AI Is Everywhere at DSAG 2026: Which Procurement Use Cases Deserve Your Budget?
At the DSAG Jahreskongress in Cologne, about one in three items in the program references AI, and SAP has built agents deep into its procurement products. A new preprint on management illusions explains why many programs still stall, and how sponsors can tell good use cases from bad ones.
AI Is Everywhere at DSAG 2026: Which Procurement Use Cases Deserve Your Budget?
I am writing this at the DSAG Jahreskongress 2026 at the Koelnmesse in Cologne (6 to 8 October). DSAG is the German-speaking SAP user group, and its annual congress is where SAP customers compare notes. This year you cannot get past artificial intelligence (AI). Almost every vendor on the exhibition floor has jumped on the AI train.
If you sponsor AI programs, especially in procurement, this is an opportunity you should not miss. It is also the moment to be strict, because not every AI use case deserves your budget. Below: what the program shows, what SAP has built, how adoption really looks, and five questions that separate good use cases from bad ones.
What the program shows
I counted the items in the printed program: every session title, plus the abstract or theme label where the program prints one (keynotes and the stage). An item counts if it names AI, KI (the German abbreviation), artificial intelligence or Künstliche Intelligenz (spelled out), agents, Joule (SAP's AI assistant) or large language models explicitly.
The spelled-out forms are rare: Künstliche Intelligenz appears in three items, artificial intelligence in none. Items that only say "autonomous", "intelligent" or "assistant" are not counted. Counting them too lifts the total to 139 (39 percent), so 35 percent is a floor. The SAP keynote abstract claims that AI has made the leap from promise to practice. The organizer sets a more careful tone: the DSAG foreword asks for "Einordnung statt Euphorie und Nutzenbewertung statt Buzzwords" (classification instead of euphoria, benefit assessment instead of buzzwords). That is exactly a sponsor's job.
SAP has built AI deep into the product
The volume is not just marketing noise. SAP has invested significantly in AI inside its product offering, and fast:
SAP's own congress program carries titles such as "All-in on AI: real AI transformation at SAP" and "Procurement 2030: the way to autonomous purchasing". Session titles in this post are translated from German.
The reality check: adoption lags the announcements
Available is not adopted. The DSAG Investment Report 2026 (198 respondents, surveyed from December 2025 to January 2026) found that 43 percent of companies have implemented AI use cases and 51 percent have not. Of those who have, 77 percent run them in production or use with non-SAP solutions, and only 3 percent rely on SAP for this. In the test phase the split is 65 percent non-SAP against 8 percent SAP.
DSAG itself asks to read these figures in the context of the survey basis, and the survey closed before SAP's May announcements. The message for sponsors holds anyway: the technology is arriving fast, and the distance between announcement and productive use is where programs are won or lost.
What a new preprint says about why programs stall
Projects stall often. Gartner predicted in June 2025 that over 40 percent of agentic AI projects will be canceled by the end of 2027, a backdrop the preprint cites. The preprint "Management Illusions About Agentic AI" (SSRN 7406178, September 2026) has nine authors from academia and industry, among them Johannes Habel and Jan Zawadzki. It lists ten misconceptions that executives hold about AI agents. The authors define agentic AI as software that perceives, reasons and acts autonomously, usually built on large language models. An assistant that a person prompts and checks at every step does not count. They validated the ten illusions with a survey of 167 technical experts. That is a convenience sample and a preprint, not a controlled study, so I read it as a well-informed warning, not as proof.
The four illusions the experts rated most critical are the downsizing, testing, scaling and end-to-end illusions. The finding that stuck with me concerns the very first step. In one large AI transformation managed by one of the authors, only about 10 percent of the identified automations needed an agent built on a large language model. The rest ran on deterministic, rule-based workflows, often to the surprise of the managers involved. The authors' advice: AI agents should be the last resort for automation, not the first.
Good and bad use cases in procurement
This table is my own assessment, applying the preprint's logic to procurement. AI earns its place where the input is messy and a person can check the output. It is a poor fit where a rule already does the job, and it is risky where the output writes to a system of record without a gate.
| Procurement use case | Fit | Why |
|---|---|---|
| Extract clauses, prices and renewal dates from supplier contracts | Good | Unstructured text, a person reviews, errors are cheap to catch |
| Classify requisition and invoice lines into material groups and accounts | Good | Reasoning on messy input, accuracy is measurable by sampling |
| Compare supplier bids across unit price, freight and payment terms | Good, buyer decides | A narrow subtask, the buyer keeps the award decision |
| Match standard purchase order invoices to goods receipts (three-way match) | Weak for an agent | A rule does this deterministically, cheaper and repeatable |
| An agent that releases payments or places orders without a human gate | Not yet | Writes to the system of record, needs approval, logging and rollback |
| One agent for the whole sourcing process, from request to award | Not yet | End-to-end illusion: reliability drops as process complexity rises |
| A business case built on fewer buyers | Rethink | Downsizing illusion: the value is quality, speed and spend under management |
One more distinction matters when you read vendor material: assistant or agent? SAP names many offerings assistants that orchestrate agents. An assistant waits for a person at every step. An agent acts on its own. The risk and the testing effort are very different, so ask which one you are buying.
Five questions before you sponsor the next use case
Testing is the one most sponsors underestimate. Agents are stochastic, so repeated tests produce a distribution of outcomes, not a pass or fail. The preprint's advice is to define an approval threshold and the residual error rate the business accepts, and to reserve domain experts from day one. In procurement those experts are your buyers and category managers, who rarely have spare hours.
On scaling, the preprint describes a pilot built to prove the concept on the cleanest data. The second implementation took considerably longer and needed substantially more resources. In procurement, look at supplier master data quality, plant-specific approval rules and local legal requirements before you promise a rollout date.
On value, the preprint cites Gartner research suggesting that up to 30 percent of workers laid off because of AI may be rehired within three years, and it notes measured productivity gains that trail expectations. A local gain, such as faster requisition handling, does not become a company-wide gain while approvals and goods receipt downstream stay the bottleneck.
Guardrails for agents that write
Procurement agents do not only read. They create requisitions, goods receipts and payment proposals. The preprint's governance advice is concrete: grant tool permissions per agent, give each agent its own identity, treat untrusted input such as a supplier's PDF or email with suspicion, log every action, put approval gates in front of high-consequence writes and keep a path to revoke and roll back. No set of preventive controls is complete, so detective and corrective controls are not optional. The authors also argue against waiting for the perfect framework: build governance step by step alongside the first pilots.
Procurement on day 3 (8 October)
If you sponsor procurement AI and you are in Cologne on Thursday, the printed program lists these slots:
Use the five questions above as a lens in each of them. Times and rooms are as printed on 1 October and may change.
Work with me
I help sponsors tell good AI use cases from bad ones, with both sides of the expertise: SAP (S/4HANA, procurement processes, data and interfaces) and AI (agents, evaluation, governance, and my PhD research on integrating AI with enterprise resource planning systems). If you plan, or need to rescue, an AI program in procurement, get in touch or find me on LinkedIn. For five patterns that work inside an SAP estate, see my earlier post Künstliche Intelligenz in SAP Systemen.
Sources and method
I extracted the item titles from the printed program and read the printed abstracts and theme labels of keynotes and stage slots. An item counts as AI if it contains AI, KI, agent (also agentic, Agenten), Joule, LLM (large language model), GenAI, generative, Copilot, Machine Learning, Chatbot, Künstliche Intelligenz or artificial intelligence. Speaker and company names are not checked, and the counts cover what is printed, not what is said on stage. The numbers for SAP's program, partner talks and espresso sessions can be off by about two. I can share the full list of counted items on request.
Paul Oesterwitz
AI & SAP Consultant · PhD Researcher