ERP + AI
What an AI agent can actually do inside your ERP
Not a chatbot bolted to the side. The useful version reads live records, proposes an action, and has a named human who approves it.
14 November 2026 · 8 min
Most AI demos in this space are a chat window that can answer questions about documentation. That is a search box with better manners. The version that pays for itself works inside the system of record, on live transactions, and it is far less exciting to watch.
Start where the volume is and the rules are knowable
The workflows worth automating first share three properties: they happen hundreds of times a month, the correct answer is derivable from data you already hold, and a mistake is caught cheaply by the next step in the process.
- Purchase order matching: invoice against PO against goods receipt, with the exceptions routed to a person
- Invoice coding: proposing the GL account and cost centre from vendor, description and history
- Month-end close support: finding the accruals nobody posted, the intercompany mismatches, the journals that look unlike last month's
- Inbound document triage: deciding what a supplier email is and which queue it belongs in
Notice what is not on that list: anything customer-facing, and anything where being wrong is expensive and invisible.
Reading is a different risk from writing
An agent that reads your ERP and proposes something is a productivity tool. An agent that writes to your ERP is a system of record change, and it needs the same scrutiny as any other integration: who did it, under what authority, and how do you reverse it.
In practice the useful pattern is narrow write access through a defined interface rather than general database access. The agent can create a draft journal, not post one. It can propose a match, not clear a payment. The approval stays with a person until the numbers justify moving the line, and the numbers come from your evaluation set rather than from how it feels.
The constraint is almost never the model
If two systems disagree about who a customer is, an agent reading both will produce a confident answer and it will be wrong.
Every failed AI pilot we have been asked to look at failed on data or process, not on the model. Duplicate vendor records. A product master where the same item exists three times with different units. A process that is actually five undocumented processes depending on which regional team is running it.
That is why we treat master data governance as a precondition rather than a follow-on project. It is also why an ERP partner who has never governed master data tends to build agents that demo beautifully and are quietly switched off a quarter later.
How to tell whether it is working
Build the evaluation set before the feature. A few hundred real cases with known correct answers, scored on every change. Without it you cannot distinguish an improvement from a regression, and every subsequent decision — new model, new prompt, new retrieval strategy — is guesswork dressed up as iteration.
The number that matters is not accuracy in isolation. It is the share of cases handled with no human touch, at an error rate the business has agreed it can absorb. Those two move against each other, and deciding where to sit on that curve is a business decision, not a technical one.
A realistic first six months
One workflow. Evaluation harness first. Four to eight weeks to build. Then live but human-approved for a month, watching the numbers, before anyone discusses autonomy. If that sounds slow, compare it with the time already spent on pilots that never reached production.

