ERP-native agents should start with exception-heavy work

The gap around the ERP is often more valuable than the happy path inside it.

Core enterprise systems are good at deterministic transactions. They are less good at the messy work surrounding exceptions: missing documents, mismatched quantities, ambiguous supplier emails, policy interpretation, cross-system investigation and coordination between people.

SAP’s 2026 Agentic Innovation Tournament in the Americas focused participant projects on real business problems including supplier management and logistics. That is telling. These areas contain exactly the kind of semi-structured exception work where agents can add value without replacing the ERP.

For many enterprises, this is the right starting point.

Do not automate the clean path first

The happy path already works:

purchase order → goods receipt → invoice → payment

It may be highly automated and well controlled.

Value leaks in the deviations:

• invoice amount differs from PO;
• goods receipt is missing;
• supplier changed bank details;
• delivery was partial;
• contract terms conflict with invoice;
• approval owner is absent;
• item code does not match;
• shipment status disagrees across systems.

Humans investigate these cases by opening ERP screens, email, documents and spreadsheets.

That is agent territory.

The ERP remains system of record

A strong design does not ask a model to recreate ERP logic.

Instead:

1. ERP emits an exception or queue item.
2. Agent receives the bounded case.
3. Agent retrieves related evidence from approved systems.
4. Deterministic checks run where possible.
5. Model reasoning handles ambiguity.
6. Agent proposes or performs an allowed resolution.
7. High-risk changes receive approval.
8. ERP records the final transaction.
9. Trace stores evidence and decision path.

The ERP keeps transactional truth. The agent becomes an investigation and coordination layer.

Exceptions have natural evaluation labels

Exception workflows are attractive because success can often be measured objectively.

Examples:

• resolved without reopen;
• correct matching;
• correct routing;
• cycle time;
• touch time;
• financial recovery;
• escalation rate;
• policy compliance;
• number of systems/handoffs avoided.

Historical exception queues can also provide an evaluation corpus.

The organisation can replay past cases against the agent before granting write access.

This is much stronger than evaluating generic assistant quality.

Start read-heavy, then earn writes

A sensible autonomy progression is:

read → investigate → recommend → prepare transaction → approved write → bounded autonomous write

Each stage creates evidence for the next.

For supplier master data, a bank-account change may always require human approval. For low-value invoice coding, the agent may eventually write autonomously after sufficient evaluation.

Autonomy should follow risk and evidence, not enthusiasm.

The integration layer should expose business capabilities

Do not give the agent a raw database or an unrestricted ERP super-user account.

Expose capabilities such as:

• get_purchase_order
• get_goods_receipt
• get_supplier_contract
• create_invoice_exception_note
• prepare_supplier_change
• submit_for_approval

Each capability can carry permissions, validation and audit semantics.

This is the same principle behind MCP-style capability layers: make business actions explicit and governable.

Exception work is ideal for human-agent collaboration

Humans are expensive at repetitive cross-system searching and excellent at novel judgement.

Agents are strong at:

• gathering context;
• comparing records;
• checking policy;
• drafting communication;
• monitoring follow-up;
• maintaining state.

A good exception system reserves human attention for the few cases where judgement or authority matters.

The goal is not “zero humans”. It is fewer low-value hand-offs.

The Lexdatum view

ERP-native agents should begin where deterministic systems stop and operational ambiguity begins.

Procurement, logistics and finance exception queues provide bounded work, clear records, measurable outcomes and natural human escalation.

That makes them unusually good places to learn enterprise agent operations safely.

Keep the ERP. Keep the controls. Add an intelligent layer around the work the ERP was never designed to resolve.

Sources

• SAP — SAP Innovation Tournament Americas 2026 mostró cómo la IA agéntica está transformando los negocios en México (2026-08-25)

---

Research note: This analysis is based on public information available on 1 September 2026. Product capabilities, prices and regulatory interpretations can change. Vendor-reported claims are identified as such. This is not legal, security, financial or investment advice.

Discuss this architecture with Lexdatum