How AI agents are governed in regulated companies

Timotheos Samartzidis

I am an enterprise architect in Wiesbaden. I build experiments and write about what has to be settled before an AI agent may act in a company where rules and audits apply.

An agent can find its own way to a result. Who approves, what is recorded and how an action is undone stays with the organisation. The six points below are what I have worked out about this so far. On this site I speak for myself and not for my employer.

Six points on governing agents

Each point links to the essay behind it. The overall idea is described in Governed Intelligence.

The path of one action
Outcome one named commitment Agent chooses the route,proposes an action System check authority, budget,required approval Service executes the action Result confirmed, failedor unknown Person approves when review is needed evidence back to the agent
  1. 01

    One outcome and its consequences

    A review starts with one outcome the agent can bring about and what follows from it, and then asks who can authorise each change.

    An agent architecture review starts with the commitment

  2. 02

    Flexible route, explicit authority

    An agent may choose how to reach a result. Who may approve a purchase, and which service may place it, is still decided by the company.

    When to use a workflow and when to let an agent choose the route

  3. 03

    Permission enforced by the system

    A model can interpret a document or propose a next step. The surrounding system decides when review is needed and enforces whether an action may proceed.

    Designing Agentic AI for the Enterprise

  4. 04

    Unknown results

    If a booking succeeds and the reply is lost, the agent sees a timeout. The system has to keep the outcome as unknown until it can check with the service that executed it.

    Authority and recovery have to survive the handoff

  5. 05

    Budgets enforced in code

    Time, tool calls, cost and repeated attempts are limits that code enforces. A model's confidence does not establish that a task is complete.

    An agent architecture review starts with the commitment

  6. 06

    Onboarding before autonomy

    An agent gets what a new colleague gets: a role, scoped access, curated context, an escalation path and a probation period.

    Why every AI agent needs onboarding before it gets autonomy

Recent writing

All writing

Experiments and workshops

Graph-walk experiment

An experiment on whether a knowledge graph can change an agent's next decision. Budget, stopping rule and failure modes are written down.

Read the essay

Character Reactor

Two characters receive the same words and react differently. A local model runs in the browser; state and decisions can be inspected.

Open the app

Klarogram

A transcription app that runs offline in the browser. The recording does not leave the device.

Product page

Workshops: a remote workshop for about 200 people in R&D on producing reviewable HTML reports from Excel data with an in-house AI portal, followed by a second one on company design and printable PDF output. Participants later applied the method to their own tasks.

About me

I work as Principal Systems / IT-Architect Engineer at R+V Versicherung AG, in the Center of Excellence Enterprise Architecture. Before that I worked in IT compliance and architecture governance, and led a team of around 17 people on connected rehabilitation products at LIME Medical.

My background is research in biometrics and computer vision at Fraunhofer IGD, followed by software and product architecture.

Areas of focus
  • AI Governance
  • Enterprise Architecture
  • Agentic Systems
  • Governance & Security
  • Systems Thinking
  • Risk Management
  • Cloud Architecture
  • MedTech

A first conversation

If you are wondering what AI is going to change in your team's work, and where to start, I would be glad to talk. A first conversation is an exchange about one concrete question. It is free and commits neither side.

Write to me [email protected] All contact details

Useful in a first message
  • What the team works on
  • What you suspect might change
  • What makes it hard to decide

Book a time

You can also book a time directly. The calendar is loaded from Google only after you allow Google Calendar in privacy settings.

Open the Google Calendar booking page directly →