In a live SAP ECC 6 system, an AI assistant answered three purchasing questions from current SAP data in about a minute, then refused a fourth because the user was not allowed to see it. This is session 1 of the SAP without SAP GUI series. It shows what it looks like to ask SAP in plain language.
The first question: which purchase orders are late?
The user asked, in plain words: "Which purchase orders were due in October and still haven't been received? Show PO number, vendor, item, quantity, due date and value."
The assistant returned five late items worth $4,245.00:
| PO | Vendor | Item | Quantity | Due date | Value |
|---|---|---|---|---|---|
| 4500019730 | C.E.B. New York | 100-100 Casings | 10 PC | 1 Oct 2026 | $1,245.00 |
| 4500019732 | Delta Steel Inc., Houston | I-1100 Pump Installation | 10 hrs | 27 Oct 2026 | $500.00 |
| 4500019733 | Delta Steel Inc., Houston | I-1100 Pump Installation | 10 hrs | 27 Oct 2026 | $500.00 |
| 4500019734 | Delta Steel Inc., Houston | I-1100 Pump Installation | 20 hrs | 29 Oct 2026 | $1,000.00 |
| 4500019736 | Delta Steel Inc., Houston | I-1100 Pump Installation | 20 hrs | 29 Oct 2026 | $1,000.00 |
It also added what a buyer would want to know next: the casings were the oldest item, 37 days overdue, and Delta Steel accounted for four of the five, $3,000 in installation services.
In SAP GUI this is a report, a selection screen and an export. Here it was one sentence.
The follow-up: asked the way you would ask a colleague
The second question did not repeat anything: "What else is on order from Delta Steel for this month?"
The assistant kept the context. It knew which vendor was meant and which month, and it listed Delta Steel's other open purchase order items due in November, with quantities, plants, due dates and a total.
This matters more than it looks. A report answers the question it was built for. A conversation lets the user follow the thread, and most real questions in purchasing are the second or third one.
Stock works the same way
The third question moved to inventory: "How many casings, material 100-100, do we have in stock in New York?"
The answer was 2,048, all of it unrestricted. Nothing was blocked, in quality inspection or in transfer.
The user did not say which table holds stock or which plant number New York is. The assistant worked that out from the SAP objects IT had approved for the role.
The refusal: the question the user is not allowed to ask
The last question was deliberate: "Show me the general ledger postings for October."
The request was refused. The demo user is a purchasing user, and general ledger data is not approved for that role. The assistant said so plainly and did not try another route to the data.
This is the part of the session an SAP team should watch most closely. Showing an AI assistant reading SAP is easy. What decides whether it can be allowed near a production system is what happens when it asks for something it should not have.
Two checks stand between the assistant and SAP data, and a call must pass both:
- The role's whitelist. IT lists the tables, reports and functions each role may use. An object that is not on the list cannot be reached or even discovered.
- The user's own SAP authorizations. The call runs as the named user, so SAP applies the same authority checks it applies in SAP GUI.
The refused request is logged in SAP like every other call, with the user's name on it.
The security page describes both checks in full.
What happened underneath
Every answer came from live SAP data, read at the moment of asking. Nothing was extracted to a warehouse or cached. The assistant called CGR MCP Server, which called CGR SAP Connector inside SAP, which read the approved tables under the user's own SAP user.
Filters ran in SAP, so only the matching rows left the system, and row limits stop any single request from pulling bulk data. No BTP, no middleware and no development on the SAP side were involved.
Try it with your own questions
A 30-minute live demo runs questions like these on a real SAP ECC 6 system, including a request the assistant is refused. Bring the questions your team asks SAP every week.
Next in the series: session 2, writing into SAP.