SAP's tech chief says AI can build an app but not run a business

AI tools can now build an app in minutes, and some people take that to mean traditional business software is finished. Philipp Herzig, chief technology officer at SAP, says that confuses building an app with running one. He made the case in an interview with The Rundown AI, done in partnership with SAP.

Building is not running

Herzig gave an example. Generating an expense app is one thing. Running it across 40 countries means keeping up with tax rules, controlling who can see what, and matching reimbursements to payments. He says AI agents, which are programs that carry out tasks on their own, need trusted data, business context and security to do that work well.

He also says this gap explains why many AI pilots never reach real use. Turning a prototype into a system a large company can depend on, while keeping it simple for the user, still takes a lot of human work.

The risk of confident wrong answers

Herzig calls the riskiest AI answers "fluent-but-wrong." They sound confident and are well worded, but they miss how the business actually runs. Company data sits in tables, and a field only makes sense with the process behind it. A supplier in purchasing, for example, has to be recognized as the same company in finance.

SAP's answer is a Knowledge Graph, which maps what a company's data means. The interview also covers a record of how past decisions were made. The excerpt cuts off before the details.

What SAP is selling

SAP's newest product, Joule Work, lets people say what they want done instead of clicking through several programs. It can also build a chart, table or dashboard that the user refines by asking. Herzig says he uses it to gather updates on what his engineers are building, so he can spend more time in conversation. He says it does not make the judgments for him.

Joule Studio is SAP's tool for building custom AI agents. Herzig says it gives developers a head start, with sign-in, permissions and compliance already in place. He sees developers' work shifting toward deciding what is worth building and checking that results are correct and safe.

What to do

SAP sells enterprise software, so it has a stake in this argument. Still, the adoption advice holds for any vendor. Herzig says adoption was harder than the technology, because useful AI changes who makes which decisions. The fastest companies gave access first to people whose judgment the organization already trusted.

• Pick one high-value task instead of rolling AI out everywhere.
• Give it first to colleagues your team already respects.
• Let their results, not a training program, persuade others.
• Name someone who owns checking that the output is correct.

The excerpt does not cover how to keep AI bills in check or SAP's models built for business data.