AI Safety and Evaluation
Get SAP ready for AI agents. Then prove they are safe.
We prepare your SAP landscape for agents, set what each one may do, and test it before it goes live. That covers Joule, Microsoft Copilot and the agents you build yourself, with one senior person working inside your team.
Three fixed steps: readiness check, safety review and evaluation, then a running programme. Start with the first and stop when you have enough.
Autonomy is earned. Drag to see how a policy changes.
example, made-up dataRuns alone 3
- Read open purchase orders
- Draft a supplier reply
- Summarise an invoice dispute
Asks a person first 5
- Create a purchase request under the limit
- Post a goods receipt
- Update a supplier address
- Change a supplier bank account
- Assign an SAP role
Never runs 2
- Pay an unknown account
- Import a transport to production
The problem
AI is being switched on in SAP landscapes that were never prepared for it.
Accounts nobody owns
Service, technical and interface accounts pile up over years. An agent can end up running with one of them.
Agents inherit access
An agent works with the roles of whoever runs it, including access left over from old projects.
Nobody tested it
Most agents are checked for whether they work. Few are checked for what they do when asked for something they should refuse.
Each vendor governs its own side
SAP and Microsoft both ship agent controls. The crossing between them is yours to design.
The agent is not the risk. An agent nobody prepared for, tested or governs is.
What we do
Ready, rules, test, run. Each step ends with something you can hand to an auditor.
- Step 01 of 4
Get SAP ready
Before any agent is switched on: give every service and technical account an owner, remove leftover access, and label the data that must not leave SAP.
- Step 02 of 4
Set what each agent may do
For each kind of action: runs alone, asks a person, or never runs. People and agents follow one set of segregation of duties rules.
- Step 03 of 4
Test before go-live
We run each agent against its rules: requests beyond its role, instructions hidden in a document or an email, and attempts to skip the approver. You see what passed, what failed and what was fixed.
- Step 04 of 4
Run it as a programme
Govern, map, measure, manage, every quarter. Limits start tight and widen as the record shows the agent acts correctly.
Step 1 of 4
What you get
Six documents your security, SAP and audit teams can use the next day.
Tap a document to see what's inside.
- Service, technical and interface accounts: who owns each and what it can reach
- Leftover and over-wide access an agent would inherit
- Data that needs a label before it leaves SAP
- Each agent, its identity, its owner and the access it runs with
- The risky actions, grouped by business process
- Ranked, so you fix the largest exposure first
- For each risky action: runs alone, asks a person, or never
- The rule that the approver is never the requester
- Which rule sits in SAP, which in Entra, which in Purview
- How to stop an agent that misbehaves, and who may do it
- The test cases, by business process
- Passed, failed and fixed, for each agent
- What to test again after every change
- What is recorded for each agent action, and where
- How long it is kept
- A sample walk-through for your auditor
- Named owners for each agent and each rule
- What to fix in 30, 60 and 90 days
- What is measured and reported to leadership
Works with what you have
Next to SAP's and Microsoft's tools, not against them.
Both vendors give you controls for their own agents. We design the rules that go into them, test that they hold, and cover the part that sits between them.
SAP gives you
A list of agents and a gateway
- AI Agent Hub: which agents exist and may run
- Joule Studio and the agent gateway: where agents are built, and where outside agents come in
- The roles and authorizations an agent inherits
Microsoft gives you
Identity and monitoring
- Entra Agent ID: an identity for each agent
- Purview: data protection and audit that now cover agents
- Sentinel: monitoring that includes SAP
We do
Readiness, rules and proof
- A landscape that is ready before the agent arrives
- The rule for each risky action, and an approver who is not the requester
- A test of every agent against those rules
- One evidence trail across SAP and Microsoft
Who does the work
One senior person, working inside your team.
Madhu Guthikonda, founder of UX4Tech
- Holds the Trusted AI Safety Expert (TAISE) certificate from the Cloud Security Alliance and Northeastern University.
- More than 20 years in SAP security, GRC and identity.
- Runs UX4Tech's own team of AI agents under the same approvals and evidence rules this page describes.
How we work
- 01
In your systems, not in a slide deck
We work in your SAP landscape and your Microsoft tenant, with the access you grant, next to your SAP and security people.
- 02
Short cycles
Something useful every week: a cleaned-up account list, a rule your team agreed, a test that now passes.
- 03
Your team keeps the result
Every rule, test and record stays with you. Your people can run the programme without us.
Who this is for
You are turning AI on in SAP, or you already have.
- You want Joule or Copilot on SAP data and are not sure the landscape is ready.
- You are turning on Joule in S/4HANA and need to know what it can reach.
- Copilot or your own agents already reach your ECC system.
- Your auditors or your board are asking how AI is governed and tested.
How to start
How to start
- 01
AI readiness check
Fixed scope. Ends with the AI readiness report.
- 02
Safety review and evaluation
The register, the approval policy, the hold design and the evaluation report.
- 03
AI safety programme
The evidence pack and the quarterly rhythm. Continues under vCISO Advisory.
Built on the NIST AI RMF: govern, map, measure, manage.
Questions
What people ask before the first call.
Do we need S/4HANA or Joule to start?
No. If Copilot or your own agents reach SAP data, on ECC or on S/4HANA, the work applies. The readiness check is useful before any agent exists.
Do you build the agents?
Not in this service. We get the landscape ready, set the rules and test the agents your team or your implementation partner builds.
Is the evaluation a certification?
No. It is a test report against your own rules and the NIST AI RMF. It tells you what to fix. It is not a certificate.
Do you install software?
No. The work is design, clean-up, testing and documents. The rules go into controls you already own in SAP and Microsoft.
Does this replace SAP's AI Agent Hub or Microsoft Entra?
No. Those tools list and identify your agents. We design what each agent may do and who approves, then place those rules in them.
Who needs to be in the room?
Your CISO or security lead, your SAP security lead, and whoever owns the AI rollout. Internal audit is welcome.
Go deeper
Our research behind this page.
Is your SAP landscape ready for an agent today?
If you can't answer in one sentence, start with the readiness check. We tell you which accounts, access and data to fix before anything is switched on.
Or write to support@ux4tech.com
Insights and Architecture Briefs
