← The journalVISION / EXPLAINER

What “AI-native ERP” should actually mean

Every vendor now claims AI. A working definition of AI-native ERP, and the questions that separate architecture from decoration.

Museum architecture with directional camera-motion blur.VISION / SUNDIAL
Photography: Harrison Lin / Unsplash ↗ · Sundial motion treatment
THE PRACTICAL TAKEAWAY

AI-native ERP means software agents read, propose and act under the same rules, permissions and audit trail as people. A chat box on top of a legacy system is not that.

Bolted on versus designed in

There are two ways to add AI to business software. The first bolts an assistant onto an existing product. It can summarize a screen or draft an email. It cannot reliably do work in the system, because the system was built to be operated only by a person clicking through forms.

The second designs the system so every action a person can take is also a well-defined operation a program can take, with the same validation, permissions and history. An agent creating a purchase order passes exactly the same checks as a buyer creating one. That is what we mean by AI-native.

Clean records come first

A model reasoning over messy data produces confident nonsense. If inventory is wrong, if a customer exists three times under slightly different names, or if invoices hit the wrong accounts, no amount of intelligence on top will fix the answer.

So an AI-native ERP is, first, a rigorous ERP. Double-entry accounting, inventory valuation tied to the general ledger, documents linked from quote to cash: that is what makes an answer trustworthy. Sundial is built on the Frappe and ERPNext engine because that core has done this work for years.

Agents propose, people approve

The most valuable near-term pattern is not autonomy. It is drafting. An agent sees an item fall below its reorder point, checks open purchase orders and recent supplier pricing, and drafts a PO. A buyer reviews it, adjusts a quantity and submits. The agent did the tedious part. The person kept the decision.

For that to hold, drafts must be real documents in the system: clearly marked, attributed to the agent that created them, and subject to the same approval rules as everything else. Approval has to be fast. A slow approval only moves the bottleneck.

Plain-language questions with traceable answers

The second pattern is asking the business a question in plain language. Which customers are past their credit limit? What did we spend on freight last quarter, by carrier? Which work orders are behind schedule? The value is not the conversation. It is skipping the hunt for the right report and filter.

Every answer should show its sources: the records, the filters, the period. If a number cannot be traced to documents a person can open, it should not drive a decision. And a question should never return data the person asking is not permitted to see.

Permissions and history are the product

An agent is only as safe as the permissions it runs under. Give each agent its own identity and role, exactly as you would a new hire. A purchasing agent can read inventory and supplier records and create draft purchase orders. Nothing else. It should never inherit the rights of whoever switched it on.

Every action it takes belongs in the same audit trail as a person's, under its own name. When a draft is wrong, you need to see what it read and when, so you fix the rule or the data, not the symptom.

How to test a vendor's claim

Ask what the AI can do that a user cannot. Be wary if the answer is anything. Ask whether an agent's actions appear in the audit trail under its own identity. Ask whether you can restrict an agent to drafting only. Ask what happens when the model is wrong, and how you would find out.

Then ask whether each AI feature ships today or sits on a roadmap. Both are legitimate. Only one belongs in your business case.

Where Sundial stands

Sundial's modules, records, role-based permissions, audit trail and redesigned workspace are in the product today. The AI capabilities described here, plain-language questions and agents that draft documents for approval, are how Sundial is designed and what we are building. Anything in preview is labeled as preview.

We think that order is right. Get the records, permissions and audit trail correct first. Then let software help your team move through them faster. To see the foundation for yourself, try the demo.