Governed AI agents, explained What is a governed AI agent?
A governed AI agent is software that carries out a defined business workflow inside limits people agree in writing before it is built. It only takes actions from an approved catalog, asks a named person to approve anything that touches money, access or customers, and logs every action, approval and exception for review.
Why governance comes first
An agent that acts needs rules before it needs autonomy.
Unlike a chatbot, an agent changes things: it codes an invoice, creates an account, sends a reminder. That makes one question more important than model choice: when should a person step in?
A governed agent answers that question in its design, in writing, before any code is written. The business owner defines the outcome, IT defines the guardrails, and the agent works inside both. If a request is unclear, incomplete or outside its catalog, it stops and asks instead of guessing.
The five controls
What makes an AI agent governed.
-
01
Decision rights up front
Who approves what, at which threshold and with what evidence is written into the design before build, not discovered in production.
-
02
A bounded action catalog
The agent only acts on the systems, roles and actions you have approved. Anything new goes through you first.
-
03
People approve what matters
Payments, access and credential changes, and anything customer-facing always follow a named person's approval.
-
04
A full audit trail
Every action, approval and exception is logged, so someone can read, question and correct what the agent did.
-
05
Ownership and portability
Code and documentation live in your Git, the agent runs in your cloud and model accounts, and the design stays portable across model providers.
Where people approve
What governance looks like in each function.
| Function | The agent does | A person approves |
|---|---|---|
| Finance | Captures, codes and routes invoices; sets up new vendors with documents collected. | New vendors, bank detail changes and every payment. The agent never releases funds. |
| Operations | Pulls signals into one morning briefing, with the next action drafted for each item. | Every drafted action. The agent doesn't change systems on its own. |
| People | Builds the access plan for a new hire, transfer or departure from the role catalog. | Exceptions and all privileged access, by a named approver. |
| Revenue | Researches target accounts and drafts first touches and follow-ups. | Every email, by the seller, before it's sent. |
| IT and security | Sorts and enriches alerts, groups related ones and drafts the response. | Every response, by an analyst. |
See every agent and its approval points on the agents by function page.
How it compares
Governed agents, RPA and chatbots.
| RPA bot | Chatbot | Governed AI agent | |
|---|---|---|---|
| What it does | Repeats fixed, scripted steps | Answers questions in a conversation | Carries a workflow through your systems |
| Varied inputs | Breaks when inputs change | Handles language, not actions | Handles varied inputs like invoices and requests |
| Approvals | Usually built around the script | Rarely applicable | Written into the design, with named approvers |
| Audit trail | Depends on the tool | Conversation logs | Every action, approval and exception logged |
Choosing a provider
Questions to ask anyone building agents for you.
If a provider can't answer these in writing before build, the agent isn't governed yet.
- Who approves what? Ask for the approval matrix: each decision point, its threshold and its named approver.
- What can it never do? Releasing funds, granting privileged access and contacting customers unreviewed should be off the table by design.
- Where does it run? In your cloud subscription and model accounts, or in the provider's?
- Who owns the code? You should get the source code, runbook and documentation in your own Git.
- Who runs it after go-live? Agents need monitoring, tuning and model updates every month. Find out who carries that work.
- How is it measured? A baseline before build, and a monthly report against it afterwards.
Questions
Governed AI agents: the straight answers.
What is a governed AI agent?
A governed AI agent carries out a defined business workflow inside limits people agree in writing before it is built. It only takes actions from an approved catalog, asks a named person to approve anything that touches money, access or customers, and logs every action for review.
What is human-in-the-loop for AI agents?
It means a person reviews or approves specific agent actions before they take effect. In a governed agent, those approval points are named in the design, with a threshold and an approver for each one.
Will a governed agent ever approve payments on its own?
No. In the agents Intras AI builds, finance approves every payment and the agent never releases funds. Access and credential changes also follow a person's approval.
What does an AI agent audit trail include?
Every action the agent took, every approval and who gave it, and every exception it raised. The monthly report summarizes what the agent did and where people stepped in.
Who owns a governed agent?
You should. Intras AI delivers source code and documentation to your Git, and the agent runs in your cloud subscriptions and model accounts.
How do we get started?
With a 30-minute Opportunity Scan. We identify three workflows where an agent can take on routine work, with a first business case for the strongest one and no commitment to build.
Managed Intelligence Opportunity Scan
Find the first workflow worth governing.
In 30 minutes we map three workflows, where a person should approve, and what an agent would change.