Process automation and AI workflows

Process automation, with or without AI

With process automation, software does a fixed piece of work for you. A workflow follows set rules: an order comes in, a confirmation goes out and the overview is updated. An AI step is something else: it has to understand what it's reading, such as an invoice in an unusual format or an email where the real question is between the lines. Many processes only need the first kind.

A process is a good candidate if it recurs every week, if a mistake costs money or if only one colleague knows how it works.

Where AI fits, and where it doesn't

AI automation for small and medium-sized businesses is not a silver bullet. An LLM is strong with messy input, such as a PDF that looks different every time or a product description that still needs writing. If a few clear rules do the job, a standard workflow is cheaper, more predictable and easier to check. In that case we leave AI out, and we'll tell you before we start.

A model can get things wrong. So a person checks wherever a mistake has a cost: the system prepares everything, someone on your team signs off. And we don't automate everything at once.

What you can automate

If you want to automate business processes with AI, look for work that lands on someone's desk every week and has to be read before anyone can act on it:

• reading a supplier invoice and entering the lines into your accounting software or webshop;
• sorting incoming email to the right team or folder;
• copying data from an email or form into a spreadsheet or system.

Not everything needs AI: a weekly report that compiles itself is an ordinary workflow.

How we build it

We build workflows in n8n, where every step is visible. If a step needs AI, we choose the model case by case: Claude, GPT or Gemini in the cloud, or a local model via Ollama. A local model runs on hardware you manage; with sensitive data, that can make all the difference. Every step the workflow takes is logged.

In the introductory call we walk through your process: where the work comes in, who touches it and where it can go wrong. Then we work out which steps can be handled with fixed rules and where AI is needed. The first building phase automates one process, and your team tests it with real work: that week's invoices or emails.

If your question is mainly about two systems that need to exchange data, our integrations and APIs service is a better fit.

Example: from invoice to product page

At De Whisky Specialist, every new delivery was manual work. Someone looked up each unfamiliar bottle, entered the details in Shopify and picked the right value for each attribute. If a distillery was spelled slightly differently, the bottle didn't show up in the right filter or collection.

The system we built in n8n shows where AI does the work and where a person decides. The team drops the supplier invoice into Slack. The system reads it and splits the lines into bottles the shop already stocks and new ones. For a new bottle, the system searches online. First, a separate agent checks whether the sources it found are really about this bottle; only then does any writing happen. That's also how the system knows when it doesn't know enough. Five agents then each fill in some of the attributes: country, region, distillery, bottler and cask type, always using the shop's existing values. That way every new bottle fits the filters and collections that already exist.

Nothing goes online without approval. Each proposed product appears in Slack, where the team accepts it, rejects it or edits it first, down to the last field. All the team has to do is supply the invoice and make the final click.

What does automation cost?

It depends on the number of steps and systems, how messy the input is, whether AI is needed and how many phases you want. A cloud model is paid per use; a local model needs its own hardware. Talking through your process and working out which approach fits is free.