To begin an AI ERP integration safely, start with a single process, in read-only mode, through the ERP’s official interfaces, using the user’s own permissions and logging every access. The AI prepares; a person approves. Only open up write access to the ERP once quality has been measured on real cases. The ERP remains the source of truth.
What AI does well with an ERP, and what to keep it away from
| Use case | Reliability today | Condition |
|---|---|---|
| Answering staff questions about orders, inventory, invoices | Good | Each answer cites the ERP record it comes from |
| Extracting data from a document (order, invoice, delivery note) | Good | Human check before data entry |
| Preparing a draft (order, reminder, customer reply) | Good | Approval by a person |
| Matching documents and flagging discrepancies | Good | Explicit business rules |
| Calculating forecasts | Best left to statistical models | Generative AI can explain a forecast, not calculate it |
| Changing critical data on its own | Avoid at first | Only after a controlled increase in autonomy |
Three levels of autonomy, in order
- Read-only. The AI queries the ERP to answer or analyze. No risk of corrupting data. This is the right starting point.
- Guided actions. The AI prepares an action (an order, an entry, an update) and a person approves it before it is saved in the ERP.
- Supervised autonomous actions. For simple, well-measured cases, the AI acts on its own within set limits, with full logging and after-the-fact review.
The move from one level to the next is based on the numbers: error rate, share of suggestions accepted without changes, incidents.
AI ERP integration: three ways to connect
1. Through the ERP’s APIs
The recommended approach when the ERP exposes documented APIs. The AI calls specific, limited functions (read an order, create a draft) without direct access to the database.
2. Through an integration layer
A middleware application (built on Mendix, for example) or an integration platform exposes a small number of authorized tools to the AI and handles permissions, errors and audit trails. This is useful when several systems are involved (ERP, PLM, CRM).
3. Through a connector or middleware for legacy systems
When the ERP has no modern API, middleware or a dedicated connector bridges the gap. It takes more work, but it avoids touching the core of the ERP.
In every case, the AI should never have direct, broad access to the ERP database.
Security: the rules to set from day one
- The AI inherits the permissions of the user working with it, never broader ones. Tasks that run without a user get a dedicated, limited identity.
- Every access and every action is logged.
- Sensitive data (payroll, health) stays out of scope until it is actually needed.
- No model training on your ERP data: the AI retrieves the information when the question is asked.
- Watch out for hidden instructions in the documents and emails the AI processes: its limited permissions are the best protection.
- Hosting: check where data flows and is stored, and review the contract terms with the model provider (GDPR).
How to run the pilot
- Choose a process that is frequent, measurable and low-risk.
- Clean up the relevant master data (items, customers, suppliers).
- Test on a copy of the ERP before any access to production.
- Run the AI in parallel with the current process, over a full cycle.
- Set the success threshold in advance, and decide what happens if it is not met.
Our experience at Tocndix
Tocndix connects business applications to ERPs (SAP, Priority) and to Siemens Teamcenter. For example, we linked the Priority ERP to Siemens Teamcenter (see our 3 options for connecting Teamcenter to an ERP). Our engineers work embedded in your teams until go-live, and every project starts with a scoping phase. To choose a provider for this kind of project, see our scorecard for comparing AI integration providers.
Our ERP / PLM connection solution · AI integration · Talk to us about your ERP
FAQ
Do you need to train an AI model on your ERP data?
Almost never. The AI retrieves information from the ERP when the question is asked, with the user’s permissions. It is safer, cheaper and always up to date.
Our ERP is old and has no API. Is it still possible?
Yes, with a connector or an integration layer that exposes the functions you need. The scoping phase lets you estimate the effort.
How long does a first use case take?
It mostly depends on data access and the number of systems involved. A first read-only use case on a single process is the fastest to set up.
Can AI enter orders directly into the ERP?
Yes, but not at first. Start with drafts approved by a person, then open up write access for simple cases once quality has been measured.