What is AI ERP?

Software initiates, person governs

A bill arrives; an agent codes it and routes it.

The work starts without a person. Requires an authority model, an audit trail, and a policy engine — which is why it cannot be retrofitted onto a screen-driven system.

The distinction that matters is not how clever the model is. It is who is responsible for noticing the work exists.

The six requirements

All six, or it is a copilot.

These are not features. They are the conditions under which software initiating financial work is safe enough to allow, and a product missing any of them should not be trusted with it.

Software initiates

Work begins without a person opening a screen. A bill arrives, a bank feed updates, a contract is signed — and something picks it up. This is the load-bearing criterion.

An explicit authority model

What the software may do, at what threshold, in what scope — enforced in code rather than described in a prompt. Without this, initiation is a liability rather than a capability.

A deterministic execution layer

The model proposes; something that is not a model decides whether it posts. A ledger needs a guarantee, and models produce likelihoods.

An audit trail built for it

Model version, context, source records, reasoning, confidence, policy applied, approver, and resulting entry — in the same schema as human actions.

One data model

An agent reasoning about a bill needs the contract, the PO, the budget, and the payment history. Six systems joined by a nightly sync cannot supply that.

Measured accuracy

Straight-through rate and confidently-wrong rate, per workflow, against your own corrections, with regression gates. Otherwise "it works well" is an assertion.

The short answer

Traditional ERP is a system of record. It stores what people did. A person opens a screen, enters a transaction, and the system remembers it accurately and permanently. That is genuinely valuable and it is the entire design.

AI ERP is a system of record and a system of action. It still stores what happened, with the same rigour — but a meaningful share of what happened was started by software: a bill read and coded, a reconciliation performed, a close task completed, a chase drafted. People move from performing the work to governing it and handling the exceptions.

The question is not how clever the model is. It is who is responsible for noticing the work exists.

What AI ERP is not

Not a chat panel. A conversational interface over reports is useful and entirely person-initiated. Nothing happens unless somebody opens it and types, which means it has not changed who does the work.
Not better autocomplete. Suggesting the account you probably want speeds up data entry. It does not remove it.
Not automation in the old sense. Rules engines have existed in ERP for thirty years. If X then Y is deterministic and brittle; it fails on the first document that does not match the pattern, which is why nobody claims their rules engine is AI.
Not autonomous finance. Nobody should want software closing periods, releasing funds, or signing assertions. Those need a person for reasons that are governance rather than technology, and a vendor promising them is describing a liability.

Why it cannot easily be retrofitted

Every incumbent will ship credible AI features, and some will be better than ours in specific places. What is hard to add later is the architecture underneath: a permission model describing what software may do rather than what people may do, an audit schema recording reasoning and confidence beside the entry, and a policy engine positioned between intention and execution.

Those are not features you add in a release. They are decisions about where the execution path runs, and a system designed around human-initiated screens has to be substantially rebuilt to accommodate them. That is the actual content of the phrase AI-native, as distinct from AI-enabled.

Our interest, disclosed

We sell an AI ERP platform, so this definition is not disinterested — it is a definition under which our product qualifies and several competitors do not. Judge it on whether the six requirements are the right ones, and note that it also excludes the autonomous finance that would be commercially convenient for us to promise.

Questions

Common follow-ups.

Is AI ERP just marketing?

The phrase is heavily overused, which is why a definition that excludes things is worth having. If a product cannot initiate work, cannot express an authority model, and cannot show you an audit record with model version and policy on it, the phrase is doing no work.

Do we need AI ERP?

Not necessarily. If your close works, you have one entity, and the pain is a reporting gap, you have a reporting problem. Agentic automation earns its place where volume of routine financial work is genuinely consuming people.

Is it safe to let software post transactions?

Under an authority model with a deterministic execution layer, yes, and that is the entire content of the question. Without them, no — and the difference is architectural rather than a matter of model quality.

Will incumbents catch up?

On features, quickly. On architecture, slowly, because a system built around human-initiated screens needs substantial rebuilding to put a policy engine between intention and execution. We would not bet on that gap lasting forever.

What about autonomous finance?

We do not build it and think nobody should. Closing a period and releasing funds are assertions that need a person behind them, and automating them removes the meaning rather than the work.