SAP's CEO said the quiet part about AI agents on his own keynote stage

An Authority Report is where we read what people with real operating responsibility said in public, in their own words, and work out what it means for a company that isn’t them.

Keynotes are the last place you expect a useful admission. At SAP Sapphire in May 2026, Christian Klein made one anyway, roughly twenty minutes after announcing that agents would run the business. For anyone in the DACH mid-market deciding what to believe about enterprise AI, that pair of statements is more informative than any analyst report published this year.

The vision, stated plainly

Klein’s framing for the event was not modest. SAP is “building nothing less than a new SAP,” he told the room in Orlando, and the company is “becoming a business AI company.” The product behind it is the SAP Business AI Platform, and the destination is what SAP calls the Autonomous Enterprise, where agents run the business.

If that were the whole speech, this piece would not exist. Every vendor said a version of it in 2026.

The threshold

The line worth writing down came when he talked about what enterprise AI has to clear before it touches anything real:

If AI runs payroll, financial close, or supply chain planning, 80% accuracy is not good enough.

Sit with the three examples. They are not customer service, not marketing copy, not the pilot use cases everyone starts with. They are the processes where an error is a legal event. Klein picked the hard cases on purpose, and then put a number on the pilot-grade performance most demos actually deliver.

Eighty percent is roughly where a competent LLM sits on a messy real-world process with no guardrails. It looks superb in a demo, because a demo is twelve examples and the failures are the ones nobody clicked.

On payroll, 80 % accuracy means one employee in five gets paid wrong.

Which is why Klein named payroll, financial close and supply chain planning rather than the pilot use cases everyone starts with

The admission

Then, in the same event’s press conversations, he said this about SAP’s own AI assistant:

Talking about lessons learned, we know that Joule is not perfect. When you are talking to some of our partners and customers, probably also here, the results are not really accurate. Is it really compliant? Is it really governed in the right way? Or does it really require a lot of handholding?

And:

I have to admit, there’s probably a lot of handholding behind to let the agents connect to the wide data fields and making sure they understand the process context.

That is the CEO of Europe’s largest software company saying, on the record, that agents need a great deal of work behind them to connect to the data and understand the process. Not “the model isn’t smart enough.” Handholding, data connections, process context. Those are integration problems, and they are exactly the problems that do not get solved by waiting for a better model.

Why the two statements fit together

Klein also gave the reason the gap exists, and it is the most concrete thing he said all week:

When we started five years ago to harmonize our data model, I was actually overwhelmed to understand, “Oh, there are one million correlations between logistics and finance…”

Five years, at SAP, with SAP’s own data. That is the size of the job underneath every autonomous-enterprise slide.

The two statements are not a contradiction. The vision describes where a system lands once its data and process context are in place; the admission describes what it costs to get there. Vendors normally publish only the first. Klein published both, and the second is the one you can plan against.

He added the constraint that matters for anyone buying:

One lesson I learned during my whole career is technology means nothing if you don’t bring it to adoption.

We can’t have our customers waiting for three years until they finish the modernization of the system to actually use AI.

What this means for a company of 200, not 200,000

Most companies in the DACH mid-market are not running SAP’s data model, and that is good news here. Your correlations are not in the millions. But the shape of the problem is identical, and three things follow from what Klein said.

Pick processes by the cost of being wrong, not by how impressive they sound. Klein’s own examples are the ones where 80% fails. Start where an error is visible, recoverable and cheap, a draft a human sends, a summary a human reads, a classification a human can override. Then earn your way toward the ones with legal consequences.

Budget for the handholding, because it is the work. If SAP needs it against its own ERP, your agent needs it against a CRM three people have been maintaining by hand. The build is rarely the model. It is the connections, the field-level cleanup, and the process context that has never been written down anywhere. That is the part we spend most of our time on, and it is why a blueprint is more useful than a demo.

Do not wait for the platform. Klein’s own objection to a three-year modernisation applies to the smaller version too: you do not need a finished data landscape to automate one chain end to end. You need one chain, its data, and a person who owns the result.

The useful reading of Sapphire 2026 is not that SAP is behind. It is that the biggest enterprise software company in Europe, with five years of head start on its own data, is describing the same bottleneck a 50-person company hits in week two. The bottleneck was never the intelligence.

Frequently asked questions