Challenge us
← Back to overview

AI Agents That May Act – but Only Through Your Process

Reading is easy, acting is delicate

Comic illustration: a clerk sends a capsule into a wall of pneumatic tube stations; in the office next door the sales director stamps incoming cases and sends one back with a correction.

An AI agent that reads applications, looks things up in tables and gives an assessment is quickly built. It gets difficult the moment it is supposed to change something: create a customer, draft a quote, send an email. Then comes the question that keeps almost every AI project from going live: who controls what the AI does?

The usual answer: give the model write access to an interface and hope the instructions in the prompt hold. In the workflow application we take a different route. The principle is: the AI decides what needs doing. Your process carries it out.

How it works

You hand the agent process steps, for example “Create customer”, “Create quote” or “Send quote”. You model each of these steps in the workflow as usual. That is where you define exactly what happens: whether data is saved, who has to agree beforehand and who gets notified.

The agent chooses for itself which step it needs and when, and fills in the required details. Then the process takes over. The agent learns the result and continues working with it.

So the AI never changes anything around your data. Every change runs along a path you drew yourself.

Example 1: Maintaining customer master data

An employee writes: “Muster GmbH has moved from Kassel to Marburg, and Beispiel AG is set up twice, please delete it.”

The agent looks up both customers and changes the address immediately. Deleting, however, it may not do directly: according to the process, this step first goes to master data management for approval. If they reject it, for instance because invoices are still open, the agent learns the reason and records it in its result.

Same agent, same request, two different paths. Which path applies is not in the prompt but in the process model.

Example 2: A quote with approval

A customer wants maintenance for 20 installations and would like a 12% discount.

  1. The agent first has the terms checked. A second agent that knows the discount scale handles this. Its answer: 8% at most.
  2. The agent creates a quote at 8% and sends it to sales management for approval.
  3. Management rejects it: “Please 5% at most, the customer often pays late.”
  4. The agent revises the quote and resubmits it. Once approved, it goes to the customer by email.

Nobody prescribed this order; the agent chose it itself. Management still had the last word.

Example 3: Several steps at once

If a customer needs two quotes, say for two sites, the agent creates both in one go and sends them for approval together. If management approves one and rejects the other, the agent receives both answers together and revises only the rejected one.

What you gain

Control. What the AI may do is in the process diagram, not in a hidden setting. Delicate steps such as deleting or sending you tie to an approval.

Traceability. The history of every case records what the agent wanted, who approved or rejected it and what was saved in the end.

Flexibility. The agent orders the steps itself and responds to rejections. You do not have to model every special case in advance.

Clear limits. Cost, run time and number of steps you cap for each agent individually. In an emergency you stop a running agent.

Why the detour through the process

One might object: why so roundabout? A model with direct database access would be connected faster.

Connected faster, yes. But then all control sits in the prompt, and a prompt is not a permission. Through the process, every action of the agent is subject to the same rules that apply to people: the same approvals, the same notifications, the same history. If a rule changes, you change the process, and the agent follows it from its next run. We described the same principle for our AI chat for an equipment database: no special rights for the AI, no unnoticed changes.

That way the AI works right inside your process, not around it.

Further reading