AI hallucinations compound across 100+ process steps in finance, warns SAP CFO

The AI hype cycle has been dominated by ever‑larger models and impressive zero‑shot benchmarks, but real‑world finance workflows can involve hundreds of sequential steps. When hallucinations compound statistically down that chain, the probability of a critical failure skyrockets. Dominik Asam’s call for ‘excruciating assurance levels’ is a rallying cry for a new branch of applied AI—one that prioritizes reliability, data grounding, and cost over raw model size.

SAP's finance chief has drawn a stark line in the enterprise AI sand, arguing that the technology must climb beyond chatbot 'low-hanging fruit' before it can deliver the returns businesses expect. Speaking after the German software giant's second-quarter 2026 results on July 23, CFO Dominik Asam told reporters that the "lion's share" of AI token consumption today is spent on relatively safe tasks like coding assistants and chatbots, where errors carry limited risk. But applying AI to finance, supply chain, and other core business processes is an entirely different challenge—one where hallucinations compound across multiple steps, raising compliance and operational risk. The remarks mark a strategic pivot for SAP, which processes an estimated 77% of the world's transaction revenue, as it seeks to embed AI deeply into its enterprise resource planning (ERP) and cloud applications without exposing customers to runaway token costs or embarrassing failures.

“

SAP itself has invested heavily, releasing its Joule AI copilot and embedding AI across its S/4HANA cloud suite.

Asam was blunt about the hurdles. In finance workflows, he noted, "If you have some hallucinations in the process, the errors will actually compound statistically over many steps. It requires much more excruciating assurance levels." This is not a theoretical concern. A single error in an automated invoice matching system, for instance, can cascade through payment runs, account reconciliations, and regulatory filings. For supply chain processes, the risks are equally tangible: an AI misclassifying a customs document or misreading a shipment notice can delay billions of dollars in goods. The first generation of enterprise AI—typified by generic large language models bolted onto chat interfaces—has been a useful proof-of-concept, Asam suggested, but the real prize lies in governed, process-specific systems that understand the unique data and rules of each company.

This message comes at a critical juncture. Global enterprise spending on generative AI is projected to soar, yet chief information officers are increasingly vocal about the gap between hype and measurable productivity gains. SAP itself has invested heavily, releasing its Joule AI copilot and embedding AI across its S/4HANA cloud suite. But Asam's warning signals an internal recognition that scaling AI in mission-critical areas demands a fundamentally different approach—one that prioritizes reliable, cost-effective inference over access to the most powerful frontier model. He explicitly called out the fallacy that "AI will solve all these problems if they are messy, legacy data silos." Instead, companies must first make their own data usable and governed, so that AI can operate with the knowledge of the company rather than guessing from a generic corpus.

Cost is a central part of the equation. Asam pointed to "extremely high token costs" when companies attempt to run large models on unstructured, ungoverned data. In practice, he predicted, enterprises will gravitate toward the cheapest reliable tool that can deliver the required outcome safely, whether that is simple software, an open-source model, or a top-tier cloud API. This aligns with a broader industry trend toward small language models (SLMs) and retrieval-augmented generation (RAG) architectures that ground outputs in a company's own verified data. For SAP, this philosophy could reshape its product strategy: rather than promoting a single AI assistant, SAP might accelerate domain-specific agents for accounts payable, demand forecasting, compliance checks, and other narrowly defined tasks where accuracy can be tightly controlled.