Can AI safely act in your ERP?
Five controls that make AI writes defensible in an ERP.
The short answer
Yes, when it acts as the person asking, inside the ERP’s own rules.
What is the risk when AI writes?
A wrong write that nobody reviewed. A bad answer can be checked. A bad write lands in the ERP and flows into planning, cost and compliance.
So is the model the problem?
Not on its own. Any model can be wrong, and so can a person. The question is whether a wrong action can reach IBM Maximo, IFS or SAP without a person seeing it.
Then what decides the risk?
The design around the model. Whether a wrong action can reach the ERP unseen is decided by identity, permissions, approvals and the record kept, which is why the controls below matter more than which model you pick.
Which controls make AI writes safe?
Each one covers a different way a write can go wrong.
-
It acts as you
Your own ERP permissions, never a shared service account.
-
The ERP decides
The ERP enforces access, so no prompt can widen it.
-
You approve writes
The writes you choose wait for a person’s yes.
-
Edits do not collide
A record changed meanwhile is caught, not overwritten.
-
Every step is logged
Who asked, what was planned, what ran, what came back.
Which work orders at site 100 are overdue? Raise a fault report for the worst one.
- 1 Read the overdue work orders at site 1007 overdue, 4 on pump P-114 IBM Maximo
- 2 Read the repair cost for P-1144,200 against the Q4 maintenance budget SAP
- 3 Draft the fault report for P-114Waiting for your yes IBM Maximo
What does approval look like?
The assistant builds the whole change, shows what it will write, and waits. The person approves, or changes a value first, which is quicker than asking again.
What goes in the audit trail?
What NgageChat records for every turn of a conversation.
| Field | What it records |
|---|---|
conversation_id, message_index | Which conversation, and where in it |
role, content | Who spoke, and what was said |
tool_invocations | Every step it ran, its inputs and what came back |
provider, model | Which model handled the turn |
has_error, error_summary | Whether a step failed, and how |
token_usage, duration_ms, created_at | Cost, time taken and when |
More questions from ERP teams
Does write access have to be all or nothing?
No. What each person can write is already set by their ERP permissions, which your administrators maintain today. The question becomes whether you trust a named group of people to do what they are already allowed to do, with an approval in front of the writes you choose.
Can we restrict what the AI may write?
Yes. Narrow a person’s ERP permissions and you have narrowed the assistant. You also choose which writes wait for an approval, and many rollouts start with questions only, then turn on one kind of write at a time.
What if it gets a value wrong?
Where a write waits for approval, you see every value before it is saved, including the defaults the ERP filled in, and you can correct one in the conversation instead of starting again.
Does the ERP still record who made the change?
Yes. Every request runs under the signed-in person’s own ERP account, so the ERP records the change against that person, as it does for work done in its own screens.
Is it the same on IBM Maximo, IFS and SAP?
The controls are the same: each person’s own permissions, approvals on the writes you choose, and every step on the audit trail. NgageChat connects through each ERP’s official interfaces.
See NgageChat answer a question from your ERP
Send us a task you have not prepared for. We run it with you, live, on a test system.