Autonomous Enterprise: The Autonomy Ladder for SAP Agents

In 30 seconds

What it is Six levels of agent autonomy, and the evidence an agent needs to climb one.

You use it when You are deciding how much an AI agent may do before a person signs off.

Next in this section Autonomous Enterprise Myths vs Reality: Seven Beliefs to Drop →

Autonomy is a dial per agent: case routing can act then report, while salary changes stay at propose.

40-second version of this explainer, no sound.

Part 2 of 26 in SAP Business AI, Explained.

SAP talks about the Autonomous Enterprise, and the phrase can sound like a business with nobody at the controls. It isn’t. In design reviews the real question is smaller: how much may this one agent do before someone signs? Here’s a ladder for answering that.

What SAP actually means by “autonomous”

SAP describes the Autonomous Enterprise as assistants and agents carrying out end-to-end work across finance, HR, procurement and more, within defined boundaries and with people involved where judgment is needed.
Autonomy isn’t one setting for the whole company. Each agent sits on a six-level ladder, L0 Inform to L5 Fully autonomous, and climbs one level only on evidence: accuracy, override rate, reversibility and value at risk.

A junior doctor doesn’t get a surgeon’s licence

In a hospital, specialist doctors (agents) treat patients, a ward coordinator (assistant) decides who sees whom, and a medical board (governance) licenses each doctor for specific procedures. A junior doctor starts without a surgeon’s licence and earns it procedure by procedure, with a supervisor signing off until the record shows they’re safe. The autonomy ladder is that licence, written down per agent.

The vision, and the six rungs

In an SAP Community post on SAP-managed public cloud ERP, SAP describes the Autonomous Enterprise as AI assistants and agents executing end-to-end workflows across finance, people management, procurement and beyond. It says plainly that this doesn’t mean removing people from the loop. It names five domains: Autonomous Finance, Autonomous Spend, Autonomous Supply Chain, Autonomous HCM (human capital management) and Autonomous CX (customer experience). It also argues that agents can only act reliably if the processes underneath them are genuinely connected.

SAP News (22 September 2026) covers the operating side. It positions SAP Signavio, SAP LeanIX, WalkMe and SAP Cloud ALM (application lifecycle management) as the backbone to deploy, govern and scale agents, and says the SAP AI Agent Hub inventories agents, large language models and Model Context Protocol (MCP) servers, mapped to capabilities and owners.

SAP doesn’t prescribe an autonomy level per agent. That part is yours. This series uses six levels:

L0 Inform. The agent watches and tells you something happened. No recommendation.
L1 Answer. It answers questions from approved sources and changes nothing.
L2 Propose. It prepares a complete action, such as a draft clearing or a ranked list, and a person decides.
L3 Act with approval. It executes, but only after a named person approves each action or batch.
L4 Act then report. It acts on its own and reports afterwards, while a person reviews samples and exceptions.
L5 Fully autonomous. It acts with no routine review. I don’t recommend L5 for money, pay or legal outcomes.

Ten agents, one workshop at NextSAP

NextSAP Utilities is the fictional Dubai utility used across this site’s SAP content: about 5 million customers across electricity, water, gas and district cooling, a shared services team in India, and SAP S/4HANA with SAP Utilities (IS-U) and FI-CA, plus SAP SuccessFactors and SAP Ariba. The company is made up, but each agent below is the kind SAP describes for these processes. Hana, the AI Governance Lead, placed all ten of its agents on the ladder in one workshop with Layla (CFO), Omar (Finance Controller), Mariam (HR Director), Rashid (Procurement Lead) and Daniel (Internal Audit).

#	Agent	Level	Why this level
1	Leave and policy Q&A	L1 Answer	It only explains policy. HR corrects a wrong answer, and no record changes.
2	New-joiner onboarding	L3 Act with approval	It creates tasks and access requests. Mariam’s team approves each new joiner batch, and mistakes are reversible.
3	Salary-change request	L2 Propose	Pay is personal, confidential and hard to undo once paid. It stays at L2 permanently.
4	Payment clarification	L2 Propose	It proposes how to apply incoming money and never posts; a clerk posts. A wrong clearing affects a customer’s account and credit.
5	Collections prioritiser	L2 Propose	It starts here. It ranks which commercial accounts to chase and can flag one for a credit block, which Elena’s collections team decides. Calling the wrong customer harms a relationship. Post 09 covers its later move to L3.
6	Billing anomaly check	L2 Propose	It flags unusual utility bills, such as odd consumption or an estimated read, before invoicing. Holding or correcting a bill changes revenue and customer experience, so Meera’s billing team decides.
7	Purchase requisition creator	L3 Act with approval	Field and depot supervisors use it to create requisitions, and the normal approval still applies before anything is ordered.
8	Sourcing and supplier risk	L2 Propose	Supplier choice has commercial and legal weight. Rashid’s buyers decide.
9	Maintenance request initiation	L3 Act with approval	It raises maintenance requests for network assets from events. The maintenance planner approves before work is scheduled.
10	Case classification and routing	L4 Act then report	High volume, low value per case, easy to reroute. Errors are cheap and visible.

Why can case routing sit at L4 when salary change stays at L2? NextSAP’s commercial customer centre gets about 1,400 cases a week: billing disputes, move-ins and move-outs (a customer starting or ending supply at a premises), outage reports and meter queries. An outage report sent to billing by mistake gets rerouted in under a minute at no cost. After three months the agent was right 98.6% of the time, people overrode it on 1.6% of cases, and every action could be undone. SAP’s Joule Agents page features Bosch with a classification agent, which is the same pattern.

Salary changes are the opposite. There are perhaps 40 a month, each worth on average AED 2,500 (UAE dirhams) a month to an employee, and once payroll runs, reversing one is painful and very visible. Even at 100% accuracy, the value at risk and the human impact keep it at L2. The agent still saves Mariam’s team real time by gathering the grade, band, budget and history into one proposal, so the approver spends five minutes instead of forty.

Nine of the ten agents sit at L1 to L3. At NextSAP, the Autonomous Enterprise is mostly agents preparing very good proposals.

When does an agent earn the next rung?

I’d let an agent move up one level only when all four tests pass, measured on your own data over a set period. These are the thresholds I’d start with:

Evidence	Question	Starting threshold
Accuracy	How often was the proposal right?	At least 98% over 90 days or 200 decisions, whichever is later
Override rate	How often did people change or reject it?	Under 2%, with no rising trend
Reversibility	Can a wrong action be undone, and how fast?	Fully reversible within one business day, with no customer or employee harm
Value at risk	What is the worst single action worth?	Below a limit set by the owner, for example AED 50,000 per action

Then the process owner and the AI governance lead both sign. At NextSAP that might be Rashid and Hana for a procurement agent, with Daniel from Internal Audit reviewing the evidence pack once a quarter.

Some agents are capped whatever the evidence says. At NextSAP, pay, legal outcomes and money movement stay at L2 or L3, and nothing goes to L5.

Limits of the ladder

The ladder makes autonomy a recorded decision per agent instead of a gut feeling, and gives auditors something to test against. It has limits, though.

It isn’t an SAP product setting. You implement it through approvals, roles and agent configuration.
It won’t make an agent accurate. It only limits the damage.
It doesn’t replace monitoring. SAP News says SAP Cloud ALM adds monitoring of agent actions to catch deviations, but someone still has to read it. This is the part teams forget once the go-live party is over.

Explain it to your CFO

Each digital helper gets a licence that says exactly how much it may do without a person signing.
Helpers earn more freedom only after months of measured, correct work, and lose it when results slip. Anything that moves money or changes pay always has a person deciding.

Setting your own defaults

What’s the starting level for new agents? I’d use L2 Propose for anything touching money, pay or customers, and L1 for pure question answering.
Who signs a climb? The process owner and the AI governance lead together, with audit reviewing quarterly.
What are your caps? No L5 for money, pay or legal outcomes, and salary changes capped at L2. Whatever you pick, autonomy is earned per agent, with evidence.

Go deeper

Technical detail for architects and developers

Each level maps to controls. L1 needs read-only access through the agent gateway and approved grounding sources. L2 needs the agent to produce a structured proposal object rather than free text, so the approver sees exactly what will be posted. L3 adds a workflow approval step before the write call, and the write should run under the approver’s or the requester’s identity instead of a shared technical user. L4 needs post-action reporting, sampling and an undo path designed before go-live.

SAP News describes the SAP AI Agent Hub as inventorying agents with their capabilities and owners, and a newly announced AI Governance Assistant adding EU AI Act and NIST compliance intelligence with automated risk classification (check availability). I’d record the autonomy level, cap and last climb decision with each agent’s entry there, or in your own register.

On risk frameworks, the AI Risk Management Framework from the US National Institute of Standards and Technology (NIST) organises work into Govern, Map, Measure and Manage functions. The climb gate is a Measure and Manage control.

The Open Worldwide Application Security Project (OWASP) Top 10 for large language model applications, 2025 edition, lists Excessive Agency as item LLM06. That’s the risk the ladder exists to contain. Prompt injection, item LLM01, is why even an accurate agent shouldn’t hold more permissions than its level needs.

For evidence, log every proposal, the final human decision and any edit. Without that log you have nothing to justify a climb with.

Words you’ll hear

Autonomous Enterprise: SAP’s vision of agents carrying out end-to-end work within defined boundaries, with people involved where judgment is needed.
Autonomy level: how much an agent may do before a person signs off, from L0 Inform to L5 Fully autonomous.
Override rate: the share of an agent’s proposals that people change or reject.
Value at risk: the worst realistic cost of one wrong action by the agent.
Human-in-the-loop: a person approves before the action takes effect, the old four-eyes or maker-checker principle.
SAP Cloud ALM (application lifecycle management): SAP’s tool for running solutions in production, which SAP News says now monitors agent actions.

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 the L0 to L5 autonomy ladder an SAP standard?

No. It is the framework used in this series to make design decisions concrete. SAP describes the Autonomous Enterprise as agents acting within defined boundaries with people involved where judgment is needed. The ladder is one way to write those boundaries down.

Does the Autonomous Enterprise mean no people in the process?

No. SAP's own description says it does not mean removing humans from the loop. Agents act within boundaries, and people stay involved wherever oversight and judgment are required.

Can a prebuilt Joule Agent start at a high level?

It can be configured that way, but I recommend that every new agent starts at the lowest level that still adds value, then climbs on evidence from your own data.

Can an agent move down the ladder?

Yes, and it should. If accuracy drops, overrides rise or the process changes, drop the agent one level until the evidence recovers.

Which agents should never reach L5?

Anything that moves money, changes pay or creates a legal outcome. In this series L5 is not recommended for those, whatever the accuracy.