What tasks should I automate first?

It happens often

Daily or weekly work adds up. A task done a few times a year rarely repays the effort of automating it, however tedious it feels.

It is rule-bound or predictable

The same kind of input leads to the same kind of action. If staff can explain the steps in a few sentences, a system can usually follow them. AI helps when the input is free text, but the decision about what to do with it should still be clear.

A mistake is cheap to catch

If the system mislabels an enquiry or sends a reminder at the wrong time, someone notices and fixes it without lasting harm. That makes it a safe place to learn.

Inputs and outputs are clear

You can point to where the work arrives (a form, an inbox, a spreadsheet) and where the result should end up (the CRM, a calendar, a shared folder).

The time it takes can be measured

You can roughly say how long it takes and how often it happens today. Without that, you cannot tell afterwards whether the automation helped.

• –Judgement-heavy work, such as pricing a custom job, deciding on a refund or giving professional advice. AI can prepare information for these decisions, but a person should make them.
• –Sensitive conversations, such as complaints, health or financial matters and anything where tone carries real consequences.
• –Rare tasks. The build and testing effort is similar whether a task runs every day or twice a year.
• –Processes built on messy data: inconsistent spreadsheets, duplicate contacts, information spread across personal inboxes. Clean up or consolidate first, or the automation will spread the errors further.
• –Processes that are still changing. If the team is still working out how something should be done, automating it fixes the current version in place.

1. 01List every task that someone on the team repeats and would happily hand off. Ask the people who do the work, not only the managers.
2. 02Score each task from one to three on five questions: how often it happens, how predictable it is, how low the risk is if it goes wrong, how clear the inputs and outputs are, and how easily you can measure the time it takes.
3. 03Cross out anything that scores lowest on risk, whatever its other scores. A frequent task with serious consequences for errors is not a first project.
4. 04Rank the rest by total score, then check the highest-ranked few against one practical question: can the systems involved be connected?
5. 05Pick one. Starting with several at once makes it hard to see which one is working.

This is not a precise method, and it does not need to be. Its job is to stop the first project being chosen because it sounded impressive rather than because it was a good fit.

Take a small accounting practice with three candidate tasks. Chasing clients for missing documents before a deadline happens constantly, follows clear rules and carries little risk, so it scores high. Sorting incoming client emails into categories is frequent and useful but less predictable, so it scores in the middle. Drafting advice on a client's tax situation is judgement-heavy and high risk, so it is crossed out.

The practice starts with document reminders: a system that checks which documents are outstanding, sends a polite reminder on a schedule and flags anyone who has not responded for a staff member to call. Email sorting becomes the second project. Tax advice stays with the accountants.

Many good first tasks do not need AI at all. If the steps are identical every time, a rule-based automation is simpler, cheaper to run and easier to test. AI earns its place where the task involves reading or sorting information that arrives in varied forms. The difference between AI tools, automations and AI agents explains how to tell which you need.

Once you have a shortlist, the next question is whether the time saved justifies the build and running cost; is AI automation worth it walks through that estimate. For the kinds of workflows that usually make good first projects, see business automation.