The Role of Generative AI in Modern ERP Systems
There is a version of the generative-AI-in-ERP conversation that is mostly theatre: a chat box bolted onto a transaction screen, demonstrated on clean data, in a sandbox, to an audience who will never use it. It demos well and changes nothing.
There is also a version that is quietly removing real work from real finance and supply chain teams. The difference between them is not the model. It is what the model is pointed at.
Where it genuinely works
The strongest results we see share a shape: a high-volume, judgement-light decision that a human currently makes from structured data, where being right 95% of the time and escalating the rest is an acceptable outcome.
- Invoice and PO matching where line items do not align cleanly and a human currently reads both documents to decide.
- Incident triage — classifying and routing SAP support tickets, which is tedious, high-volume and consistently done badly by rules engines.
- Master data creation, where a new vendor or material record must be assembled from a supplier email and an existing taxonomy.
- Drafting — variance commentary, audit narratives, specification documents. Not the analysis, the writing up of it.
Note what these have in common. None of them asks the model to be the system of record. Each asks it to close a gap between two things the system already knows.
Where it does not
Anything that must be exactly right every time, anything where the reasoning must be reconstructable in an audit two years later, and anything where the training signal is thin. Generative models are probabilistic. An ERP is, by design, the place where the organisation keeps its non-probabilistic truth. Confusing the two is how you end up explaining an AI system to a regulator.
The unglamorous constraint
The limiting factor is almost never model capability. It is master data. A retrieval system over a vendor master with four spellings of the same company, inconsistent material classifications and half-migrated org structures will produce confident, wrong answers — and it will produce them faster than a human could produce the right one.
AI does not fix bad master data. It industrialises the consequences of it.
This is why our AI engagements frequently start somewhere that sounds unrelated: data governance. Catalogs, lineage and quality frameworks are not an AI project. They are the thing that decides whether the AI project works.
How to start without wasting a year
Pick one process. Measure how often it currently needs a human. Build the narrowest possible intervention inside your existing authorisation model on SAP BTP, so the AI inherits the permissions the user already has. Then measure the same number again. If it has not moved, you have learned something cheaply. If it has, you have a defensible case for the next one.
That measured-rate approach is the whole of our enterprise AI practice. It is less exciting than a demo and considerably more likely to survive contact with your month-end.