Five Ways to Use AI Agents in Business Processes
One word, many meanings
”AI agent” is the phrase on every product page in 2026. It means different things: a chat window, an automation step, a system that changes data on its own. For the question of what you can and may use in your own organisation, that vagueness does not help.
So we sort by a simple question: who triggers the agent, what may it do, and who has the last word? This yields five ways of using agents that we distinguish in the workflow application.
1. The assistant in a business application
A person asks a question, the assistant answers. It queries data in natural language, compares it, analyses documents and prepares changes.
- Triggered by: a person in a dialogue
- May: read and propose changes
- Last word: the user, who explicitly confirms every change
What matters is that the assistant gets no rights of its own. The application’s existing permission system applies. What this looks like in a production equipment database is described in An AI chat for a grown business application.
2. The checking agent as a process step
Here no person asks. The process reaches a step and the agent takes over: it reads an application including its attachment, looks things up in approved tables and knowledge sources and delivers a result into the fields of the workflow form. If something is missing that only a person knows, it asks instead of guessing.
- Triggered by: the process
- May: read, look up, ask, deliver a result
- Last word: the process model. The agent can only trigger one of five fixed exits: result, question, handover to a person, limit reached or error.
Typical tasks are preliminary checks: an investment request against the budget, an invoice against the purchase order, an enquiry against internal documentation. Details on the page AI agents in the process.
3. The acting agent
The next step: the agent does not only read, it acts. It creates a customer, drafts a quote, triggers an approval. It gets no database access for this, but process steps you have modelled yourself, such as “Create customer” or “Send quote”.
- Triggered by: a person or the process
- May: choose which process step to call and when, and fill in the details
- Last word: the process step. Whether data is saved, who has to agree beforehand and who is notified is defined in the workflow.
If sales management rejects a quote, the agent learns the reason, revises and resubmits. How this works in detail is shown in AI agents that may act – but only through your process.
4. Several agents working together
A single model always sounds convinced, even when it is wrong. At important steps, several agents therefore work independently of one another, if you wish with models from different providers. The process evaluates their results by a fixed rule, such as unanimity or majority.
- Triggered by: the process
- May: each agent judges on its own; a further agent prepares the decision if they disagree
- Last word: the rule if they agree, a person if they do not
Working together also means dividing work: one agent can ask a specialised agent for information, for example which discount the scale allows. More in Several AI agents, one process.
5. The AI that builds the process
The fifth kind does not work in live operation but before it. An AI assistant such as Claude or Codex creates the process itself through an open interface (MCP): flow, forms, logic, and if required the agent steps within it. It plays the finished process through as every role, and validation reports missing exits and approvals.
- Triggered by: the person who needs a process
- May: design, build and test a process
- Last word: the person who accepts the process and puts it into operation
Background on the page AI in business processes.
At a glance
| Kind | Triggered by | May | Last word |
|---|---|---|---|
| Assistant in the application | person in a dialogue | read, propose changes | user confirms |
| Checking process step | process | read, look up, ask | fixed exits in the model |
| Acting agent | person or process | call modelled process steps | the process step, with approval if set |
| Several agents | process | judge independently | rule or person |
| AI builds the process | person | build and test a process | acceptance by a person |
What all five have in common
None of these gives the AI a path that bypasses the existing rules. The boundary always lies outside the model: in the permission system, in the fixed exits, in the modelled process step, in the evaluation rule or in the acceptance. A prompt can give an instruction. It does not replace a permission.
Which kind fits your case depends less on technology than on who has the last word today. The AI should not change that.