Managing 10 AI agents in SAP with SAP AI Agent Hub

In 30 seconds

What it is Ten agents, one heatmap, one lifecycle and a quarterly scorecard for each.

You use it when You run, or are about to run, several AI agents across departments and need one way to manage them.

Next in this section SAP AI Agent Hub explained: the HR register for your digital workforce →

Ten agents become manageable when each has an owner, a status, a lifecycle stage and five numbers reviewed every quarter.

40-second version of this explainer, no sound.

Part 7 of 26 in SAP Business AI, Explained.

One AI agent you can keep in your head. Ten across HR, finance and procurement is a portfolio, and you manage it like staff: one list, one owner each, clear stages from hiring to retirement, and a regular performance review. Below are all ten agents at NextSAP Utilities, the fictional Dubai utility I use across this site’s SAP content, plus what SAP AI Agent Hub turned up on day one.

If you only read one section

Put every agent on one heatmap with its owner, risk tier, autonomy level and verification status. SAP AI Agent Hub provides the inventory and the status, according to SAP’s documentation.
Give every agent the same lifecycle: propose, verify, pilot, scale, review, retire. That’s my practical version of SAP’s planning-to-decommissioning lifecycle.
Each quarter, look at five numbers (accuracy, override rate, cost per task, incidents and value) and retire the agents that don’t earn their place.

Why ten feels so different from one

Go back to the HR register. With one employee, the manager knows everything. With ten across three departments you need a register, a probation period and an annual review, or people with no manager and no job description start to appear.

Agents go the same way. SAP’s documentation says that when business units build agents independently, organisations lose sight of which exist and who owns them, which leads to duplicates and governance gaps.

Running the portfolio, quarter by quarter

Connect and discover. Hook up every platform where agents run. Discovered agents wait in a review inbox before they join the inventory.
Fill the heatmap. Record owner, risk tier, autonomy level and status for each agent. Every discovered agent starts as unverified (SAP Help and the SAP Community Agent Mesh post both say so).
Run the lifecycle. Each agent moves through propose, verify, pilot, scale, review and retire, and only verified agents move to pilot.
Measure. Observability and process data feed the scorecard. SAP’s product page lists goal completion rate, tool call correctness and agent mining, which checks whether an agent follows the path it was built for.
Decide. Every quarter, every agent is kept, fixed or retired.

Day one at NextSAP: ten agents and a few surprises

NextSAP serves about 5 million customers across electricity, water, gas and district cooling from Dubai, with shared services in India. It runs SAP S/4HANA with SAP Utilities (IS-U) and contract accounting (FI-CA, short for Contract Accounts Receivable and Payable), plus SAP SuccessFactors and SAP Ariba. A portfolio like this is a realistic size for a utility of that scale. Hana, the AI Governance Lead, owns the portfolio, and Priya, the IT Platform Owner, connects the platforms.

The ten agents on day one

#	Agent	Business owner	Risk	Autonomy	Status on day one
1	Leave and policy Q&A (Human Capital Management, HCM)	Mariam, HR Director	Low	L1 Answer	Verified
2	New-joiner onboarding	Mariam	Medium	L3 Act with approval	Unverified
3	Salary-change request	Mariam	High	L2 Propose	Rejected
4	Payment clarification	Omar, Finance controller	Medium	L2 Propose	Unverified
5	Collections prioritiser	Elena, Collections	Medium	L2 Propose	Verified
6	Billing anomaly check	Meera, Billing	Medium	L2 Propose	Verified
7	Purchase requisition creator	Rashid, Procurement	Medium	L3 Act with approval	Unverified
8	Sourcing and supplier risk	Rashid	Medium	L2 Propose	Verified
9	Maintenance request initiation	None named	Medium	L3 Act with approval	Unverified
10	Case classification and routing	None named	Low	L4 Act then report	Unverified

That’s four verified, five unverified and one rejected. The autonomy labels come from the ladder in post 02.

What discovery found that wasn’t on the list

Priya connected the SAP platforms and the third-party platform field operations uses. Discovery came back with 13 agents, not 10.

Two duplicates. The India shared services HR team had built its own leave policy bot, doing the same job as agent 1, and the billing team in Dubai had its own check on estimated meter reads (bills based on a guess when no one could read the meter) that overlapped agent 6. Each duplicate meant more prompts and tool permissions to secure.
One shadow agent. Field operations had built a supplier onboarding agent on a third-party platform, outside the IT (information technology) team. It could read supplier bank details, and procurement had no idea. Frankly, this is the finding that gets a finance chief’s attention.
One rejected agent. The salary-change agent (agent 3) failed verification. It could read the whole salary table rather than just the requester’s team, and its tool list included a write action an L2 Propose agent doesn’t need. Hana rejected it, and Mariam’s team narrowed its access and resubmitted.
Two ownership gaps. Agents 9 and 10 were prebuilt patterns IT had switched on, agent 9 with NextSAP’s own SCADA alarm trigger added. Priya was technical owner, but no business owner had signed, and Hana won’t verify an agent without one.

The fixes took three weeks. The duplicate leave bot was retired and its users moved to agent 1. The two billing checks were merged into agent 6. The shadow agent was merged into the vendor onboarding workflow from post 04, with Rashid as owner, and then retired. Agents 9 and 10 got business owners from network maintenance and customer service.

The quarterly scorecard

At the first quarterly review, Hana, Daniel from Internal Audit and Layla, the Chief Financial Officer (CFO), looked at five numbers per agent. Here are four of the agents. All figures are illustrative for NextSAP, with money in AED (UAE dirham).

Agent	Accuracy (sampled)	Override rate	Cost per task	Incidents	Value this quarter	Decision
4 Payment clarification	92% of 300 sampled	9%	AED 1.80	0	About 750 clerk hours moved to exceptions	Keep, raise sample
5 Collections prioritiser	88% of 200	17%	AED 0.90	0	AED 2.1 million overdue collected earlier	Fix: retrain on overrides
7 Purchase requisition creator	95% of 250	6%	AED 1.20	1 (duplicate requisition)	640 hours saved	Keep, fix duplicate check
1 Leave and policy Q&A	97% of 400	2%	AED 0.15	0	3,800 HR tickets avoided	Keep

Agent 5 was retrained on its overrides after this review. Post 09 picks up its story from there.

Here’s how I read those columns:

Accuracy is checked by a person on a random sample. The agent doesn’t report it.
Override rate is how often a human changed the agent’s proposal. A high rate isn’t a failure. It tells you where the agent’s rules are wrong.
Cost per task is model calls plus platform cost, divided by completed tasks.
Incidents are anything that reached a customer, an employee or the ledger by mistake.
Value is measured against the process baseline rather than estimated. The SAP Community Agent Mesh post describes SAP Signavio comparing process KPIs (key performance indicators) before and after an agent’s rollout.

SAP’s Joule Agents page describes Uniper’s custom procurement agent built in Joule Studio and Bosch’s classification agent. In a real portfolio, each would be a row on the heatmap.

Six stages from proposal to retirement

Stage	Gate question	Who decides	Status in the hub
Propose	Is there a process problem worth solving?	Business owner	Not yet registered, or unverified
Verify	Owner, risk, autonomy, tools and tests all present?	AI governance lead	Verified or rejected
Pilot	Does it work on real cases with a small group?	Business owner	Verified
Scale	Do the pilot numbers hold?	Business owner and CFO office	Verified
Review	Keep, fix or retire this quarter?	Governance lead, audit, owner	Verified, or set back for re-review
Retire	Is it still worth its cost and risk?	Business owner	Decommissioned in the lifecycle

These stages are my recommendation, not SAP’s, but they fit inside its planning-to-decommissioning lifecycle.

What the hub can and can’t do for a portfolio

It finds agents on connected platforms, including ones IT didn’t build, and holds one status per logical agent so sandbox and production don’t drift apart. SAP’s documentation also mentions dashboards and predefined reports on portfolio, approval coverage and risk exposure.

What it won’t do:

Find agents on platforms you haven’t connected.
Merge duplicates or name owners. People do that.
Set your scorecard targets. The SAP Community Agent Mesh post is clear that the owner, criteria and autonomy thresholds stay with the enterprise.

Explain it to your CFO

We have ten digital workers. Each now has a named manager and an approval stamp, or it isn’t allowed to work.
On day one we found three we didn’t know about, and we’ve retired or merged all three.
Every quarter we check whether each one is accurate, cheap enough and worth keeping, and we switch off the ones that aren’t.

Policies to settle before agent number eleven

Who owns the portfolio? My default is one AI governance lead, with Internal Audit invited to every quarterly review.
What happens to shadow agents? No punishment, but they go through intake within 30 days or they’re switched off.
When do you retire an agent? After two quarters in a row below target on accuracy or value, or when it’s gone 60 days without a business owner.

Behind all three sits one rule: manage agents as a portfolio, not as projects.

Go deeper

Technical detail for architects and developers

Discovery scope. SAP’s product page lists automatic discovery from SAP, Microsoft, Google, AWS, Databricks and ServiceNow. The SAP Community Agent Mesh post adds that SAP agents can be imported from the reference content catalog in SAP LeanIX, and third-party agents can be created as fact sheets through an SAP LeanIX API (application programming interface). Plan for platforms with no connector.

One status per logical agent. SAP Help says that for agents discovered from an SAP environment, verification applies to the logical agent, so production and sandbox instances share one status. Name agents so two genuinely different ones don’t collapse into one logical record.

Runtime enforcement. SAP’s product page says only verified MCP (Model Context Protocol) servers can be called in production workflows and that access is revoked when status changes. SAP’s Business AI page describes the enterprise agent gateway handling identity, access and policy at runtime. Check where your landscape enforces agent status, not just tool status.

Monitoring and value. SAP News says SAP Cloud ALM (Application Lifecycle Management) AI agent monitoring tracks every agent action after deployment. The Agent Mesh post describes Agent Mining in SAP Signavio, which treats agent traces as an event log to check conformance. The same post notes that, as of October 2026, Agent Mining and parts of monitoring were not yet generally available, and that the hub documentation carried an Early Adopter Care notice. Check the current status in the Joule capability watch.

Risk classification. SAP News describes a new AI Governance Assistant (check availability) with European Union (EU) AI Act and NIST (US National Institute of Standards and Technology) compliance intelligence and automated risk classification. I’d use it as an input to the risk tier and keep a human sign-off for high-risk agents like the salary-change one.

Words you’ll hear

Agent portfolio: the full list of agents a company runs, managed together like a product portfolio.
Heatmap: a table coloured by risk and status, so problems stand out at a glance.
Shadow agent: an agent built outside the agreed process, often by a business team, with no register entry.
Override rate: the share of agent proposals a human changed before approving.
Verification status: unverified, verified or rejected, recorded against each agent in SAP AI Agent Hub.
Key performance indicator (KPI): a number that shows whether a process is doing well, such as cycle time.
Decommissioning: formally retiring an agent, which means removing its access, its identity and its schedule.

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

How many agents before I need SAP AI Agent Hub?

There is no fixed number. In my experience the need appears when agents come from more than one team or platform, because that is when duplicates and unowned agents start.

Can SAP AI Agent Hub find agents that IT did not build?

It can discover agents from the platforms you connect. SAP's product page lists SAP, Microsoft, Google, AWS (Amazon Web Services), Databricks and ServiceNow. An agent on a platform you have not connected will not appear.

What does a rejected status mean?

SAP's documentation says rejected means the asset was reviewed and not approved for use. I treat it as a fix-and-resubmit decision, not a permanent ban.

Which numbers should a quarterly agent review use?

I recommend five: accuracy on a sample, human override rate, cost per task, incidents and measured value. SAP lists observability measures such as goal completion rate and tool call correctness that can feed the first two.