TheLeverage
← Writing
November 2, 2026·7 min read· AI· Finance

SAP's Own CFO Says AI Must Move Past Chatbots: What the Q2 2026 Earnings Split Means for Finance Teams

By Michel Escoda — Independent Architect & SAP FICO Consultant
Share
Summary

SAP's own CFO, Dominik Asam, told Reuters after Q2 2026 results that enterprise AI must move beyond chatbots into complex business processes before it pays off — and that messy legacy data, not model quality, is what actually breaks the ROI case. This piece unpacks both claims, weighs the obvious self-interest against what SAP's own quarter shows, and turns the earnings-call language into four concrete sequencing moves for finance teams running on, or migrating to, S/4HANA.

SAP's own chief financial officer just told the market that chatbots are not where enterprise AI earns its keep. Dominik Asam, speaking to reporters after SAP's Q2 2026 results on 23 July, said enterprise AI "must move beyond chatbots and coding tools into more complex business processes." He added that the "lion's share" of AI token consumption today still goes to the easy stuff, coding assistants and chat interfaces, where a wrong answer costs you a few minutes of annoyance, not a misstated balance sheet.

That is not a line from a competitor trying to sell against SAP. It is SAP's own finance chief, on the company's own earnings call, drawing a line between AI that is fun to demo and AI that actually changes how a finance function runs. For a CFO or controller who has been fielding chatbot pitches for two years and wondering whether any of it moves a KPI, that is permission to ask a harder question of every vendor in the room: what business process does this actually touch, and what happens when it is wrong.

Why the distinction is harder than it sounds

Asam's reasoning is specific, and it is worth sitting with rather than skimming past. "If you have some hallucinations in the process, the errors will actually compound statistically over many steps," he said. "It requires much more excruciating assurance levels." A chatbot that occasionally invents a fact is embarrassing. An agent embedded six steps into a reconciliation or a close process that occasionally invents a fact can propagate that error through every downstream step that trusted it, and nobody notices until the numbers do not tie out.

A decision chain — an error propagating silently through several linked steps Hallucinations "compound statistically over many steps" — an error propagating through a multi-step agent chain rather than staying a single bad answer.

This is not a new idea to anyone who has run SOX controls or an internal audit program. It is the same logic that makes finance teams nervous about any automation that touches a general ledger without a human checkpoint: errors in finance do not fail loudly, they fail quietly and then compound. Asam is restating a control principle finance already lives by and applying it to agentic AI specifically. That gives finance leaders a vocabulary they can use internally that does not sound like they are dismissing AI, because they are using the vendor's own framing.

The data-quality line that should land harder than the chatbot one

The second Asam quote matters as much as the first, maybe more, for anyone actually doing SAP finance transformation work. "The idea that AI will solve all these problems if they are messy, legacy data silos is not true," he said, adding that trying to force AI to compensate for bad data comes with "extremely high token costs."

Translate that out of earnings-call language: you cannot buy your way out of a dirty chart of accounts, fragmented master data, or a Universal Journal full of inconsistent postings by throwing a frontier model at it. The model will try, and it will burn an expensive amount of compute doing something a clean data model would have made unnecessary. This is, functionally, the business case for the unglamorous work that precedes any AI rollout: chart of accounts harmonization, master data governance, clean core discipline, Universal Journal hygiene. SAP's own CFO just validated, on an earnings call, that the data-readiness work is not a prerequisite nice-to-have. It is the thing that determines whether the AI spend pays off at all.

A CFO with reasons to say exactly this

It would be naive to treat Asam's comments as disinterested truth-telling and nothing more. SAP sells Joule agents and the Business AI Platform it launched at Sapphire in May 2026, and steering enterprise AI budgets toward "complex business processes" steers them toward precisely the category SAP monetizes. Chatbots and coding copilots, the category Asam downplays, are also where the most visible competitive AI products already live, many of them not SAP's.

But that self-interest cuts in an interesting direction, not just against the argument. SAP is not claiming its own agents are risk-free. Asam described the company's strategy as "ambidexterity," deterministic, low-cost applications running alongside probabilistic agentic AI, and he was explicit that Joule's agents run on the same category of large language models everyone else uses, including frontier models from multiple providers through Joule Studio. The hallucination risk he is warning finance teams about applies to SAP's own agent stack too. The honest reading is not "trust our agents instead of chatbots." It is "any agent touching a financial process needs governance and human-in-the-loop checks that chat interfaces do not," and SAP happens to be selling tools built around that premise. Readers should weigh the message on its logic, not on who is saying it, and the logic holds up on its own.

The numbers behind the statement

The quotes landed inside a quarter that tells its own two-speed story. SAP's current cloud backlog hit 22.9 billion euros, up 27 percent reported. Cloud revenue reached 6.3 billion euros, up 22 percent, with the Cloud ERP Suite alone accounting for 88 percent of that cloud revenue. Free cash flow rose 27 percent to 3 billion euros. By the usual growth metrics, this was a strong quarter.

At the same time, SAP cut its 2026 non-IFRS operating profit outlook by 0.1 billion euros, down to a range of 11.8 to 12.2 billion, to absorb the drag from its Dremio and Prior Labs acquisitions, with management flagging a "very low triple-digit million euro" hit still to come in the second half. Software support revenue fell 8 percent and software licenses fell 32 percent, the last visible traces of the on-premise business winding down as cloud and ERP Suite take over completely.

Put those two halves together and you get a company living out loud the exact tradeoff it is telling customers to expect: AI investment that strengthens the long-term business while visibly pressuring near-term margin. CEO Christian Klein underlined the ambiguity directly on the same call, noting that AI token spend "doesn't mirror outcomes," a tacit admission that SAP cannot yet draw a clean line from its own AI spending to its own AI returns, even though it reported that R&D token usage is tracking with roughly 30 percent productivity gains internally. If the company selling the agents cannot fully quantify the payoff of its own agent investment yet, that is a useful data point for anyone being asked to sign off on an enterprise AI budget this year.

What this means for the work in front of you

None of this is an argument against AI in finance. It is an argument for sequencing it correctly, and SAP's own quarter makes the sequencing case better than any consulting deck could. The practical implications for a finance organization running on, or migrating to, S/4HANA:

Prioritize the processes where errors are forgiving before the ones where they are not. Invoice capture, AP matching, and anomaly flagging in the general ledger tolerate a human catching a mistake before it posts. A multi-step close process that chains several agent decisions together does not forgive the same way, because by the time something looks wrong, three other steps have already trusted it.

Treat data readiness as the AI budget line, not a separate project competing with it. Master data cleanup, chart of accounts harmonization, and clean core discipline are not preconditions you check off before the real AI work starts. Per Asam's own framing, they are the work that determines whether the token spend buys anything at all.

Ask every vendor, SAP included, what "excruciating assurance" looks like in their specific product. Human-in-the-loop checkpoints, audit trails, and deterministic guardrails around the probabilistic layer are the actual control story, not a footnote to it. A pitch that cannot describe its assurance mechanism in the same specificity Asam used on the earnings call is not ready for a financial process yet.

Watch how token costs get passed through in your own contracts. SAP disclosed that it charges its own internal functions for the tokens they consume. Enterprise customers should expect SAP, and every other AI vendor, to meter and bill the same way eventually. Model that line item into any multi-year AI business case now, before it shows up as a surprise in a renewal.

The headline from SAP's Q2 was never really about chatbots. It was a finance chief using his own earnings call to tell the market, and by extension every CFO watching, that the easy AI wins are mostly spent and the harder, more valuable work is still ahead. For SAP finance teams and the consultants advising them, that is less a prediction than a checklist.

Sources

Frequently asked

What did SAP's CFO say about AI and chatbots?

Dominik Asam told Reuters after SAP's Q2 2026 results that enterprise AI must move beyond chatbots and coding tools into more complex business processes before the returns show up, because the 'lion's share' of today's AI token consumption still goes to low-risk, easy-to-demo use cases rather than process-level automation.

Why does SAP's CFO say hallucinations are riskier in finance processes than in chatbots?

Because a chatbot's occasional wrong answer costs a few minutes of annoyance, while an agent embedded several steps into a reconciliation or close process can have an error compound statistically through every downstream step that trusted it — a risk that requires, in Asam's words, 'much more excruciating assurance levels.'

Can AI fix messy or fragmented SAP data on its own?

No. Asam was explicit that the idea AI will solve messy, legacy data silo problems on its own 'is not true,' and that trying to force a model to compensate for bad data comes with extremely high token costs. Data readiness — chart of accounts harmonization, master data governance, clean core discipline — is a precondition for AI ROI, not a side project.

Is SAP's CFO just talking down competitors' chatbot products?

There is an obvious self-interest angle, since SAP sells Joule agents and a Business AI Platform aimed at the complex-process category Asam is recommending. But Asam also said Joule's agents run on the same category of large language models as everyone else, including frontier models from multiple providers, and that the same hallucination risk applies to SAP's own agent stack. The caveat cuts both ways rather than only serving SAP's pitch.

Need this in your organisation?

I work with a small number of clients each quarter on ERP strategy and IT-department automation. If the questions raised above are live in your team, get in touch.

Start a conversation →
Share