When "AI" and "ERP" are mentioned together, the first thing that comes to mind is usually automation: a form filled in automatically, a repetitive task running without a human. The "agentic" approach increasingly discussed in ERP circles describes something different. It isn't about triggering a pre-defined chain of rules; it's a way of working that understands an intent expressed in natural language, checks the relevant data, prepares a proposal, and leaves the final decision to a person. In this piece, we look at the difference between classic automation and the agentic approach, and how approval works within that model.
What classic ERP automation does — and doesn't do
Classic ERP automation, as the name implies, runs on pre-defined rules: "email the supplier when stock drops below this level," "create the accounting entry automatically once an invoice is approved." These rules are extremely valuable for work that is clear, repetitive, and free of exceptions — they reduce human error and speed up the process.
But the model has a limit: a rule only covers the scenario it was written for. When something unexpected happens — missing information, two conflicting records, an exception nobody defined in advance — the automation stops, and the task lands back on someone's desk. Automation doesn't know "what to do"; it knows how to repeat what it was already told.
What is an AI agent?
The concept of an "agent" describes a different layer. An agent doesn't start from a fixed list of rules; it starts from an intent expressed in natural language. When told, "Create an order for this customer," the agent first understands the intent, then works out which data — customer, product, stock, price, order history — is relevant, checks the related records, and, if it notices something missing or inconsistent, asks a person about it.
This is fundamentally different from classic automation: automation executes a script written in advance, while an agent builds its own steps each time by looking at the data in front of it. If the situation differs from what was expected, the agent pauses, asks, and clarifies — it doesn't fail silently or proceed on a wrong assumption.
The human-approved transaction model
The most critical feature of the agentic approach is that the decision belongs to the person, not the agent. When an agent receives an intent, it doesn't act on it directly; it first prepares a draft, presents that draft as a preview, and only completes the transaction once explicit approval is given. For actions like orders, returns, or stock movements, this approval step is never skipped.
The model can be thought of in three stages: preparation (the agent gathers the relevant data and builds a proposal), preview (the user sees what is about to happen and what it's based on), and approval (the transaction only goes through with the user's explicit decision). The agent prepares and shows its reasoning; the person always makes the call.
The difference between automation and the agentic approach
- Classic automation executes a pre-defined rule; the agentic approach understands an intent and builds its own steps.
- Automation stops and fails on an unexpected situation; an agent asks a question and clarifies.
- In automation, approval is usually noticed after the fact; in the agentic model, approval happens before the transaction, through a preview.
- Automation can stay limited to a single record or table; an agent can evaluate several relevant data areas together.
- In both cases, the final decision and authority remain with the person — the agentic approach changes how preparation happens, not who holds authority.
A worked example: "Create an order for this customer"
A concrete example makes the difference clearer. Under classic automation, such a request would likely be entered by hand, screen by screen; the system only checks the data that was entered and blocks or approves the transaction against pre-defined rules — credit limit, stock sufficiency, and the like. When the request arrives in natural language, classic automation cannot process it directly; a person is still needed to turn that sentence into a transaction.
Under the agentic approach, the same request — "create an order of 100 units for this customer" — is first understood, and then the relevant customer, product, stock, and price records are checked. If stock is insufficient, or if there's a flag in the customer's payment history, the agent notices it and asks the user: should the full quantity ship, or should a partial order be prepared instead? Once the answer is clear, a draft is prepared; the user reviews that draft, and the transaction completes on approval. The difference is that the same request is now prepared by the agent and handed to the person, rather than being entered by the person screen by screen.
Where caution is still needed
The agentic approach is a powerful way of working, but its limits need to stay clear. An agent cannot go beyond a user's existing authorization in the ERP; it cannot see data or perform actions the user themselves couldn't. This isn't an expansion of authority — it's the existing authority being used faster and more accurately.
For critical, hard-to-reverse, or regulation-bound decisions — a large cancellation, a legal notification — the agent's proposal should always pass through human review. The value of the agentic approach isn't in replacing the person; it's in reducing the preparation burden the person carries before making a decision.
In the end, agentic ERP doesn't replace automation — it covers the gap automation leaves behind: unexpected, multi-variable requests expressed in natural language, handled under human approval. You can take a closer look at how this flow runs in daily operation next.