SAP Replaces Costly AI Models as It Focuses on Profit per Task

SAP is replacing some expensive artificial-intelligence models with lower-cost alternatives where performance remains adequate, finance chief Dominik Asam told Reuters on Wednesday. The changes include its internal consulting assistant. The business-software company still expects total AI spending to rise as use expands. Its objective is healthier margins as the business grows.

That distinction puts a concrete business issue behind the debate over AI expenditure. A company can pay less for an individual operation while spending more overall because customers use the service more often. For shareholders, the relevant comparison is the revenue and cost attached to useful work, rather than whether the aggregate technology bill rises or falls.

SAP supplies systems used to manage finance, employees and supply chains. Its AI tools operate within those business processes, where an answer may need to become an approved invoice, a purchasing action or another completed workflow. TTN's analysis of inference economics explains why the price of running a model is only one part of the cost of delivering that result.

Customer billing and model costs are different measures

SAP's current pricing documentation separates included Base AI, Premium AI and custom-built applications. Premium services use AI Units, with agents charged by action and other features by request or processed record. Joule for Consultants uses a per-user monthly arrangement. A business task can involve multiple actions, so one completed workflow should not automatically be treated as one billable unit.

This matters when assessing the benefit of a cheaper underlying model. The customer's bill follows the applicable product and contract terms, while the supplier pays to operate the service. Lower operating expense does not necessarily mean an immediate reduction in every customer's price. Conversely, a low model price does not establish the total bill for an implementation.

Replacing a model is therefore an operating decision as well as a technology choice. If the required output can be produced at lower expense without extra rework, more revenue can remain after delivery costs. If a cheaper service needs repeated attempts or more human review, part of that saving can disappear. Performance must be judged against the business requirement being served.

The latest financial statement supplies a useful baseline

In results published October 21, SAP reported third-quarter cloud revenue of €6.668 billion, up 26% in reported currency. Its IFRS cloud gross margin was 73.9%, compared with 74.6% a year earlier. Reported operating profit nevertheless increased 18% to €2.930 billion. These are group and cloud-business measures, not a separate profit statement for its AI products.

The figures illustrate why faster sales growth alone cannot settle a margin question. More revenue can coexist with a lower percentage retained after delivery costs, while the absolute profit of the wider company still increases. The statement does not isolate the effect of Wednesday's model-selection details, so attributing the entire change in cloud margins to AI would go beyond the disclosed evidence.

Our guide to interpreting earnings makes a related distinction between growth, profitability and expectations. The next disclosures will be more informative if they show how changing delivery costs interact with customer usage and pricing, rather than relying solely on the number of announced AI features.

Adoption creates an execution test

SAP's October 26 product update described assistants that coordinate specialized agents within existing permissions and audit processes. Its examples included creating invoices and identifying supply shortages. These are the company's descriptions of intended functionality, rather than independently measured productivity gains for every customer using the software.

The practical test is whether those workflows become reliable enough to use repeatedly at an acceptable cost. TTN's enterprise-agent return analysis examines that conversion from capability to operating value. Wednesday's report adds evidence that model expense is already being managed at the product level. It does not establish a universal winning model or prove that every new AI feature will earn an adequate return.