How to Test a Fiori-Style SAP App with an AI Agent

See how an AI agent can read plain-English test cases, drive a real browser on a Fiori-style SAP app, and catch regressions with screenshot evidence.

Wed Nov 18, 2026

Regression testing is the part of every SAP release that everyone agrees matters and nobody has enough time for. Scripts break when the UI changes. Manual runs get squeezed into the last two days before a transport goes to production. And the test cases that do exist often live in a spreadsheet nobody has opened since UAT.

AI agents offer a different approach. Instead of writing a brittle script for each scenario, you write the test case the way a functional consultant already would, in plain English, and an agent works out how to run it in the browser.

This post explains how that works, what it looks like on a Fiori-style Sales Orders app, and where the limits are.

What "AI test automation" means here

The setup has three parts:

  1. A plain-English test case. For example: "Create a sales order for customer C200 with material M-1001, quantity 100. Expect a 3% discount and the net value to be shown."
  2. A browser the agent can control. A headless Chromium instance, driven through Playwright using the Model Context Protocol (MCP). The agent can open pages, read what's on screen, click, type and take screenshots.
  3. An AI agent with rules. Built with Google's Agent Development Kit (ADK) on Gemini. It reads the test case, decides the steps, runs them in the browser, compares what it sees with what was expected, and returns a verdict: PASS, FAIL or ERROR, with evidence.

The important shift is that the test case is the source of truth, not the script. If a button moves, the agent still finds it, because it reads the page the way a person would.

A worked example: catching a pricing regression

Imagine a Fiori-style Sales Orders app built on OpenUI5, with a few business rules taken from a functional spec:

  • Orders of 100 units or more get a 3% discount.
  • If the order value pushes a customer past their credit limit, the order is saved with status Credit block and a warning is shown.

You write eight test cases from that spec. Two of them are a boundary pair:

  • TC-04: quantity 100 → discount applied.
  • TC-05: quantity 99 → no discount.

Now a new release goes out, and a developer accidentally changes the discount threshold logic.

When the agent runs the suite:

  • On the stable release, all 8 tests pass.
  • On the defective release, TC-04 fails (the discount shows ₹0 instead of 3%) while TC-05 still passes.

That pattern is exactly what a good tester looks for: the boundary moved. The agent saves a screenshot of the order screen and a JSON report with the expected and actual values, so a developer can see the problem in seconds.

Why boundary pairs matter more than the AI

The agent is only as good as the test cases you give it. In practice:

  • Happy-path-only test sets miss most regressions. If you only test "quantity 50, no discount", you'll never notice the threshold moved.
  • Write expected values, not just steps. "Check the discount" is vague. "Expect discount 3%, and net value = quantity × price × 0.97" is testable.
  • Cite the rule. Each test case should say which line of the functional spec it checks. When a test fails, you know which business rule broke.

AI makes running tests cheaper. It doesn't replace good test design. If anything, it rewards consultants who already write clear specs.

Guardrails: keep the agent on a short leash

An agent that can click anything in a browser needs firm limits, especially near business systems:

  • Origin lock. The browser is only allowed to open the practice app's address. Navigation anywhere else is blocked by code, not by a polite instruction in the prompt.
  • Tool filtering. Browser tools the agent doesn't need (file uploads, arbitrary JavaScript) are removed.
  • Private services. The browser and the agent run as private Cloud Run services with least-privilege service accounts.
  • Test systems only. Point this at a QA client with test users and resettable data. Never at production.

The common mistake is to treat the prompt as the security control. The prompt asks for good behaviour. The guardrails enforce it.

FAIL vs ERROR: reading the results properly

A useful agent reports two different kinds of "not passed":

  • FAIL: the app did something different from what the test expected. That's a real defect, or a wrong test case.
  • ERROR: the agent couldn't complete the steps (a page didn't load, an element wasn't found). That's an environment or test-design issue, not necessarily a product bug.

Keeping these separate stops your team from chasing false alarms.

Where this fits in an SAP project

  • Regression before each transport: run the suite on the QA system after every release.
  • UAT support: turn agreed UAT scripts into agent-run checks, so business users only review the failures.
  • Hypercare: re-run critical scenarios daily in the first weeks after go-live.

It won't replace your test management tool, and it isn't a substitute for an enterprise automation suite on a large programme. But for a functional consultant who wants to understand and demonstrate AI testing, it's the fastest way to get hands-on.

Try it yourself

Our Lab 3: AI Agent That Tests a Fiori-Style SAP App walks you through the whole build on your own Google Cloud project, in 7 modules:

  • the practice app
  • a locked-down browser
  • the ADK test agent
  • test cases from a functional spec
  • a regression run on two releases
  • evidence and an auto-checker

It has a 17-page guide and a capstone where you write your own test case. Every step is checked before you move on.

→ Lab 3 on Brainix Tech

Not affiliated with or endorsed by SAP SE or Google. The app, data and customers in the lab are synthetic.