How to find a workflow worth automating

Businesses have plenty of tasks they do every day, but not all deserve automation. Some repeat but take only minutes. Others take hours but differ every time. Some have clear rules, yet a mistake could cause irreversible damage.

So do not start by asking “What can AI do?” Ask:

Which part of the work repeatedly consumes time and already has a stable enough method for a system to assist?

Map one real process first

Choose work that staff often complain about, that builds up in a backlog, or that requires repeated entry. Observe how it actually happens. Do not just copy the standard process document; record the extra actions people take to finish the job.

Daily email orders, for example, might involve:

Receive email
Download attachments
Check customer, products and quantities
Ask for missing information
Enter valid data into the order sheet
Put attachments in the designated folder
Notify the reviewer
Confirm receipt to the customer

“Handling email” sounds like one task, but only four steps may be suitable for automation. Breaking it down reveals a scope with the right balance of value and risk.

Six signals for assessing candidates

The more of these characteristics a process has, the more suitable it is for assessment:

1. Frequent. It happens daily or weekly, not once a year.
2. Identifiable inputs. Required data, files or triggers can be listed.
3. Reasonably stable rules. Most cases follow the same decisions.
4. Checkable results. A person can verify whether the system acted correctly.
5. Exceptions can return to a person. Uncertain cases have a named owner.
6. Delays or errors have real costs. The benefit is not merely looking advanced.

This is not a score that obliges you to automate. It helps filter out ideas with unclear boundaries.

Look for waiting and copying before complex decisions

The most useful starting point is often reducing data movement and unnecessary waiting, rather than replacing difficult judgement. For example:

• Write form submissions into customer records automatically.
• Check required fields and send standard reminders for missing information.
• Name and file attachments using agreed rules.
• Remind the owner when a status exceeds its deadline.
• Generate a fixed-format summary from the same data.
• Synchronise confirmed results with the next tool.

These actions are not clever, but they interrupt people every day. Removing those interruptions can be more useful than automatically generating text.

Document one normal case and three exceptions

Before discussing a solution, prepare:

• One normal case from start to finish.
• The three most common exceptions.
• Who currently decides each exception.
• How mistakes are detected and corrected.
• Which data must not leave the existing system.

If the team cannot describe even the normal case consistently, the first step is usually clarification, not automation.

Make the first release promise one observable improvement

Good first-release goals are specific:

• Reduce time spent re-entering order data.
• Improve completeness of required information in new enquiries.
• Make every request left unhandled for more than a day visible.
• Reduce misplaced attachments or inconsistent names.
• Focus human review on exceptional cases.

“Improve efficiency across the business” cannot be meaningfully accepted. “Remove copying each order from email into a spreadsheet while retaining human confirmation” is much clearer.

A simple candidate list

List five to ten repetitive tasks and record:

Item	What to record
Frequency	Daily, weekly or occasional
Time per case	Including waiting, searching and repeated confirmation
Normal-case share	How many cases follow the same path
Error impact	Minor rework, customer impact or major risk
Checkability	Who can confirm the result is correct
Exception owner	Who takes over when the system is uncertain

Prioritise frequent, time-consuming work with a stable normal path and recoverable errors. It may not be the most eye-catching problem, but it is often the best place to start.

---

This article provides a general method for selecting workflows. It does not assume automation or AI is always the best answer.