Start automating with the process that is repetitive, rule-based, frequent and costly when it goes wrong, for example copying order details from emails into a spreadsheet and then into accounting. Leave decisions that need judgement with people, and let automation prepare those decisions instead of making them.
How do you find the processes worth automating?
Walk through a normal week and list every task where someone copies, re-types, chases or checks something. Then score each one:
| Question | Score 1–3 |
|---|---|
| How often does it happen? | Daily scores highest |
| How long does it take each time? | Longer scores higher |
| How rule-based is it? | Same steps every time scores highest |
| How costly are mistakes? | Lost orders, wrong invoices score highest |
| How many systems does it touch? | More systems, more copying |
Start with the highest total. It is usually less glamorous than expected, and saves the most.
Common first automations
- Order intake: an order email or form becomes a record in the system automatically.
- Approvals: requests route to the right person, with reminders and a record of who approved.
- Invoicing: delivered orders become invoices in the accounting system.
- Reminders: payment, renewal and maintenance reminders go out on schedule. Factories use this for compliance and maintenance tracking.
- Reports: the weekly report builds itself from live data.
Where does AI genuinely help?
AI is useful where the input is messy: emails, PDFs, scanned forms, free-text requests. It can read a customer's email, understand that they want to change a delivery date and add units, find the order and draft the change. A person approves anything above a set limit. Used this way, AI removes the reading and re-typing, and people keep control of the decisions.
What should you leave alone?
- Processes that change every time.
- Rare tasks that take minutes a month.
- Decisions with real consequences and unclear rules. Automate the preparation, not the decision.
- Broken processes. Automating a bad process only produces bad results faster; fix the steps first.
How do you make automation reliable?
Log every run, alert someone when a step fails, keep a manual fallback, and review the results regularly at first. Automation that fails silently is worse than none.
What does it need underneath?
Usually connected systems: automation moves data between tools, so those tools must be reachable. Sometimes the right first step is a custom system that replaces several spreadsheets.
Find what to automate first in a short discovery call.