AI agent sprawl: stop 10 agents becoming 100

In 30 seconds

What it is One front door for new agents: check the need, reuse first, register, retire.

You use it when Business teams have started asking for, or quietly building, their own AI agents.

Next in this section Multi-agent conflicts in SAP: stop loops, clashes and double actions →

Most agent ideas do not need a new agent. Check the need, reuse first, register what you build, and retire what no longer earns its place.

Part 26 of 26 in SAP Business AI, Explained.

AI agents are easy to build now, so they multiply. Without a front door, ten useful agents turn into a hundred half-owned ones, each with its own access and its own risk.

The fix I recommend is a light intake path. Every new agent comes through one door, shows it fixes a real process problem, reuses what already exists, gets registered with an owner and has a planned end.

Why sprawl happens, in brief

Sprawl is a known problem. SAP’s documentation says agents built independently across business units lead to lost visibility, duplicate agents and governance gaps.
Intake takes five steps, from a one-page request through reuse checks and registration in SAP AI Agent Hub to planned retirement.
Adoption doesn’t happen by itself. SAP News cites Gartner: 87 percent of information technology (IT) leaders said AI adoption did not happen automatically. A front door is how you steer it.

The everyday picture

Back to the human resources (HR) register for digital employees. No well-run company lets managers hire whoever they like, on any contract, with any building access. There’s a headcount request, a check that the role doesn’t already exist, a contract, a badge, and an exit process when someone leaves.

Agent intake does that for digital employees. It isn’t there to say no, only to give every new hire a job, a manager and a desk, and get the badge back from leavers.

From request to retirement in five steps

A one-page request covering the problem, the process step, the business owner, the systems the agent would touch and the autonomy level it needs, from L1 Answer to L4 Act then report.
A “where does it help” check against process data, before anyone builds anything. According to SAP News, the SAP Signavio Process Consulting Agent looks at how processes perform in SAP and other systems, points out where work gets stuck and suggests where an agent would pay off. If the data shows no problem, park the idea.
Reuse before build. I check, in order, the agents you already run, prebuilt Joule Agents from SAP, and the catalog of standard SAP AI agents that SAP’s documentation describes in SAP AI Agent Hub. Build only when none fits.
Register and verify. New agents go into SAP AI Agent Hub with an owner, capability, risk tier and autonomy level, starting as unverified and moving to verified or rejected after review, as SAP Help describes.
Plan the exit. Every approved agent gets a review date and a retirement trigger. SAP describes the hub’s lifecycle as running from planning through decommissioning.

Hana’s first quarter of intake

A year into running ten agents, NextSAP Utilities, the fictional Dubai utility used across this site’s SAP content, was getting new requests weekly. Hana, the AI governance lead, set up intake with Priya, the IT Platform Owner, and Daniel from Internal Audit.

The company and its numbers are invented, but the shape of this first-quarter funnel is what I’d expect:

40 requests came in through the one-page form, from HR, finance, procurement and field operations.
22 passed the “where does it help” check. The other 18 showed no measurable bottleneck in the process data or were really reporting requests.
9 of the 22 were met by reuse. Five field supervisors wanted a leave policy bot for their crews, which the existing leave and policy agent already was. Three procurement requests matched the purchase requisition creator, and one a prebuilt agent pattern.
7 were parked, with no business owner willing to sign.
6 new builds were approved and registered in SAP AI Agent Hub as unverified.
5 were verified and moved to pilot. One was rejected for asking for write access to supplier bank details at L4 Act then report. It came back later at L2 Propose.

At the same review, Hana retired two agents. One was a duplicate billing check for estimated meter reads that had been merged into the main billing anomaly agent in the first weeks (as post 07 describes) but was still sitting in the register with live access. The other was a reporting bot unused for 60 days, whose owner had left. Retiring them removed their access and schedules.

The quarter ended with five agents in (one sent back), two out, and no new agent without an owner. Finance estimated reuse alone avoided about AED (UAE dirham) 180,000 of build and run cost in the year (an illustrative figure).

Where each kind of request should go

Request looks like	Route	Why
“Answer questions about a policy”	Reuse an existing Q&A agent, add content	Same job, new knowledge
“Do what SAP already offers for this process”	Prebuilt Joule Agent	Less to build and maintain
“Automate a step unique to us”	New custom agent, full intake	Real gap
“A dashboard or report”	Not an agent	Reporting tools fit better
“An agent someone already built quietly”	Register, verify or reject	Bring it into the light
“An agent with no owner any more”	Retire	Remove access and identity

What a front door can’t catch

Intake stops duplicates before anyone builds them, every new agent arrives with an owner, a purpose and a risk tier, and Internal Audit gets one list to test instead of ten.

It can’t catch agents that skip the front door, though. Discovery in SAP AI Agent Hub helps, but only for the platforms you connect. Intake doesn’t judge business value on its own either. The process check and quarterly review do that. And keep it light, because if intake takes three months, teams will route around it.

Explain it to your CFO

Before anyone builds a new agent, we check the problem is real and whether one we already pay for can do the job.
Every new agent gets a named manager and an approval before it starts.
Agents that stop earning their keep get switched off.

Three rules to agree up front

Who runs intake? I’d give it to the AI governance lead, with a ten-working-day decision target.
What’s the minimum to register? A named business owner, the process step, systems touched, autonomy level and a retirement trigger.
What’s your retirement rule? No business owner for 60 days, two quarters below target, or a better agent doing the same job.

Reuse before you build, and retire before you add.

Go deeper

Technical detail for architects and developers

Discovery is your safety net. SAP’s documentation says discovered agents land in a review inbox before entering the inventory, and SAP’s product page lists automatic discovery from SAP, Microsoft, Google, Amazon Web Services (AWS), Databricks and ServiceNow. Use it to catch agents that bypassed intake and send them through the same verification.

For reuse, SAP’s documentation lists an SAP solution explorer, a catalog of standard SAP AI agents and applications, and SAP suggestions for SAP AI agents in reports. The SAP Community Agent Mesh post adds that SAP agents can be imported from the reference content catalog in SAP LeanIX. Check them before approving a custom build.

SAP Help says reviewers see the fact sheet attributes that administrators configured as verification criteria. Line your intake form up with those attributes and the request becomes the first draft of the register entry.

On process evidence, SAP News describes the latest release of the SAP Signavio Process Consulting Agent introducing process mining capabilities. The Agent Mesh post describes SAP Signavio comparing process KPIs (key performance indicators) before and after an agent’s rollout. Record the expected effect at intake to give the quarterly review a baseline.

Some SAP AI Agent Hub features become generally available in phases, according to SAP’s documentation. Check the current status in the Joule capability watch.

Words you’ll hear

Agent sprawl: uncontrolled growth in the number of AI agents, many without clear owners.
Intake: the front door for new agent requests, with one form, one review and one decision.
Reuse before build: checking existing and prebuilt agents before approving a new one.
Shadow agent: an agent built outside the intake process, often by a business team.
Process mining: analysing system event data to see how a process really runs and where it slows down.
Decommissioning: formally retiring an agent, including its access, identity and schedules.

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

Won't an intake process slow teams down?

A light one shouldn't. I recommend a one-page form and a decision within ten working days. You're trying to stop duplicates and unowned agents, not ideas.

How do I check whether an agent will actually help?

Start from the process. According to SAP News, the SAP Signavio Process Consulting Agent looks at how processes perform in SAP and other systems, points out where work gets stuck and suggests where an agent would pay off.

What if a team has already built an agent without asking?

Bring it in through the same intake, without blame. SAP AI Agent Hub can discover agents on the platforms you connect, and every discovered agent starts as unverified until someone reviews it.

When should an agent be retired?

When it has no business owner, when it repeatedly misses its targets, or when another agent does the same job better. SAP describes the hub's lifecycle as running through to decommissioning.