AI agent development for work that is never the same twice

An agent is software with a language model that gets a goal and chooses its own steps to reach it, using the tools you give it access to. A fixed workflow does the same steps every time; an agent adjusts to what it finds on the way. Here is what that means inside a business.

Not every row needs a full agent. A weekly report is really a workflow with one AI step: the numbers come from your systems, the model only writes the commentary. Drafting quotes from a site note is its own page: AI quote generator. An agent that prepares your marketing is worked out in AI for marketing, other uses of AI in a business on AI for business, and an agent inside a larger platform is custom software.

Small companies rarely lack tools, they lack connections. The inbox does not know what is in the CRM, and only the owner knows everything. An agent works on those seams: it reads in one system and acts in another.

Take an installation company with twelve people, a shared inbox with forty emails a day, fifteen new enquiries a week and an owner who chases quotes himself. These are assumed figures. Run them again with your own week.

That is 12.25 hours a week. Over 46 working weeks it adds up to 563.5 hours. At an assumed cost of 40 euros an hour, that is 22,540 euros a year.

Against that stands the model cost. Say the agent calls the model four times per email, each time with about 3,000 tokens in (the mail, the file, your instructions) and 400 tokens out. With Claude Sonnet 5.5, at 2 dollars per million input tokens and 10 dollars per million output tokens, that is 4 US cents per email. For 880 emails a month it comes to 35.20 dollars, about 31 euros, or some 375 euros a year. With the faster Claude Haiku 5.5, at 0.10 and 0.50 dollars, it drops to 1.76 dollars a month. With the heaviest model, Claude Opus 5.5 at 4 and 20 dollars, it is 70.40 dollars. Run your own volume through the AI cost calculator, and see which model fits which task in ChatGPT vs Claude.

Build the same flow in a tool and a subscription comes on top. Zapier Professional costs 19.99 dollars a month on annual billing for 750 tasks, counting every step after the trigger. n8n Starter costs 20 euros a month on annual billing for 2,500 executions.

What does that mean for your budget? An agent that frees 563 hours a year and costs a few hundred euros in model fees pays back an investment of about 22,000 euros in a year, and one of 44,000 euros in two. Whether building beats buying for your case is laid out in build vs buy software.

An agent is only as good as its limits. It gets the permissions it needs and no more, it never sends, pays or deletes without approval, and it writes every step to a log. On top of that come two European laws. The EU AI Act (Regulation 2024/1689), amended by the AI Omnibus in force since 27 July 2026, also applies to companies outside the EU when the output of their AI system is used in the EU. GDPR applies as soon as the agent touches personal data.

For most business agents this comes down to three agreements: who approves, which data never goes to the model, and what the agent says about itself in a conversation. The EU AI Act checker tells you in ten questions which risk class your use falls into. Whether you need a written policy is answered in do I need an AI policy. What changed for chatbots in August is in EU AI Act transparency obligations, and the new timeline for high-risk systems in the EU AI Act high-risk delay.

• Too much at onceAn agent that has to handle mail, quotes and planning does all three badly. Start with the task that costs the most hours and carries the least risk.
• Permissions that are too wideAn agent allowed to delete your whole inbox will one day do so. Reading and drafting is almost always enough.
• The model as the source of truthA price, a date or a stock level comes from your system, never from the model's memory.

We are based in Ghent and work in English, by video call and shared screens. Belgium runs on Central European Time, so a full working day overlaps with the Netherlands, Germany and the Nordics and the morning overlaps with the UK. Prices are in euro. We host in an AWS region in Europe, so your data stays under EU rules, and you own the code and the data.

For AMOSI, our own marketing software that went live in July 2026, we built blocks in which a model does work on its own: a website scan that collects the facts about a company, reading what a company uploads, an overview of competitors and the sector. That is where we learned where a model may choose and where a person has to sign.

For DA Logistics, a Belgian haulier, we built a platform the office and about thirty drivers use every day, with five queues for the office and a mail layer that sends its own reminders: 185 emails in thirty days. That is automation without AI, and it shows where an agent lands most easily: in software where the tasks and the data are already in place.

What we do for you:

• A call about your real work: your inbox, your CRM and your week. Afterwards you know which task comes first and what it is worth.
• Limits on paper: what the agent may do, what it asks, and what never.
• A first version in a few weeks, measured against your own examples.
• Maintenance: models and prices change, and we tell you when a cheaper model does the same job.

The full offer is on AI automation. Book a call and we will look at whether a workflow is enough or an agent makes sense.