It is ten in the morning and the customer service person at a forty-person machine shop has three questions waiting. A customer wants to know whether their brackets ship Friday. The owner wants to know what went out last month to the shop's biggest account. Purchasing wants to know what is short for next week's work orders. Every answer is somewhere in the system. Getting it means opening the customer order, the manufacturing order and the stock screen, filtering and comparing, and for the Friday question, walking out to the floor anyway.
One question, three different jobs
"Where is this order?" sounds like one question. It is three.
The first is a lookup: what does the system say? Which manufacturing order, which operation it is sitting at, whether the material is allocated. That is clerical, and it is where most of the clicking goes.
The second is a check: is the system right? The routing says the parts are at deburr, but the operator finished them an hour ago and has not booked it. Only someone who knows the floor, or goes and looks, can answer that.
The third is a promise: will it actually ship Friday? That is judgment. It depends on whether second shift is fully staffed, whether the plater turns the batch on time, and how much this customer matters. It is the part the customer is really asking, and it comes last, after the lookup has eaten the time.
As of September 2026, in at least one small-manufacturer ERP, software can do the first part. It cannot do the other two, and that line is the useful part of the story.
What MRPeasy says it added
In a blog post dated September 28, 2026, MRPeasy, which sells manufacturing ERP software, says its built-in assistant, Mr. Peasy, can now be given read-only access to an account's operational data. Until then the assistant answered from MRPeasy's knowledge base, explaining how features work. With access switched on, an employee asks in plain language and gets an answer drawn from the account's records, including reports, tables and charts; the tables can be exported as CSV or PDF. MRPeasy says the assistant's querying is comparable to what its API already offered, so the data was reachable before. MRPeasy's own changelog for May 2026 also lists a connection that lets outside AI tools read account data through version 2 of its API, then in beta. What changed is that the assistant built into MRPeasy can now answer, for someone who has never set up either.
The limits are in the vendor's own description. Access is switched off by default, and an account administrator has to turn it on. The assistant follows each user's existing permissions, so it cannot show anyone what they could not already see. It is strictly read-only: it cannot create, edit, delete or approve anything, and a mode that can make changes is only planned. It runs on Google Gemini; MRPeasy says the names of customers, suppliers, products, employees and similar records are swapped for random identifiers before a request leaves MRPeasy, and that, under Google's Gemini API terms, prompts and responses are not used to train Google's models.
Only MRPeasy's own account of this was available; nothing here has been independently confirmed. As of early October 2026, MRPeasy's user manual says the assistant is "available to all users", with data access off by default; it carries no beta or preview label.
Who will push back, and why they are right
The production scheduler, or whoever in your shop is the person everyone asks. They will keep walking the floor, and they are right to, because the system only knows what was booked. In a shop where operators book operations at the end of the shift, or not at all, an assistant reading the system answers the first question fast and the second one wrong, confidently, in a clean table. The scheduler knows that, and will quietly keep a whiteboard more current than the ERP.
The other holdout is the owner who does not want shop data going to an outside AI service. Swapping names for codes is the vendor's answer to that. Whether it is enough is your call, and the setting stays off until someone makes it.
The honest no: if work orders are not booked as the work happens, do not turn this on yet. Asking a stale system faster does not make it right. Fix the booking discipline first, then see how many questions are left. And if one person fields a handful of these a day and knows the answers cold, the clicking is not your bottleneck.
If this is your Thursday afternoon, the ten agent roles are in the Library.