Article

Automating a business process with AI: a step-by-step guide

A practical step-by-step guide for automating a business process with AI: from documentation to pilot, including the mistakes you want to avoid.

Written by Loek Delahaye, founder, Delahaye Solutions · ex-CTO with 10+ years in software and AIPublished:
Short answer
  • Automation most often fails not because of a bad tool, but because of a poorly documented process. Start with the process, not the tool.
  • Five steps: document the process, identify the repetitive steps, choose the level of automation, run a two-week pilot, and monitor the results.
  • The three most common mistakes: automating a broken process, choosing an AI model when a simple rule would work better, and skipping the pilot.
  • AI automation for a specific business process at Delahaye Solutions starts from €1,500 for a single automation.

Why do so many AI automation projects fail?

Most failures in automating business processes are not caused by the AI. They are caused by the process. A tool that sends calendar invitations or processes invoices only works reliably once you know exactly what the input is, what the output should be, and what the exceptions are. Businesses skip that step, build an automation based on how they think the process works, and discover three weeks later that the exceptions account for half the cases. That is not an AI problem; it is a documentation problem. The good news: it is solvable, and you do not have to make a large project out of it.

Step by step

Five steps for automating a business process with AI

This step-by-step plan works for any SME business process: from email handling and invoice processing to lead qualification and report generation.

1. Document the process as it currently works

Write down exactly what happens, step by step, including the exceptions. Use two concrete examples from the past week. This feels time-consuming, but it is the step that determines everything. Without proper documentation you automate an assumption, not the real process. A simple list in a text document is enough: input, what happens to it, output, and who or what makes a decision along the way.

2. Identify the repetitive steps

Which steps in your documentation are exactly the same every time, or nearly the same? Mark those separately. Those are your automation candidates. Steps that are different every time, or that require someone to make a judgment call based on context, are less suitable. Start with the steps that cost the most time and require the least judgment.

3. Choose the right level of automation

Not every repetitive process needs AI. There are three levels: a simple rule ('if X then Y'), a trained model that recognises patterns, and an autonomous AI agent that carries out multiple steps independently. Start with the lowest level that solves the problem. A rule is cheaper, faster, more reliable and easier to understand than a model. AI adds value when the input is too varied for a fixed rule.

4. Run a two-week pilot

Build the automation at limited scale: for part of the volume, or for part of your clients. Run the automation for two weeks in parallel with the manual process. Compare the output, count the deviations, and determine whether the difference is acceptable. Two weeks is enough to see exceptions you missed during the documentation phase, without moving too much volume too early.

5. Monitor and improve

After the pilot, track how the automation performs: output volume, error rate, and time saved per week. Build a simple dashboard or export the counts to a spreadsheet weekly. Set a threshold: if the error rate rises or the volume drops, you want a signal before it becomes a problem. Automation is not 'build it and forget it', but in the maintenance phase it costs far less time than the manual work did.

Mistakes to avoid

The three most common mistakes in AI automation

Mistake 1: automating a broken process. If a process already has problems when done manually, automation does not fix them. It makes them faster. Fix the manual process first. Mistake 2: immediately choosing an AI model. Teams reach for ChatGPT or a model when a simple rule would have been enough. A rule that says 'if subject starts with ORDER then assign to sales team' is faster to build, cheaper and easier to debug than a language model. Use AI only when the input is too varied for a fixed rule. Mistake 3: skipping the pilot. 'We know how the process works, so we roll out immediately.' Three weeks later it turns out that 30% of the cases are exceptions that only surface once you run real volume. A two-week pilot is cheaper than two weeks of emergency repairs after a full rollout.

Frequently asked questions

Frequently asked questions about automating business processes with AI

Which business processes are best suited for automation?
Processes with high repetition, low variation, and a clear input and output work best. Practical examples: invoice processing, email classification, lead qualification based on form data, report generation from fixed data sources, and appointment scheduling. Processes that depend on human judgment based on context, or where the input varies significantly each time, are less suitable as a starting point.
How long does it take to automate a business process?
A single automation, such as processing incoming emails or invoices, typically takes two to four weeks from intake to pilot. The documentation phase takes one to two days, the build one to two weeks, and the pilot two weeks. More complex processes with multiple systems or approval flows take longer. We always provide a fixed price and timeline upfront.
What does automating a business process cost?
A single AI automation at Delahaye Solutions starts from €1,500. What determines the price: the complexity of the input, the number of systems that need to be connected, and how many exceptions there are. We estimate the scope in a free intake and give a fixed price before we start. No open-ended billing.
Do I need technical knowledge to have an automation built?
No. What you do need is a thorough understanding of your own process: who does what, when, with what input, and what the result should be. We translate that into a technical solution. The intake largely consists of walking through the process together so we know all the exceptions and edge cases before we start building.
What is the difference between an AI automation and a regular automation?
A regular automation follows fixed rules: if X then Y. That works well for processes with limited input variation. AI adds value when the input is too varied for fixed rules: recognising intent in an email, classifying a document based on content, or summarising customer feedback. In practice you combine the two: fixed rules for the structure, AI for the variable parts.
Read more

Want to know what can be automated in your process?

Book a free intake. We look together at your process, what can be automated and what it costs — free and without commitment.

Book a free intake →

Free and without obligation · Reply within one business day · Fixed price up front, no surprises