AI Agent Governance Operating Model: Who Signs Off, and When

In 30 seconds

What it is Six roles, eight steps, one accountable signature per step, evidence in the register.

You use it when You are setting up, or fixing, how your organisation approves and oversees AI agents.

Next in this section EU AI Act, NIST and SAP AI Governance: What Compliance Teams Need →

Tools keep the register, but people carry the accountability. Every agent needs a named owner, and every autonomy increase needs a named signature.

40-second version of this explainer, no sound.

Part 9 of 26 in SAP Business AI, Explained.

Agents are arriving faster than anyone takes responsibility for them. On projects, the question I hear most is who signs for an agent, not which one to build. Here’s who proposes, approves, tests, signs off, watches and retires one, and where SAP’s tools keep the evidence.

AI agent governance is an agreed set of rules about who may let a digital worker start, do more or stop, and who checks afterwards that the rules held.

Governance on one page

It’s an operating model you agree, with six roles, eight lifecycle steps and one accountable person per step.
SAP AI Agent Hub keeps the register of agents and owners. SAP has also announced an AI Governance Assistant for risk classification (check availability), and SAP Signavio helps decide where agents are worth it.
The riskiest moment is rarely go-live. It’s the quiet decision, months later, to raise an agent’s autonomy, which deserves a formal sign-off with evidence.

What a hospital board would do

Think of a hospital, with specialist doctors as the agents. A new doctor doesn’t start operating because one department head likes them. A medical board checks credentials, grants privileges for specific procedures and can suspend a licence.

Artificial intelligence (AI) agents need the same board, and the AI Agent Hub is the human resources (HR) register for these digital employees. Raising an agent’s autonomy is like granting new surgical privileges, and it deserves the same formal review.

Who sits at the table, and what they sign

Name six roles. Each agent needs people in these seats:
- Business owner, who owns the process and its outcome (collections, say) and answers for both value and harm.
- Agent owner, the named person who answers day to day for the agent’s behaviour, tuning and incidents.
- IT (information technology) platform owner, who runs the platform, the agent’s identity, deployments and technical monitoring.
- Risk and compliance, who classify the risk and check obligations such as data protection.
- Internal audit, which stays independent: it tests that controls work and never approves what it will later audit.
- AI governance lead or board, who set policy and approve new agents and autonomy changes.
Use one lifecycle for every agent: propose, approve, build, test, go live, raise autonomy, monitor, retire. Prebuilt and custom agents differ only in effort.
Give each step one accountable person. A RACI grid (Responsible, Accountable, Consulted, Informed) makes this explicit. If a row has two “A”s, accountability is already blurred.
Keep the evidence where the agent is registered. SAP describes AI Agent Hub as an inventory of agents, large language models and Model Context Protocol (MCP) servers, mapped to capabilities and owners. Agents deployed from Joule Studio are registered there automatically, after a built-in evaluation step.
Review on a rhythm. I’d suggest a monthly agent dashboard for owners, a quarterly board review and an annual internal audit sample.

An SAP Community post on agent management warns that tooling gives you a registry, verification and measurement, but some properties aren’t available out of the box: the agent owner, the verification criteria, the autonomy level from which a human approves, and which business activities an agent may take over. Those are the decisions this operating model exists to make.

Who signs when the collections agent asks for more

NextSAP Utilities, the fictional Dubai utility used across this site’s SAP content, runs a collections prioritiser, agent 5. It’s a prebuilt agent with an extension, scheduled and at L2 Propose. Each morning at 06:00 Dubai time it reviews about 2,400 overdue open items in the commercial and property-management (COMM) portfolio, worth AED 18.6 million (UAE dirham), and proposes about 150 calls, ranking who to chase, each with a suggested next step within dunning procedures 01, 02 and 03. Collectors on the shared-services team in India, working for Elena, the collections lead, accept or change each one. It can flag an account for a credit block, but Elena’s team decides; the agent never blocks on its own.

Its first quarterly scorecard (post 07) showed 88% accepted and a 17% override rate, so it was retrained. After the retrain it ran another 90 days, and this time the numbers look good. Collectors accepted 98.4% of about 9,800 proposals unchanged, a 1.6% override rate, and the few overrides cluster on large key accounts such as Al Noor Property Management, whose collective bill account rolls many premises (the individual properties supplied) into one bill. That passes the evidence gate from post 02. Elena asks to move the agent to L3 Act with approval. After a collector approves a batch, the agent would send standard reminder letters and set follow-up dates itself.

Here’s who signs what:

Who	Role	What they do for this change
Collections process lead in Elena’s team	Agent owner	Prepares the evidence pack: acceptance rate, override reasons, incident log, letter templates
Priya	IT platform owner	Confirms the agent can send reminders only, and cannot change credit limits or post write-offs. Tests the stop switch
Risk and compliance	Risk	Checks whether any accounts belong to individual consumers and, if so, reviews data protection and European Union (EU) rules where relevant
Daniel	Internal audit	Consulted, not approving. Asks for a weekly 5% sample of executed actions for the first eight weeks
Elena	Business owner	Responsible: signs the business case
Hana’s governance board, with Layla signing as CFO (Chief Financial Officer) for finance agents	Governance	Accountable: approves L3 with conditions

The board approves the move to L3 with conditions: reminders only, never write-offs, and credit blocks stay a human decision. Accounts above AED 250,000 stay at L2, so a person still acts on them. And if the override rate climbs above 15% in any week, the agent drops back to L2 automatically and the board takes another look.

None of this is exotic. SAP lists customers such as Uniper building a custom procurement agent in Joule Studio, alongside prebuilt agents, and once both kinds run side by side you need one register with one owner per agent.

How heavy should each gate be?

Every agent follows the same strip, with signatures at the human gates.

Gate weight should follow risk. Here’s where I’d start:

Decision	Low-risk agent	Medium-risk agent	High-risk agent
Approve a new agent	Governance lead	Governance board	Board plus formal risk sign-off
Go live	Business owner, lead informed	Business owner, board consulted	Board, after audit walkthrough
Raise autonomy	Business owner and lead	Board, with evidence pack	Board, and never above L2 for pay, money or legal outcomes
Monitoring rhythm	Monthly	Weekly for the first 60 days	Weekly, plus audit sampling
Retire	Agent owner	Business owner	Board

SAP tooling slots in at the edges. SAP describes Signavio’s Process Consulting Agent as recommending where agents would create the most value, which feeds the “propose” step, and SAP Cloud ALM (Application Lifecycle Management) as tracking agent actions once live, which feeds “monitor”. AI Agent Hub holds the register throughout.

Where the operating model runs out

A good one gives every agent named owners, makes go-live and autonomy increases signed decisions, and leaves evidence an auditor can follow. It also settles in advance who switches an agent off. It has limits, though:

It won’t make a badly designed agent safe, and it doesn’t replace security controls such as least-privilege identities.
A tool alone can’t deliver it. The tool records decisions, and people still make them.
It doesn’t settle legal classification. That needs your compliance team.

Explain it to your CFO

Every digital worker gets a manager, a job description and a review, and it needs a signature before it handles money or gets wider powers.
If it misbehaves, we already know who switches it off and who explains what happened.

Three rules to agree before the next agent

Who signs autonomy increases? The governance board is accountable, the business owner signs the business case, and finance agents need the CFO on the board.
What evidence is enough to raise autonomy? At least 60 to 90 days at the current level, acceptance and override rates, an incident log and a permission review.
Can an agent run without a named owner? I’d say no. An agent with no named owner doesn’t run, and an agent doesn’t get more autonomy without a signature.

Go deeper

Technical detail for architects and developers

Register and verification. The SAP product page calls AI Agent Hub vendor-agnostic discovery, inventory and governance. According to the SAP Community agent mesh post, every agent the hub discovers starts unverified, and a verification record with configurable criteria leads to verified or rejected. The same post says the runtime (Joule Studio, SAP Integration Suite and the agent gateway) must enforce that status, and notes that AI Agent Hub documentation carried an Early Adopter Care notice in October 2026. Check the current status in the Joule capability watch.

Build and test evidence. SAP’s Joule Studio post describes a native evaluation framework before production, covering technical and business quality, with offline evaluations sandboxed away from live systems. Deployment registers the agent in Joule Work and AI Agent Hub, and builders can transfer ownership to a colleague. Treat that transfer as a governance event and update the register and the RACI the same day.

AI Governance Assistant. SAP News (22 September 2026) describes a new AI Governance Assistant that embeds EU AI Act and National Institute of Standards and Technology (NIST) compliance intelligence into the agent inventory, automates risk classification and reassesses risk when governed systems change. I’d treat its output as input to the risk role, not a decision. Check its availability in the same capability watch before you plan around it. The same article cites Gartner research that 87% of IT leaders say AI adoption did not happen automatically. The operating model is what drives it.

Framework mapping. The NIST AI Risk Management Framework is intended for voluntary use and has four core functions. NIST describes Govern as “infused throughout AI risk management”, and in this model it’s the roles and the RACI. Map sets context and supports an initial go or no-go (propose and approve). Measure covers testing before deployment and in operation (test and monitor). Manage covers responding to risks and incidents (go live, raise autonomy and retire).

ISO/IEC 42001:2023 comes from the International Organization for Standardization and the International Electrotechnical Commission. In plain terms, it describes what an organisation needs in place to run AI under a managed system, and to keep making that system better over time. If you already run ISO-style management systems, this model slots in as the agent-specific part.

Words you’ll hear

Operating model: the agreed roles, steps and decision rights for running something, here AI agents.
RACI (Responsible, Accountable, Consulted, Informed): a grid showing who does the work, who answers for it, who is asked and who is told.
Agent owner: the named person answerable day to day for one agent’s behaviour and incidents.
Autonomy level: how much an agent may do on its own, from L0 Inform to L5 Fully autonomous.
AI Agent Hub: SAP’s register of agents, models and MCP (Model Context Protocol) servers with their owners.
NIST AI RMF (Artificial Intelligence Risk Management Framework): a voluntary US framework built on Govern, Map, Measure and Manage.
ISO/IEC 42001: an international standard setting requirements for an AI management system.

Keep going

Independent explainer, not affiliated with or endorsed by SAP SE. SAP, Joule and other SAP product names are trademarks of SAP SE.

Frequently asked questions

Is AI agent governance a tool I can buy?

No. Tools such as SAP AI Agent Hub keep the register, the owners and the evidence. The decisions about who may approve, raise autonomy or switch an agent off are an operating model your organisation has to agree.

Who should be accountable for an agent?

Name one business owner for the outcome and one agent owner for day-to-day behaviour. A governance lead or board approves new agents and autonomy increases. Internal audit stays independent and never signs off its own subject.

Do I need ISO/IEC 42001 or the NIST AI RMF to start?

No. They are useful structures. NIST describes its framework as intended for voluntary use, and ISO/IEC 42001 sets requirements for an AI management system. Start with roles, a register and sign-off rules, then align to a framework.

What is the most overlooked governance moment?

Raising an agent's autonomy level. Go-live gets attention, but the later decision to let the agent act instead of propose often happens informally. Treat it as a formal change with evidence and a named signature.

Does SAP's AI Governance Assistant replace the compliance team?

No. SAP describes it as adding EU AI Act and NIST compliance intelligence and automated risk classification to the agent inventory. People still decide and sign. Check the current status before you plan around it.