Künstliche Intelligenz in SAP Systemen

Some corners of SAP look like they shipped in another decade. The screens can read old, and it is easy to conclude the whole system is finished. I read the SAP PRESS handbook *Künstliche Intelligenz in SAP Systemen* (Sebastian Sav, Michael Schwarzenberger, 2026) cover to cover, and the book says the opposite. This post is the report: where a Large Language Model actually reaches into a live system, and what that takes.

The uncomfortable question

The standard move today is to bolt a chatbot in front of everything and call it innovation. The book's quiet counter-argument: a chat answer is not the job. On an ERP, an output only belongs to the business once it has handed off to a real process, and a person can inspect the result. So the honest test for every feature is simple. Does the answer reach the ledger, the order, or the org chart, with the evidence attached?

That test cuts through the marketing in both directions. It is also why this matters now: the newer releases embed the intelligence inside the processes that already hold the data, instead of adding a layer of chat beside them. The trend runs further still, from assistant features to agentic systems that plan and run whole workflows on their own, while accountability stays with the people who steer the process.

The strategy underneath

None of this is a bolt-on. The book draws a single spine: the SAP Business Technology Platform as the ground, with data management, model training and operations under one architecture, and foundation models orchestrated and served up through Joule or your own apps. An Intelligent Scenario Lifecycle Management framework carries every scenario from idea into production, with monitoring, roles and endpoints kept visible, so every project and every model version stays traceable end to end. SAP's own ethical frame for the programme runs on three words: relevant, reliable, responsible. An AI that adds nothing to the process does not belong in it, and an output nobody can inspect never ships.

The harness is coming

The whole thread runs through one layer that SAP is still rolling out: Joule Studio. The Agent Builder reached general availability in late 2025. The next version, where intent-based development really pays off, went into Early Adopter Care in June 2026 with general availability targeted for the third quarter. State the intent and the platform generates the agent that calls the real system's APIs and actions, plus a pool of some two hundred packaged agents across finance, supply chain and people to build on instead of starting from nothing. The same rule that makes the five cases below trustworthy is the rule this layer is meant to carry into your own custom work.

Five use cases, one pattern

1. Provisions with a trail

The period-end close is the most time-critical stretch in the finance calendar, and provisions for recurring expenses are where it gets messy. The rules for what belongs to which period are written as prose, not as if-then rules, and a rule-based accrual engine cannot interpret sentences. The system already automates accruals for goods receipts and services from structured data; what the rule-based generation could never do is read a guideline, and reading guidelines is exactly the work here.

So the guideline goes in as a document, grounded by SAP Document AI, which turns the prose into context: the text is split into semantically meaningful passages and mapped so the model can look up what applies to which expense category. From there the agent compares guideline against the current state of the subledger, flags the gaps where a provision should exist and does not, and assembles each posting proposal with reasoning, historical values and source attached. The entry stays a draft until a person reviews and releases it, and the reversal at the start of the next month stays a deliberate decision. Nothing reaches the books on autopilot.

2. Reports you can act on

The Review-Booklet-App Konzernabschlüsse (F6133) turns the consolidated income statement, balance sheet, cash flow and cost centers into summaries you can drill into with quick actions. No extra tool, no request to IT, no analyst needed for every follow-up question. A variance becomes a claim you can click through to the data underneath. The book names the role shift too: analysts stop being processors of data and become shapers of the result, because causes and correction proposals now arrive in the same pass as the numbers. Decision quality rises not because the model is clever, but because the evidence travels with the summary.

3. Documents that find their home

Supplier quality certificates and invoices arrive in every format imaginable. The document pipeline extracts the fields that matter, matches them against purchase orders and supplier master data, and files each document with the right record. Where the match is ambiguous, the case stays open for review rather than forcing a guess into the books. The same extraction feeds downstream quality and procurement processes, so one pass earns its keep twice.

4. Procurement with a forecast

A regression scenario trained on one hundred thousand to one million records of purchasing history predicts when material actually arrives: supplier performance, material, order size, seasonality included. A consolidated view of in-transit and at-risk items shows a bottleneck while there is still time to act. The model is retrained as fresh data arrives.

5. People, not just positions

In SuccessFactors, the gap between today's role and a target role shows up as competencies, courses and mentors, and simple bullets grow into measurable objectives. Recruiting drafts keep their quality while the repetitive typing is automated away. Roughly thirty minutes per posting across some two hundred a year is real money, and the fix is a workflow, not a rewrite.

And one more, same logic

SAP Cash Application closes the circle in receivables: payments have to be matched against open items at the end of the sales process, and the application does with machine learning and precise pattern matching what a clerk once did by eye. Matched items clear automatically, unclear cases surface as clarification work, and the reminder process keeps its rhythm. Data quality and compliance are not checked afterwards; they are built in, because closing and reporting lean on exactly this data.

The through-line

All five follow the same rule. The model interprets, prepares, proposes. A human reviews, decides, and releases. The evidence stays where the business already knows to look. That is what makes the difference between intelligence in the loop and intelligence as a lottery.

And the pattern generalizes beyond SAP: anything that must be booked, cleared, hired or defended needs the same thing. An AI that can be trusted is one that leaves the paper trail.

The real point about legacy

Writing the system of record off as old school confuses the look of the screen with the function of the system. The core transactions stay, precisely because they are proven. Contemporary artificial intelligence does not evict them. It moves through them. The work is not replacing what you have. It is making the processes around it faster and honest about what is decided, and by whom. Where screen and model both feed the same ledger, the ledger wins, and the ledger is the asset.

Sources