AI for ERP: how AI works with the ERP you already run
AI for ERP answers questions from the ERP that you already run and drafts entries for a person to approve.

What AI for ERP means
An ERP, short for enterprise resource planning, is the business system that holds a company’s orders, inventory, purchasing, production, and accounting. You can keep the ERP as the system of record, which holds the official version of each record, and add AI on top. The AI layer reads the ERP data, answers questions about that data, and prepares work inside the ERP. The table shows the four jobs that the layer does and who signs off on each job.
| Job | Example | Who signs off |
|---|---|---|
| Answer a question | “Which open orders for this customer will ship late, and why?” answered with the order numbers | No one, because the answer names the source records for anyone to check |
| Flag an exception | A supplier’s price rose on the last three invoices, or an item’s on-hand quantity went negative | The person who owns that supplier, item, or account |
| Draft an entry | A sales order from a customer’s emailed purchase order, a purchase order from a reorder suggestion, or a supplier bill from an invoice | A named person approves the draft before the entry is saved or sent |
| Operate the screens | An agent keys an approved order into an older ERP the way a clerk would | The same approval, given before the agent types |
One rule keeps AI for ERP safe. The AI sees only what the person asking may see, and the AI prepares only drafts. A named person approves anything that changes a record or leaves the building.
Answering questions with the source record
The right AI tool for ERP questions reads the ERP with your own permissions and names the source record in every answer. Many questions that people ask their ERP are lookups spread across several screens. Examples include what a customer ordered last year, which purchase orders are late, and what a part cost the last three times. An AI system answers such questions in one of two ways. One way is to write a database query from the question and run the query. The other is retrieval-augmented generation (RAG). AWS describes RAG as having a model reference “an authoritative knowledge base outside of its training data sources” before the model answers. AWS adds that RAG does this “all without the need to retrain the model.”
Either way, an answer about your business should name the source records, such as the order number, the invoice, and the receipt date. A number with no source cannot be checked. An assistant that reads the ERP through an administrator’s account can show anyone who asks the margin on every customer. So the answer should also respect who is asking. Microsoft’s MCP server for Business Central performs all operations “with your user identity and permissions, ensuring audit trails show who performed each action.”
An ERP AI chatbot built into the ERP answers from the ERP data only. A plant’s question often spans more, such as the ERP export, the customer’s email, and the drawing attached to that email. A data warehouse is one database that gathers data from several systems. A chat that answers from a warehouse holding all three sources can name the record behind every number, whatever system the record came from.
Drafting entries for a person to approve
Drafting targets the documents that people retype. Customer purchase orders are keyed as sales orders, supplier invoices as bills, and reorder lists as purchase orders. A customer set up for EDI (electronic data interchange) sends each purchase order as an X12 850 instead, which needs no retyping. AI reads the document, finds the customer, supplier, and items in the ERP, and fills in a draft.
The vendors that ship drafting build in the same approval. Microsoft’s Sales Order Agent “always involves designated Business Central users to review and approve all outgoing messages before it sends them to customers,” Microsoft says. The agent always creates a sales quote first, because quotes “have no impact on planning, reservations, availability calculations, or cash flow forecasts.”
Microsoft’s Payables Agent “creates drafts only and never automatically posts invoices.”
Four ways to connect AI to an ERP
How AI reaches your ERP decides what the AI can do and how much can go wrong. There are four routes, and many companies use two of them.
| Route | How the route works | Good for | Watch for |
|---|---|---|---|
| API or MCP server | An API (application programming interface) is the ERP’s own connection for other software. An MCP server is a connection built on the Model Context Protocol, an open standard for AI tools | Reading and writing with the ERP’s business rules applied | Usually needs a cloud or recent version of the ERP. Check which records and actions are exposed |
| Read-only database connection | A reporting account reads the ERP’s database tables directly | Fast answers and reports on an ERP installed on your own server | Skips the ERP’s business rules, so keep the connection read-only. Ask the vendor or your IT support first |
| Scheduled exports and imports | The ERP writes reports to files on a schedule, and approved entries go back through the ERP’s import tool | Older systems with good reports | Answers are only as fresh as the last export |
| Screen agent | Software reads the screen and uses the keyboard and mouse like a clerk | Older ERPs with no API, for keying approved entries | Slower than an API, and a person approves what the agent types |
The Model Context Protocol (MCP) is “an open standard that defines how AI applications communicate with data sources and tools,” in Microsoft’s words. MCP matters because one connector can serve several AI tools. Microsoft’s documentation names Copilot Studio, GitHub Copilot, Claude, and ChatGPT as clients for the Business Central server. By default, the server “provides read-only access to all exposed Business Central API pages.” Write access needs a deliberate grant, page by page. Keep the same default on any ERP connection: read-only until a specific write is needed.
What the major ERP vendors ship
Microsoft Dynamics 365 Business Central, Oracle NetSuite, SAP, Sage X3, Epicor Kinetic, Infor CloudSuite, and Acumatica all build AI assistants or agents into their products. The descriptions in the table are the vendors’ own, checked September 28, 2026. Features change with each release, so check your version.
| ERP | What the vendor ships | In the vendor’s words |
|---|---|---|
| Microsoft Dynamics 365 Business Central | Sales Order Agent, Payables Agent, and an MCP server | The Sales Order Agent “uses AI to analyze customer requests received via email and locates the customer in Business Central” |
| Oracle NetSuite | AI Connector Service and Bill Capture | “NetSuite has adopted the Model Context Protocol (MCP), which boosts its integration capabilities and enables secure interactions between AI models and data systems” |
| SAP | Joule Assistants | “Joule Assistants use deep role and process context to coordinate AI agents and autonomously execute complex workflows across your business” |
| Sage X3 | Sage Copilot for Sage X3 | Starts with helping sales teams “query customer orders, shipping timelines, and fulfilment issues,” with questions asked in English, French, or Spanish |
| Epicor Kinetic | Epicor Prism | “Engage in natural language conversations with Epicor Prism to gather information, request insights, or execute actions” |
| Infor CloudSuite | Infor GenAI assistant | “a conversational assistant enabled with industry-specific AI agents,” embedded in Infor CloudSuites |
| Acumatica | AI Assistant and AP (accounts payable) document recognition | “The Acumatica AI Assistant helps employees ask questions about business data in plain language” |
Three questions apply to every vendor. First, how is the AI billed? Microsoft bills selected Business Central agents per use in Copilot Credits, through prepaid capacity or pay-as-you-go. Microsoft’s billing documentation explains the model. Ask every vendor for the billing model, and check the vendor’s own pricing page. Second, which records and actions can the AI reach, and under whose permissions? Third, where are prompts processed? Microsoft’s Payables Agent FAQ covers environments in regions where Azure OpenAI is not available. In those environments, “administrators must allow data to move across geographies.”
Which route fits your ERP
Tell us which ERP you run, which version, and where the ERP is hosted. We will tell you which of the four routes the ERP supports and what AI can do through each route.
What to do when your ERP has no API
If your ERP was installed on your own server years ago and has no modern API, the vendor’s cloud agents are usually out of reach. AI can still work with that ERP.
- Ask the vendor or reseller what exists. Some older systems have an add-on API, an ODBC driver, or an import tool that nobody at the company has used. An ODBC driver is a standard way for reporting tools to read a database.
- Read from a copy. A read-only database connection, or a nightly export of the reports that people already use, is enough for questions, answers, and exception flags.
- Write back through the screens, with approval. A computer use agent, software that works the screen, keyboard, and mouse like a person, can key approved entries. Anthropic describes the computer use tool as providing “screenshot capabilities and mouse/keyboard control for autonomous desktop interaction,” and recommends precautions. One precaution is “Asking a human to confirm decisions that might result in meaningful real-world consequences.”
- Keep the ERP as the record. Whatever the route, the ERP stays the system of record. The AI layer logs what the AI read and drafted, and who approved each draft.
Running AI next to an old ERP also records the questions that people ask and the documents that they retype. That record is a useful list of requirements if you later replace the ERP.
Where to start
Pick one job with a clear before and after. Two kinds of job make good first projects. One is a question that someone answers by hand every day, such as late orders or stock for a customer. The other is a document that someone retypes every day, such as customer purchase orders or supplier invoices. Measure the time and the error rate for a month by hand. Then measure again with the AI drafting and a person approving.
How ThriveAI helps
ThriveAI is an AI engineering company in Ottawa that builds private AI systems for manufacturers in Canada and the United States. Each system runs on the client’s own data. ThriveAI works with the ERP that a company already runs, through whichever route that ERP supports, including screen agents for older systems. Each answer names the source record, and drafts go to a named person for approval before anything is saved or sent. Each client has a copy of the data reserved for the client’s team. ThriveAI hosts that copy, or the client hosts the copy on its own servers. The client chooses a model where the copy is hosted or a hosted model under a written zero data retention agreement. Under that contract, the provider keeps no copy of a request or of the answer. A hosted model may process requests in another country. Derik Lawlis, the founder, leads every project and stays close to the build. For more information, see About ThriveAI.