Business process automation: what to automate, which tool to use, and how to measure the return

Automating a business process pays off when three conditions are met: the process runs often, its rules are stable, and an error is expensive. The right tool then depends on the process, not on fashion: a feature already built into your software, a no-code connector (Zapier, Make, n8n), an AI step, or custom development. And the return on investment is calculated with your numbers, measured before you start, never with an average found online.

Most guides on this topic end with a vendor's percentage. This one gives you a scoring grid to choose what to automate, an honest comparison of the families of tools, and an ROI method you can present to your board or your bank.

A task, a workflow, a process: three scales

"Automation" covers very different things:

• a task: send a reminder when a quote has had no reply after two weeks;
• a workflow: a website form creates the contact in the CRM, alerts the salesperson and adds a row to the pipeline report;
• an end-to-end process: an order travels through sales, production, logistics and invoicing, with approvals at each stage.

The larger the scale, the more the choice of tool matters. A task is often set up in an hour. An end-to-end process held together by twenty chained connectors becomes fragile, and that is where custom software enters the discussion.

What to automate: a six-question grid

For each candidate process, answer yes or no:

1. Does it run at least weekly, and for several people?
2. Have its rules been stable for several months?
3. Is the data structured (fields, amounts, dates) rather than free text?
4. Does it involve retyping from one tool into another?
5. Is an error expensive: a dispute, a penalty, a lost customer?
6. Are exceptions rare, roughly fewer than one case in five?

Question 2 is the real deal-breaker. Automating a process that changes every month means paying to reprogram it every month.

Simplify before you automate

Before any tool, map the process step by step with the people who do it, exceptions included. You will almost always find steps that serve no purpose: an approval nobody ever refuses, a copy into a spreadsheet nobody reads, an email sent "for information". Removing them costs nothing. Automating a useless step makes it permanent.

Examples by department

If many of these processes currently live in shared spreadsheets, our guide to replacing Excel with a custom application covers that first step.

The families of tools

1. What your software already does

Before buying anything, look at the rules and automations built into the tools you already pay for: scheduled reminders in your invoicing software, mailbox rules, CRM workflows. They are often free, maintained by the vendor, and underused.

2. No-code connectors

Zapier, Make and n8n link applications that have an API, with rules of the "when this happens, do that" kind. Their public pricing, checked on 7 November 2026:

They are excellent for starting fast and testing an idea. Their limits show up over time:

• the bill follows volume: a workflow that runs ten times more often costs more every month;
• complex rules become unreadable: twenty conditional branches in a visual editor are something nobody reviews;
• failures can go unnoticed if nobody watches the alerts;
• data passes through a third party, which belongs in your GDPR records of processing. n8n can be self-hosted, which changes the picture on that point.

Our comparison of n8n and custom code goes deeper into this choice.

3. Robotic process automation, for tools with no API

RPA tools drive an existing application's screens the way a person would: clicking, copying, pasting. They are a workaround for old software that cannot be integrated any other way. They work, but they break when a screen changes, so treat them as a bridge while you replace or open up the underlying tool, not as a foundation.

4. AI, for what is not structured

Reading an email to work out what the customer wants, extracting the lines of a scanned purchase order, summarising a file before a meeting: these are steps that fixed rules cannot handle and AI handles well. Two precautions. It is sometimes wrong, so its important outputs go through a human check. And documents sent to an AI model are data leaving your systems. Our guide to what a custom AI agent costs covers both points.

5. Custom development

It becomes the right choice when the process sits at the core of your business: numerous rules specific to your company, several departments working on the same file, a need to trace who did what, high volumes. The code then builds on the APIs of your existing software. Our explainer what is an API describes that mechanism and what drives the cost of an integration.

The comparison

Most companies combine these families. The rule: one family per type of step, and a single place where everyone can see what is automated.

Calculating the return, with your numbers

You will find spectacular percentages online. They say nothing about your company. Here is a method you can apply yourself in half a day.

Step 1: measure before

For two weeks, record for the chosen process:

Step 2: calculate the annual gain

• Time saved = time per occurrence × occurrences per month × 12 × fully loaded hourly cost × share actually automated.
• Errors avoided = errors avoided per year × average cost of an error.
• Faster cycle time: count it only if it has a measurable effect (invoices issued sooner, so paid sooner; orders shipped faster).

Be conservative about the "share actually automated": exceptions stay manual, and time freed up only becomes a gain if it is redeployed to something else.

Step 3: compare with the total cost over three years

Add up setup, subscriptions (connectors, AI models, licences), hosting, maintenance, and the internal time spent describing the process, testing and training. Divide by the annual gain: that is your payback period, calculated with your own data. Our guide to what a custom business application costs breaks down the development side.

Step 4: measure after

Take the same measurements three months after go-live. That is what validates the calculation, and what justifies, or not, the next automation.

The traps that sink an automation

• Automating a mess: a vague process, automated, stays vague, only faster.
• Silent failures: a workflow that stops running without telling anyone. Every automation must alert a person when it fails.
• Sprawl without an owner: dozens of workflows created by different people, with no list and no documentation.
• Forgotten personal data: every third-party tool handling customer or employee data belongs in your GDPR records, under a suitable data processing agreement. Our GDPR checklist lists what to check.
• Leaving teams out: an automation imposed without the people concerned gets bypassed at the first exception.
• Removing every human check on decisions that commit the company: credit notes, discounts, refusals.

Where to start

1. List the processes that run every week and score them with the six-question grid.
2. Pick one, the highest scoring, and map it with the people who run it.
3. Measure it for two weeks.
4. Try the lightest solution that fits: built-in feature, then connector, then custom development.
5. Name an owner and set up failure alerts before go-live.

If your process calls for custom software, see the tool before you buy it. We build a free visual prototype, presented on a video call within 5 business days, so your team can see the automated process screen by screen before any commitment. Our article see your software before you pay explains how it works.