SAP

Contribution to AI

A machine can only reason about a business it can read. The systems built at Walldorf from the mid-1970s onward imposed one schema on functions that had previously lived in incompatible ledgers and departmental files — accounting, payroll, materials, orders — and when R/3 spread through the world's large corporations after 1992, the practical result was that firms rearranged their own procedures to match the software rather than the other way round. A purchase order, a cost centre and a material master came to mean approximately the same thing in Osaka as in São Paulo. That sounds like an administrative fact rather than an AI one, but it produced the largest body of structured, semantically consistent records of industrial activity anywhere, and later work depends on that consistency. Process mining is the clearest instance: the technique works only because such systems timestamp every step of a transaction in a predictable format, which is what Celonis was founded in Munich in 2011 to read. The database side mattered too. A 2009 SIGMOD paper from the company argued that transactions and analytics should share a single in-memory column store instead of copying data between them, and HANA, shipped from 2010, made querying a live business a matter of seconds rather than an overnight batch. Grounded enterprise AI now assumes such a layer exists.

Drafted with AI and edited by hand (claude-opus-5, reviewed 2026-08).