Which processes are worth automating, how to find them, how to calculate the return and where the limits are. A guide from first thought to a working rule - with no technical prerequisites.
In practice the word automation means something far smaller and more useful than it sounds. It isn't about robots or replacing people. It's about no longer doing by hand something nobody wants to do by hand - chasing a deadline, creating a record, sending a handover report, checking that a finished job has an invoice.
This guide runs from the question “what should we automate” to a working rule, linking to detailed pieces at each step.
1. How to find what to automate
Don't start with technology, start with a list. Three approaches work:
Watch what repeats. For a week, note down activities you do more than twice in the same shape. Copying details out of an email, checking unpaid invoices, contacting clients before a deadline.
Ask what annoys people. The best candidate for automation is work nobody wants to do and which therefore sometimes doesn't get done. Forgotten invoicing is the classic.
Look for where things get lost. When something falls through, it usually means a person was watching for it. Automation watches more reliably.
What doesn't work: starting with a tool's feature list and looking for something that would fit. That's how you end up with automations nobody needs.
2. What an automation is made of
Practically every automation, in any tool, has three parts:
Trigger - when it should happen. A record changing status, a new record, an incoming email, a submitted form, or simply time (every Friday at 15:00).
Condition - when it applies. The job is finished and has no invoice. Payment is due in three days and hasn't arrived.
Action - what happens. Send an email, change a field, create a record, generate a document, notify someone.
Write a candidate from step one into those three sentences and you'll see immediately whether it's feasible - and often discover you're missing the field the condition would need.
In detail: Automation: the complete introductory guide.
3. Five automations almost everyone benefits from
Regardless of industry, these pay back fastest:
Watching for uninvoiced jobs. A job has been “done” for more than three days and has no invoice attached → alert. One forgotten invoice a year pays for this many times over.
A deadline reminder in advance. Not on the day, but several days before. The difference between “this was due today” and “this is due in three days” is the entire point.
Escalation on no response. Alert → wait → second alert, this time to a manager. A pattern that works for tickets, enquiries and approvals alike.
An end-of-week completeness check. Who hasn't logged a timesheet, which job has no documentation, which client is missing a contract.
Creating a record from an incoming message. An enquiry from email or a form creates itself and waits for a person to confirm.
In detail: Automatically issuing invoices, Deadline tracking and AI reads the email and creates the job.
4. Inside the system, or via a connector?
The fork in the road that decides both cost and reliability.
An external connector (Zapier, Make) moves data between tools. It's billed per operation, it can break halfway through, and it's a fourth thing to maintain alongside the three systems you wanted to connect.
Built-in automation works with the data in place. Nothing moves, because there's nowhere to move it from or to.
Same outcome, two different routes
Four hops between tools are four places it can break. Inside the system nothing moves, because the data is already there.
A practical rule: a connector for what genuinely goes outside (online shop → accounting). Everything concerning your own data belongs inside.
A test that settles it in ten seconds: if those two tools were one tool, would this scenario exist at all?
In detail: Are Zapier and Make dead?.
5. Where AI fits in
AI adds one capability classic automation doesn't have: it understands unstructured text. In practice that means the trigger can be an ordinary email, not just a submitted form.
What it reliably handles today: extracting details from an incoming message, sorting requests into categories, summarising a long thread, drafting text from data already in the system.
What to leave to a person: anything that leaves the company or moves money, without confirmation. The rule that addresses most of the worry is AI proposes, a human confirms.
In detail: 5 things AI can do in your company.
6. How to calculate the return
A simple calculation is enough:
(occurrences per month × minutes each × hourly rate / 60) + the value of errors it removes
An example: checking for uninvoiced jobs happens weekly, takes 20 minutes, at a €20/hour rate. That's about €28 a month. On its own, not much. But one forgotten invoice of €2,700 a year makes the second half of that equation dominate completely.
So: for automations that catch errors, don't count the time saved. Count the cost of the error they catch.
7. Automation that isn't a black box
The most common reason a team doesn't trust an automation is that they can't see inside it. When something doesn't happen, nobody knows why.
Three things fix that:
Run history. For every record it should be visible what happened to it, when and why. Not in a separate tool, but on that record.
Comprehensible naming. “Automation 3” tells nobody anything. “Watch for jobs without an invoice” does.
Tested edge cases. What happens when a required value is missing. What happens with a record that's already been processed. When a flow shows a sensible result in those situations too, the team starts trusting it.
In detail: Automation without the black box.
8. Where the limits are
Automation doesn't make sense when:
- The process keeps changing. Automating something that will be different in a month is wasted work. Stabilise the process first.
- There are more exceptions than rules. When every other case needs manual intervention, automation just adds a step.
- The decision requires judgement. Approving a discount, assessing a complaint, granting an exception. AI can prepare the material here, but a person decides.
- You have no data for the condition to test. The most common blocker. The fix isn't better automation, it's adding the missing field.
Common questions about process automation
Do we need a programmer for automations?
No. The trigger → condition → action pattern is assembled visually and can be done by someone who understands the process. A programmer is only needed for connecting to outside services with unusual interfaces.
How many automations should we have?
Fewer than you'd expect. Most companies manage with five to fifteen that run reliably. Fifty automations nobody understands is a worse state than five good ones.
What if an automation makes a mistake?
That's why the last step is left to a person for anything that leaves the company or moves money. And it's why run history matters - an error has to be traceable and understandable, not just fixable.
When should we start automating?
Once the system has run for a few weeks. Only then do you know which steps genuinely repeat. Automations set up on day one usually address the process you think you have, not the one you have.
Will automation replace a role?
In small and mid-sized companies, generally not. What actually happens is that admin shrinks - retyping, chasing, hunting - and time appears for work there was never room for.