Which tasks are worth automating?

Start where work repeats itself

An order arrives by email. Someone copies it into a spreadsheet, creates a ticket and notifies a colleague. Each step is small. Together, they take attention away from the work the customer pays for.

This is an illustrative example. Look for information people repeatedly move, check or enter. First ask whether a step can be removed or simplified. Automation becomes useful when the workflow improves.

Choose a task using five questions

• How often does it happen? Count actual cases during a typical week.
• How much active work does it take? Separate handling time from waiting for a reply.
• How similar are the cases? Clear rules make routine cases easier to automate.
• What happens if it goes wrong? An internal draft is easier to correct than an incorrect payment.
• Who owns the workflow? Someone needs to handle exceptions and monitor quality.

Moving complete form submissions into a ticketing system can be a sensible first candidate. Automatically approving contracts with ambiguous terms is a much harder starting point. Consider the consequence of errors alongside potential time savings.

Calculate the time you can actually recover

Number of cases × minutes saved per case ÷ 60 = hours freed up.

Suppose you handle 120 cases a month and reduce processing from eight minutes to three. That gives 120 × 5 ÷ 60 = 10 hours. If checking and exception handling take two hours, eight hours remain.

This is a worked example, not a forecast or measured client result. Time recovered is not automatically cash saved. Decide how you will use it: faster responses, more deliveries or less overtime. Compare that value with implementation, licences, hosting and maintenance.

Measure quality too. Fewer missed cases may matter more than faster data entry.

Do you need AI?

A conventional integration may be enough when the task follows fixed rules. A complete form can create a ticket and send a confirmation without a language model.

AI may be useful when variable text needs interpretation, classification or summarisation. Test where it misunderstands the input and decide what a person must review. Start with suggestions someone approves before they are used.

Use fixed rules for fixed conditions. Consider AI where interpretation is the task and checking the result is practical.

Test one complete, small workflow

Give the pilot a clear start and finish, such as a contact form becoming an assigned ticket. Specify required information and what happens when it is missing.

Test normal cases, duplicate submissions, empty fields and an unavailable system. Decide who receives failure alerts and how work continues manually. The solution must make errors visible.

Compare active handling time, corrections and completed cases before and after. Choose a test period that captures normal variation. Expand only when the result holds beyond the easiest example.

When should you wait?

Pause if the process changes every week, source information is unreliable or nobody can describe a correct result. Clarify the workflow first. Otherwise, automation may repeat the same misunderstanding at greater scale.

Bring a few anonymised examples, an estimate of volume and a description of where information starts and finishes to your first conversation. That is enough to discuss a bounded first delivery.

Explore AI and automation or use our assessment preparation checklist.

Worksheet: calculate your own workflow

Use a typical month. Prefer observed handling time and label estimates.

Do not count review twice. Decide how recovered time will be used before assigning financial value. Compare a conservative scenario with your main estimate, such as lower volume and more checking.