours-times-rate.

When we tell people not to automate

Three cases come up often enough to name.

**The task is unclear, not repetitive.** If nobody can write down how the work is done, an agent cannot do it either. What is needed first is a documented process. Sometimes writing that down removes the problem entirely, and there is nothing left to build. That answer is free, and we give it.

**The volume is too low to matter.** If something happens four times a month, a well-made checklist beats a system that needs monitoring, updating, and paying for. Automation has a running cost, not just a build cost. Below a certain volume, that cost is the larger number.

**The task is about to change.** Automating a process that is being restructured next quarter means building the wrong thing carefully. Wait for the process to settle.

What good candidates look like

The work we take on tends to share a shape:

• It runs on a trigger, not a decision — a form arrives, a file lands, a date passes.
• The rules can be written down by one person in one sitting.
• Failure is visible. If the agent gets it wrong, somebody notices soon, not in a quarterly review.
• Volume is high enough that a bad week is felt.

Applicant screening fits. Invoice matching fits. Report assembly fits. "Handle our inbox" almost never does, because an inbox is fifty different tasks wearing the same coat.

The practical test

Before your next automation decision, try this. Take the task and write down every step, then mark each one **rule** or **judgement**.

If the rules cluster at the start and the judgement sits at the end, you have a strong candidate: automate the first eighty percent and hand a human the decision, already prepared. That pattern — the agent does the gathering, the person does the deciding — outperforms full automation more often than most vendors will admit.

If judgement is scattered throughout, leave it alone for now. You would be buying a system that needs supervision at every step, which is a job, not a saving.

Automation is worth it when it removes interruption, delay, or error. If it is only removing hours, check the number again.