An assignment
The goal, what a good result looks like and what it must explicitly avoid assuming.
AI agents in a business process
About this source
Gen8
Source: AI agents in a business process
Managed AI service for SMEs. We set up, connect and manage; you stay in control.
© 2026 Gen8 VOF · Dutch Chamber of Commerce 92978789 · Delft
Platform
An agent is not a conversation. It carries out work within boundaries.
An AI agent receives an assignment, a set of tools, access to knowledge and a boundary within which it may act. It makes a plan, carries out steps, checks its own work and submits the proposal to a person.
- Human approval
- A log of every step
- Processing in the EU
Handle an incoming customer request
Plan01Request classified and relevant agreements found
Action02Context retrieved from the customer file, within the handler's permissions
Action03Draft prepared with a reference to the source used
Reflect04Deviation from standard terms identified: referred back as a question
Draft ready · awaiting human approval
Document to the right file
Plan01Document type and party involved identified
Action02Data extracted and compared with the existing file
Reflect03Two possible files found: choice submitted for review
Proposal ready · 1 question open
Components
What an agent consists of
An agent is not a standalone model. It is a combination you configure for each process, and that combination determines what it can do and what it will never do.
Change the tools and you change what it may do. Change the knowledge and you change what it relies on. That makes it possible to evaluate.
Tools
The actions it may perform in your systems: read, search, create, update. If an action is not included, it cannot perform it.
Knowledge
The documents, agreements and examples it relies on, with the corresponding permissions.
Control
For each action: read only, propose with approval, or execute automatically with a log.
| Chatbot | Fixed automation | AI agent | |
|---|---|---|---|
| Works with | Questions from a user | A fixed rule for each situation | A signal plus context |
| When there is an exception | Gives a general answer | Fails or does nothing | Flags it and submits the choice for review |
| Delivers | Text in a window | A completed action | A proposal in the system where the work happens |
| Best use | Answering simple questions | Stable, well-described work | Work involving a lot of context and handoffs |
An agent allowed to do everything is not an agent but a risk. Its boundaries are its design.
That is why we choose the control level for each action, rather than for an entire project. Reading can usually happen automatically; writing to a system in a way that affects a customer requires a person.
The Gen8 emblem reflects how an agent works: one assignment at the centre, reach that grows around it, and an edge where it stops. The farther from the centre, the lighter the line — and that is exactly where an agent should ask a person instead of continuing.
Boundaries
What an agent is not
It does not replace judgment. An agent prepares, proposes and records; advice, assessment or a commitment to a customer remains human work. It does not guarantee an outcome: a model can make mistakes, which is why seeing the proposed change before approval matters more than a promise that nothing will ever go wrong.
And it is not an end in itself. If a fixed rule works better for a process, we build that fixed rule.
What happens when the agent does not know something?
It should say so instead of inventing something plausible. When in doubt, it refers the question back to a person, with the sources it did find.
Does an agent run continuously or only when we start it?
Both are possible. An agent can respond to a signal, such as an incoming email or a completed conversation, or run on request. What fits depends on the process.
Can one agent use several systems at once?
Yes, that is the point: the work usually sits between systems. The actions it may perform in each system are defined for each integration.
How do we know what the agent has done?
Through the log: sources consulted, proposals made, who approved them and what was written afterwards.