A guide for management and IT leads in mid-sized companies. No jargon, with the questions that belong on the table before a decision.
A chatbot writes text that a person reads and then decides what to do with. An AI agent receives a task and access to systems. It can send an email, change a record in the CRM, prepare an invoice or place an order. The difference matters for liability: with a chatbot a person acts in the end, with an agent the software acts. That is why with an agent the question is who allowed it to do what.
An agent pays off where the same work comes up often and the rules are known: pre-sorting requests, reconciling data between two systems, checking and summarising documents, preparing quotes from templates. It gets difficult with decisions that two experienced colleagues already disagree on today. There an agent helps as a preparer, but the decision should stay with a person. You can roughly assess whether your case fits with the tools in our Lab.
Costs come in two parts. The first is development: connecting to your systems, rules, approvals, tests. The second is operation, meaning compute for the language model. With cloud models you pay per use, with a local model once for hardware and then power and maintenance.
Concretely with us: a discovery workshop clarifies in one working day whether an agent pays off for your case, and costs €2,900 net. It is fully credited against a project from €25,000 net. Custom AI applications start at €25,000 net. If you want several agents to run through one shared control and approval layer, AXYON is available as a perpetual licence: €7,990 one-off as a macOS app for one workstation, €39,990 one-off as an appliance for the first production environment. All prices net plus VAT.
The key question does not come with the first error but when someone asks why the agent did something on a certain day. That needs an answer you can back up. These building blocks belong to it, regardless of who builds the agent. More detail using WordPress as an example on the AI plugins for WordPress page.
The agent has a described task and may only use the systems and actions needed for it. Everything else is blocked, even if the tool could technically do it.
Anything that moves money, reaches customers or deletes data goes to a person first. You set the approval level, not the agent.
An agent must be stoppable immediately without anyone touching the code.
Who triggered what, who approved it, with what result. In a cryptographically chained form, later changes are detected.
Agents read emails, documents and web pages. These can contain instructions the agent mistakes for yours. Separate permissions and approval before every critical action help against this.
Task, permissions, owner, approval levels and changes belong in one operations record. It is what you present in an audit.
If a task can be fully described with fixed rules, classic automation is often cheaper and more predictable than a language model. An agent only adds value when text has to be understood, cases classified or unstructured documents read. We clarify that in the discovery workshop too, and the result can be that you don't need an agent.
Pick a time that suits you — simple and direct.
The call takes place via Zoom — you’ll receive the link automatically by email after booking.