ERP AI Integration: How to Integrate AI with SAP, Microsoft Dynamics, NetSuite and Odoo
ERP AI integration means connecting a language model to the system that holds your orders, stock, invoices and production records. People ask questions in plain language, incoming documents are read for them, and replies arrive as drafts. The short answer on AI for ERP: it is a good reader and preparer, and it should not be the one who posts. Start read-only, in one process, with the permissions of the person asking.
The ERP stays the system of record. Its pricing, tax and posting logic is the result of years of configuration, and a model has to call that logic, never imitate it. Below: what the combination is good for, the integration patterns for SAP, Microsoft Dynamics, Oracle NetSuite and Odoo, public versus local models, and a staged rollout.
What AI for ERP is good for in practice
- Questions in plain language over orders, stock and invoices. "Which open orders for this customer are late, and why?" The assistant turns the question into API calls and answers with the document numbers, so it can be checked.
- Document intake. Supplier invoices, purchase orders and delivery notes arrive as PDF or email. A model extracts header and line data, matches it against the purchase order and prepares a draft for a clerk to confirm.
- Drafting replies and reports. A delivery-status reply, a payment reminder, a weekly summary of overdue receivables: written from live ERP data and reviewed by a person.
- Anomaly spotting. Duplicate invoices, prices that deviate from the contract, unusual quantities. Rules and statistics flag them, the model explains and sorts the findings.
- Assistants for quality and maintenance records. Finding similar past non-conformities, summarising the history of a machine, pointing to the right work instruction.
What it is not good for
- Unsupervised postings. A model that books goods receipts or releases payments on its own will sooner or later act on a misread document, and a wrong posting is expensive to reverse.
- Replacing the ERP's own logic. Pricing, tax and availability checks belong to the ERP. If the assistant needs a price, it asks the ERP.
- Arithmetic over large result sets. Totals are computed by the database or the ERP's reporting. The model receives the result and explains it.
How to integrate AI with SAP, Dynamics, NetSuite or Odoo: three patterns
1. A read-only data layer first. SAP publishes its APIs, many of them OData services, in the SAP Business Accelerator Hub. Dynamics 365 Business Central has a documented REST API. NetSuite offers SuiteTalk REST web services including SuiteQL queries, and Odoo has an external API. Where an API is too narrow for analytical questions, database views or a reporting replica do the job, with read-only credentials and only the columns the use case needs.
2. Retrieval over documents. Contracts, specifications, work instructions and quality reports live beside the ERP. They are indexed and searched at question time (RAG), and the answer cites its passage. The access rights of the source documents carry over into the index.
3. Tool calls with a narrow, audited write path. Only when reading works reliably does the assistant get to prepare changes: a few named actions such as "create a draft purchase requisition". Each one goes through the ERP's own API and validation, is logged with user, input and result, and takes effect only after a person approves it. No free-form SQL writes, no generic "update record" tool.
SAP and Microsoft also ship their own assistants inside their products, Joule and Copilot in Dynamics 365. Evaluate them first for tasks that stay inside one product. A separate integration makes sense when questions span several systems, when the vendor assistant is not available for your edition, or when the data has to stay with a model you run yourself. See also how to integrate AI into your company.
Data quality and permissions
The assistant must respect the ERP's role model per user. A sales clerk who may not see purchase prices in the ERP must not get them from the assistant either. So the request runs under the identity of the person asking, through single sign-on, or the integration layer filters by that user's roles before anything reaches the model. One technical user with broad read access, shared by everyone, is how salaries and margins leak inside a company.
The second prerequisite is data the model can interpret: clean master data, descriptions for custom fields and a glossary of your status codes. An assistant on top of duplicate customer records gives fluent answers about the wrong customer.
Public model or local model for ERP data?
ERP data includes prices, margins, customer and supplier details, and often personal data from payroll.
- Hosted models such as Anthropic Claude or OpenAI GPT, called over their API or through your existing cloud provider. This fits when the data may be processed by a provider under a data processing agreement. Send only the fields a question needs. Details: Claude and OpenAI integration and the Claude API integration guide.
- Local models, open-weight models such as Qwen, Llama or Mistral on your own servers. This fits when regulation, contracts or internal policy rule out an external API. You take on hardware and operations, and you check answer quality with your own test questions. Details: private AI with local models and the local LLM on-premise guide.
Many companies use both, decided per data category. The legal side is covered in GDPR and AI integration.
A staged rollout
- Stage 1: one process, read-only. Pick a process with many repeated questions, such as order status or open items. Collect a few dozen real questions with known answers as a test set.
- Stage 2: documents and drafts. Add retrieval and document intake. All output is a draft.
- Stage 3: approvals. Add a few narrow write actions, each with human approval and an audit log.
- Stage 4: more processes. Widen only when accuracy on the test set holds and there are no permission violations.
Cost and operations
Cost has four parts: integration work, model usage (tokens for a hosted model, hardware for a local one), the ERP side and operations. On the ERP side, check API limits and how your licence treats access by external systems.
In operations, every question and tool call is logged. The test set runs again after each model change and each ERP update, because upgrades change fields and APIs. Rate limits keep the assistant from loading the production ERP during month-end closing.
Common mistakes
- Starting with write access.
- One shared technical user with full read rights.
- Letting the model write SQL against production tables without a view layer.
- No test set, so nobody notices when an update makes answers worse.
- Treating text inside incoming documents as instructions. An invoice PDF is untrusted input and must never trigger an action.
Checklist before you start
- One process chosen, with a named owner in the business
- Available interfaces listed: API, OData services, database views
- Read-only credentials; per-user permissions tested with a restricted account
- Hosted or local model decided with the data protection officer
- Test set of real questions with correct answers
- Logging of questions and tool calls; approval step for every write action
- Usage limits, and a test run after every ERP update
Frequently asked questions
What is ERP AI integration?
It connects a language model to the ERP through its APIs, database views and documents, so employees can ask questions in plain language, have documents read and get drafts. The ERP remains the system of record.
Can we integrate AI with SAP or Microsoft Dynamics without modifying the ERP?
In most cases yes. The assistant reads through existing interfaces such as OData services, REST APIs or read-only database views. Inside the ERP you usually only add users, roles and, later, a few approved write interfaces.
Should an AI assistant be allowed to post in the ERP?
Not on its own. It prepares a posting as a draft and a person approves it. Write access is limited to a few named actions that pass the ERP's own validation and are logged.
Do we need a local model for ERP data?
Only if regulation, contracts or internal policy forbid sending the data to an external provider. Otherwise a hosted model under a data processing agreement is often sufficient. Many companies combine both.
Where should we start with AI for ERP?
With one process that produces many repeated questions, for example order status, in read-only mode. Build a test set of real questions, measure accuracy, and add drafts and approvals afterwards.
Bottom line
AI for ERP pays off where people spend their day looking things up, retyping documents and writing the same replies. It becomes a risk where it posts without review. Read-only first, permissions per user, a narrow and logged write path, and a test set: with those four in place, the rest is ordinary engineering.
Planning an ERP AI integration? This is what our AI integration work covers: we connect hosted and local models to ERP, ticketing, documents and databases behind your SSO, with least-privilege access and defined data boundaries. For Harput Holding that meant a family of assistants over ERP, quality-management and business data. Talk to us.
Sources and further reading
- SAP Business Accelerator Hub: catalogue of SAP APIs
- SAP News Center: announcement of the Joule assistant
- Microsoft Learn: agents, Copilot and AI capabilities in Dynamics 365 apps
- Microsoft Learn: API reference for Dynamics 365 Business Central
- Oracle NetSuite: overview of SuiteTalk REST web services
- Odoo documentation: external API