Photo, audio, work order opened in Protheus. 90 seconds.

It’s 2:31 PM at a factory in Jaraguá do Sul.

The operator on line L2 photographs a defective part that just came off the press and sends a short audio message on WhatsApp:

“This defect showed up on production order 4471, problem with the bearing.”

2:32 PM. Kris replies:

“Identified: PO 4471 · line L2 · assembly sector. Symptom: SKF 6204 series bearing. Should I open a corrective maintenance work order in Protheus and notify Cassio, the shift mechanic?”

2:32 PM. Operator: “Confirmed.”

2:33 PM. Kris replies:

“Corrective maintenance work order #8839 opened in Protheus. Cassio notified. I’ll let you know when he responds.”

Audit log · record #45219.

Ninety seconds. Photo, audio, work order opened inside the existing ERP, shift mechanic notified, everything logged in a permanent audit trail.

Before that “opened in Protheus,” what happened is what still happens in nearly every plant in the region: the operator wrote it in a notebook, told the supervisor at lunch, the supervisor opened the work order manually in the afternoon — when he remembered. Between “I saw the defect” and “work order opened,” hours passed. Sometimes the entire production order ran through before the work order was ever created.

Why this isn’t a chatbot

The difference between what Kris does and what a chatbot does is categorical — but it’s easy to confuse, because the input channel looks the same.

A chatbot reads text and responds with text. The cycle ends inside the channel.

Kris reads the channel — text, audio, photo — and acts inside the ERP. Opens a real work order in Protheus, with a real work order number, assigned to a real mechanic, tied to a real production order. The cycle ends inside the actual operation, not inside the messaging app.

Three things have to work for this to function:

• Understand the request in any format. Photo of the part + audio in colloquial Portuguese + production order number mentioned out loud. Not isolated OCR, not isolated speech-to-text — it’s joint interpretation with the context of the operation (shift, line, equipment series).
• Know the house. “PO 4471 is running on L2” is not data the chatbot has — it’s data that lives in the MES, in the shift production queue, in the equipment history. Kris sees it.
• Act with permission. Can the L2 operator open a corrective maintenance work order without supervisor approval? That depends on the house’s RBAC. Kris inherits the permission of whoever sent the audio — it doesn’t act beyond that.

That’s the difference between “AI that converses” and “AI that does the work inside the ERP.” Kris is the second category.

The channel doesn’t matter. The engine does.

One plant uses WhatsApp because the operator already has WhatsApp. Another uses Teams because IT standardized on Teams. A third uses Slack because it’s a multinational industrial group. A fourth uses a tablet on the bench because the shop floor doesn’t allow personal cell phones.

In all four cases, the AI engine is the same. The only difference is the channel adapter — and a channel adapter is the easiest problem for a platform of this kind.

What matters isn’t how the message arrives. It’s:

• What the platform knows about your operation (data model, MES, ERP, catalog, history).
• What the codified house judgment is (standard procedure, approval policy, shift escalation).
• What the real integration is with the target system where the action happens (not “an API Protheus exposes,” but “a work order opened with the right fields in the right module, assigned to the right resource”).

When those three are in place, the channel becomes plumbing. WhatsApp today, Teams tomorrow, natural voice after that — the execution cycle doesn’t change.

Permission as design, not an option

Here is the difference that separates AI acting inside operations from “a copilot that suggests things.”

Kris only executes within the permissions of the user who triggered it. It doesn’t inherit admin permission. It has no permission of its own. If the L2 operator cannot open a corrective maintenance work order without supervisor approval, Kris doesn’t open it either — it escalates, asks, waits for the supervisor’s “OK” to come through their channel.

And every confirmation is recorded — not as a debug log, but as a compliance-grade audit trail:

• Who sent the audio.
• What action was suggested.
• Who confirmed it.
• What permission the person who confirmed it had.
• What action was actually executed, in which system, with which ID.

That record is what separates AI applied to industrial operations from a demo experiment. Without it, any action Kris takes becomes a liability on the first day of an audit.

What changes in operations

Three direct effects:

• Latency drops to minutes. The “I saw the defect → work order opened” cycle goes from hours to 90 seconds. In corrective maintenance, that’s direct margin — a stopped machine costs by the minute.
• Informal coordination recedes. Notebook, lunch, “I’ll tell you later” — all the word-of-mouth knowledge traffic that depended on physical presence becomes a digital protocol with a trail.
• The operator gains leverage. Instead of waiting for the supervisor, they trigger the system directly — within what their permission allows. When approval is needed, the request reaches the supervisor already with full context ready.

None of these gains require replacing Protheus, SAP Business One, Sankhya, or Senior. They only require that the real action, inside the ERP the plant already uses, can be triggered by an AI layer with the right judgment, context, and permission.

---

The team behind F7 KORE has been automating industrial processes for over a decade — including ERP integration in cases like Degraus (equipment rental) and Quinabra (regulated chemicals). Details on how it works on the FAQ page.

If you operate a mid-to-large industrial business, have an established ERP, and want to see “Photo + audio → work order opened in 90s” running against your system — schedule a 30-min conversation.