Many of the requests we receive start with a similar sentence: "we'd like to automate the way we work, with AI too". It's the right goal, but put that way it's too big to become a project. A company isn't automated all at once: it's automated one process at a time, starting from the one that costs the most.
This guide collects the method we use in first calls with owners and managers of SMEs. It helps you work out where to start before talking about tools.
Start from the process, not the technology
The most common mistake is choosing the tool first: a ChatGPT subscription for everyone, a chatbot on the website, a new CRM. It rarely changes anything, because the tool isn't connected to the point where the company loses time.
The starting point is always a concrete question: which repetitive activity takes up the most hours each week, and who does it? The answer usually comes within minutes. It's the office retyping orders, the salesperson updating a spreadsheet in the evening, the accounts team transcribing suppliers' delivery notes.
How to recognise a process worth automating
We use three criteria. If a process meets all three, it can almost always be automated, at least in part.
- Frequency. It repeats every day or every week. An activity that happens twice a year isn't worth the investment.
- Regularity. It follows rules a person could explain to a new colleague, even with a few exceptions.
- Written information. The data arrives in writing: messages, emails, PDFs, spreadsheets, fields in a management system. Even messy ones.
Typical examples in SMEs:
- customer orders received on WhatsApp or by email and entered by hand into the management system;
- suppliers' delivery notes and invoices transcribed line by line;
- quote requests arriving from the website that need to be sorted and recorded;
- customers and deals tracked in spreadsheets, with follow-ups that get lost;
- weekly reports put together by copying data from several programs.
Where artificial intelligence is really needed
AI isn't always needed. It's needed when information arrives written in non-uniform ways and today a person has to read and interpret it.
A real example: a B2B food distributor we work with receives 80–100 orders a day, 70–80% of them on WhatsApp. Every customer writes in their own way: "the usual on Tuesday", nicknames for products, quantities in cases or in kilos. Here a language model does what the operator used to do: it reads the message, compares it with the customer's history, the price list and stock levels, and writes the order into the management system. Ambiguous orders stay highlighted for a person. We told the story here.
When the data is already structured, for example a website form that needs to end up in the CRM, AI adds nothing: a classic integration between two pieces of software is enough, cheaper and more predictable.
Automation, custom software or an off-the-shelf program
Once the process is chosen, there are three routes. The choice depends on how specific your way of working is.
- An off-the-shelf program makes sense when the process is the same as at thousands of other companies: invoicing, payroll, email marketing. It costs little and works straight away.
- An automation makes sense when the programs already exist but don't talk to each other, and someone acts as the bridge by copying data. The existing tools are connected, often with AI in between. Here's how we work on automation.
- Custom software makes sense when the tool itself is missing: your process doesn't fit any program on the market and today lives in spreadsheets. This is the case for many CRMs: for a window and door company, for example, we built a CRM around the way they follow customers and deals instead of adapting a generic one. Learn more about custom software.
Often the answer is a combination: you keep the management system and build only the missing piece.
What it costs to automate a process
To give an order of magnitude, these are our starting prices:
| Work | From | What it includes |
|---|---|---|
| Automation of one workflow | €1,000 | One process, a few steps, software that exposes APIs |
| Custom software, first version | €1,000 | One process, a few roles, data imported from spreadsheets or the management system |
| Automation with AI reading | on proposal | Price tied to message or document volumes and integrations |
On top of these come running costs: for AI usually a few cents per message or document, for hosting and database a few euros a month, in the company's name.
The right way to assess the spend is to compare it with today's cost. If a person spends an hour a day retyping orders, that's about 220 hours a year: the return is quick to work out.
How a project runs
- 30-minute call. You tell us about the process, the volumes and the programs you use.
- Written proposal. Scope, price and timeline. What's left out is stated just as clearly.
- First version in production in 2–4 weeks, used on real work.
- Improvements based on use, not on a list drawn up at a desk.
Mistakes to avoid
- Starting with everything. A project that wants to automate the whole company never reaches production. One process at a time.
- Removing human control. Where a mistake costs money, the system prepares the work and a person approves it. Exceptions must be visible.
- Asking customers to change their habits. If they order on WhatsApp, let them. The automation works on what happens next.
- Buying a tool before understanding the problem. It's the fastest way to pay for a subscription nobody uses.
- Not measuring. Before starting, estimate the time saved. If the return isn't clear, it isn't worth going ahead.
Frequently asked questions
Which process should we start with?
The one that takes up the most hours each week, repeats often and receives information in writing. In SMEs it's almost always orders, delivery notes, email requests or managing customers in spreadsheets.
Does our data need to be organised already?
No. What's already in the management system, emails and documents is often enough. Understanding which data exists and in what form is part of the first analysis.
Can artificial intelligence make mistakes?
Yes, like a person. That's why in delicate steps the system flags uncertain cases and a person confirms before the data enters the management system.
How long before we see a result?
A first automation on a single process goes into production 2–4 weeks after confirmation. The savings can be measured from the first weeks of use.
Is the company's data used to train the AI?
No, if the models are used through APIs with terms that exclude training. The accounts stay in the company's name.