EU AI Act, NIST and SAP AI Governance: What Compliance Teams Need

In 30 seconds

What it is Classify each agent by use case, document the reasoning, and map controls to NIST.

You use it when You need a first, reasoned view of which AI agents may fall under EU AI Act obligations.

Next in this section Human-in-the-Loop Approval Tiers for AI Agents, by Risk and Value →

The same agent can be low or high risk depending on the job it does. Classify each use case, write down why, and map controls to NIST.

Part 24 of 26 in SAP Business AI, Explained.

The European Union (EU) AI Act cares about what you use an AI system for, not what the product is called, so one SAP agent can be low risk in one job and high risk in another. Below, a compliance team reasons through ten agents, and I point out where SAP’s tooling helps.

Not legal advice. This is a practitioner’s reasoned view to help you prepare questions for your legal advisers. It isn’t a legal opinion, and how any real system is classified depends on facts this post can’t see.

The first job is to place each artificial intelligence (AI) use in one of the Act’s risk classes and write down why.

The short version for compliance teams

In practice the Act has four buckets: prohibited practices, high-risk uses, uses with transparency duties, and everything else.
Employment decisions and individuals’ creditworthiness are listed high-risk areas, so human resources (HR) and some finance agents need a close look.
SAP has announced an AI Governance Assistant that adds EU AI Act and NIST (National Institute of Standards and Technology) risk classification to its agent inventory. It drafts, and your team still decides. Check availability before relying on it.

The everyday picture

Back to the hospital. A medical board grants privileges per procedure, not to “a doctor” in general, so the same surgeon takes blood samples freely but needs a second sign-off for heart surgery.

The Act works the same way. A prebuilt SAP agent answering leave questions and the same platform proposing pay rises sit in very different classes.

Five checks, in order

Prohibited practices first. Article 5 bans a short list of uses, and one matters for HR: systems that infer the emotions of people in the workplace, except for medical or safety reasons.
Then the high-risk lists. Article 6 makes a system high-risk if it’s a safety component of a regulated product, or if its use appears in Annex III. Annex III includes employment (for example AI for recruitment, or for decisions affecting terms of work-related relationships) and the creditworthiness of natural persons.
Next, the narrow exceptions. Under Article 6(3), an Annex III system may not be high-risk if it doesn’t materially influence decisions, such as a narrow procedural or preparatory task. Profiling of individuals removes that exception, and a provider relying on it must document the assessment.
Transparency duties. Article 50 expects people to be told they’re dealing with an AI system, unless that’s obvious from context.
Record the result and reasoning in the agent register, and revisit it whenever the use changes.

On timing, the Article 6 page lists 2 December 2027 as the date Annex III high-risk obligations apply. The Act has been amended since adoption, so check the official text in force before you plan around any date.

Hana’s first pass over ten agents

NextSAP Utilities, the fictional Dubai utility used across this site’s SAP content, is headquartered in the United Arab Emirates (UAE), with shared services in India. Hana, the AI governance lead, starts with a sharp question: does the EU AI Act apply at all?

Article 2 covers providers placing AI systems on the EU market wherever they’re based, and providers and deployers outside the EU where the system’s output is used in the Union. NextSAP is in Dubai, so the Act matters mainly if an agent’s output is used in the EU, for example through a planned EU services joint venture, or where an EU-based provider’s system is involved. Both are possible, so Hana takes the cautious view and classifies now.

NextSAP is fictional, and its agents are built from scenarios SAP has published. UAE and Indian law are out of scope here. For local requirements, check official government sources and your own advisers.

With Daniel from Internal Audit and Mariam from HR, she works through the ten agents. Their reasoned first pass:

#	Agent	Likely class	Reasoning
1	Leave and policy Q&A	Transparency duty	Talks to employees, so tell them it’s AI. Answers policy and decides nothing
2	New-joiner onboarding	Borderline: legal review	Employment area, but creates tasks and access requests by role, not by behaviour. Possibly a narrow procedural task, so document the reasoning
3	Salary-change request	Likely high-risk	Pay is a term of the work relationship, an Annex III employment use
4	Payment clarification (receivables)	Likely minimal	Matches payments to invoices and decides nothing about a person’s rights
5	Collections prioritiser	Borderline: legal review	Today it covers the commercial (COMM) portfolio. If it scored individual households on payment behaviour, it would edge towards profiling and creditworthiness, and households depend on electricity and water, so I’d take extra care
6	Billing anomaly check	Likely minimal	Flags billing errors for people to review
7	Purchase requisition creator	Transparency duty	Conversational with employees, otherwise an internal procurement task
8	Sourcing and supplier risk	Likely minimal	Assesses companies, not individuals. Check sole traders
9	Maintenance request initiation	Likely minimal	Raises requests for a planner to approve. Annex III also lists safety components in the supply of water and electricity, so confirm it never acts as one
10	Case classification and routing	Likely minimal	Routes customer cases such as billing disputes, move-ins and outage reports, but doesn’t decide outcomes

Half the agents (4, 6, 8, 9 and 10) come out likely minimal risk. Two carry a transparency duty, two need legal review and one is likely high-risk.

Two design rules came out of the session. No HR agent will infer employee emotions from text or voice, which Article 5 prohibits in the workplace outside medical or safety reasons. And the collections prioritiser stays on commercial (COMM) customers only until legal review closes.

Where NIST fits alongside the Act

The Act says what’s expected, and the NIST AI Risk Management Framework helps organise how. Its four functions map neatly onto SAP tooling and the Act’s classes.

NIST function	What it means	Where it lands at NextSAP
Govern	Culture, policy and accountability, across everything	Hana’s board, the RACI (responsible, accountable, consulted, informed) chart, the agent register
Map	Context, purpose and impact, supporting a go or no-go	The classification table above, per agent
Measure	Testing before deployment and during operation	Evaluations, override rates, sampling
Manage	Responding to risks and incidents	Pause rules, rollback, retirement

What this exercise gets you, and what it doesn’t

Done well, every agent gets a recorded, reasoned risk class, the few needing legal attention surface early, and one control structure (NIST) covers every agent.

It won’t replace a legal opinion or cover UAE, Indian or other non-EU rules. The bigger risk, in my experience, is quieter: an agent’s use expands bit by bit and the classification never gets re-run.

Explain it to your CFO

European rules sort AI by what it’s used for. About half of our agents are routine, one is sensitive and a few need a legal look.
Anything touching pay or people’s jobs gets the strictest treatment and stays with people.
We write down why each agent sits where it does, so we can defend it later.

Questions to settle with legal

Do we classify even if the Act may not apply? I’d say yes. It’s cheap, and markets and customers change.
Who signs the classification? Risk and compliance proposes, the governance board approves, and legal reviews every high-risk or borderline case.
When do we re-classify? Whenever an agent’s purpose, data, tools or autonomy level changes.

Go deeper

Technical detail for architects and developers

SAP News (22 September 2026) describes SAP AI Agent Hub as inventorying agents, large language models and MCP (Model Context Protocol) servers, mapped to capabilities and owners. It announces a new AI Governance Assistant that embeds EU AI Act and NIST compliance intelligence into that inventory, automates risk classification and reassesses risk when governed systems change. An SAP Community post on agent management describes an EU AI Act extension that classifies the risk tier and tracks obligations, 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.

What I’d insist on: classify the use, not the product, and write down why. A risk class without reasoning can’t be audited, so for each agent record the intended purpose, the Annex III area considered, whether profiling of individuals happens, any Article 6(3) condition relied on, and who signed. Re-run it when the register shows a change in tools, data or autonomy.

Provider or deployer? With SAP prebuilt agents you’re probably using someone else’s system, while agents you build in Joule Studio may give your organisation provider duties. The obligations differ, so ask your legal advisers agent by agent.

For borderline agents, design for the strictest class you might land in. Build in human oversight, logging, override and stop from day one, because retrofitting after a reclassification is far harder. The approval tiers post covers the oversight design.

NIST describes Govern as cross-cutting, informing the other three functions. Map supplies context for an initial go or no-go decision, Measure includes testing before deployment and in operation, and Manage covers response, recovery and communication when incidents happen.

Words you’ll hear

EU AI Act: the European Union regulation that sets rules for AI systems by risk class.
Annex III: the Act’s list of high-risk use areas, including employment and creditworthiness.
Provider and deployer: roughly, whoever develops and markets an AI system, and whoever uses it.
Profiling: in plain terms, automatically assessing personal aspects of someone, such as behaviour or reliability.
Transparency duty: telling people they’re dealing with an AI system.
NIST AI RMF (Artificial Intelligence Risk Management Framework): a voluntary US framework built on Govern, Map, Measure and Manage.
AI Governance Assistant: a new SAP capability, announced to add risk classification to the agent inventory. Check availability.

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

Does the EU AI Act apply to a company headquartered in the UAE?

It can. Article 2 brings providers and deployers outside the EU into scope where the output of their AI system is used in the Union, and covers providers placing systems on the EU market wherever they are based. Whether it applies to you is one for your legal advisers.

Is an HR agent automatically high-risk?

Not automatically, but employment is an Annex III area. AI used for decisions on terms of the working relationship such as pay, promotion or termination is listed. Treat any agent touching those decisions as likely high-risk until legal review says otherwise.

Will SAP's AI Governance Assistant classify our agents for us?

SAP has announced it as a new capability, describing it as automating risk classification and embedding EU AI Act and NIST compliance intelligence into the agent inventory. I'd treat its output as a strong first draft that your compliance team reviews and signs. Check availability before relying on it.

Is the NIST AI RMF a legal requirement?

No. NIST describes it as intended for voluntary use. It's still a useful structure for the controls a regulation expects, under four functions: Govern, Map, Measure and Manage.