# How to use ChatGPT with Infor M3 > Source: https://netray.co/integrations/infor-m3-chatgpt > Publisher: Netray (netray.co), AI and integration engineering for ERP and manufacturing systems. > Section: Integrations ## Short answer There are three ways to use ChatGPT with Infor M3: paste or upload exports for one-off analysis, build a custom GPT with actions that call a wrapper around the M3 API, or use a ChatGPT enterprise connector. For regulated or export-controlled data, a private model behind the same wrapper is usually safer. ## System Infor M3 ## Capability chatgpt ## Intro Many M3 users already try ChatGPT informally: an item list exported from the grid, a pasted customer order report, a question about a weird MRP message. It works for text that is not sensitive, and it shows quickly how useful natural-language help would be on ERP data. The questions that follow are practical. Can ChatGPT see live M3 data, or only what is pasted? Who is allowed to send what, and does it count as a disclosure of customer pricing or controlled technical data? And if the answer needs to be right, how do you keep it from inventing figures? This page covers the realistic paths, what each costs in effort and risk, and where a private model replaces ChatGPT without changing the user experience. ## Use Cases ### Explaining MRP messages **Description:** Paste a sanitised planned order message set and ask for a plain explanation and likely cause, useful for new planners. ### Drafting supplier emails from PO data **Description:** Provide a PO line summary and let ChatGPT draft the chase or reschedule request, which the buyer reviews and sends. ### Analysing an exported grid **Description:** Export open order lines or inventory balances to CSV and ask for slow movers, outliers, or a summary by facility. ### Report logic translation **Description:** Describe an Infor report or a Compass query and ask ChatGPT to explain or rewrite it, with a developer checking the SQL. ### Custom GPT for item lookup **Description:** A custom GPT with an action calling an internal endpoint that returns item and balance information for named items. ### Training and documentation **Description:** Generate role-based how-to guides for M3 programs from your own procedures, edited by process owners. ## System Surface Exports: M3 grids and panels can export lists to Excel or CSV, and many sites schedule extracts. This is the lowest-effort route, and it gives a static snapshot, not live data. M3 API (MI programs): a thin wrapper service can call transactions such as MMS200MI or OIS100MI and return clean JSON. Expose that wrapper to ChatGPT, not M3 itself. ION API Gateway: on cloud tenants, calls are authenticated with OAuth 2.0 service accounts via an .ionapi credential. Keep these credentials only in the wrapper, never in a GPT configuration. Infor Data Lake and Compass: curated, read-only datasets in the data lake are a safer source for analytical questions than the live transactional database. Custom GPT actions require an OpenAPI schema and an HTTPS endpoint reachable from OpenAI, which usually means a DMZ-hosted service for on-prem M3. ## How It Works ### Classify the data **Detail:** Decide which M3 data is allowed to leave your network: item descriptions and public pricing are different from customer-specific prices or controlled drawings. ### Pick the access pattern **Detail:** Exports for ad hoc work, a custom GPT with actions for repeatable lookups, or an internal tool using the OpenAI API with your own interface. ### Build the wrapper API **Detail:** Create a small service exposing a few read-only endpoints with fixed parameters, row limits, and your own authentication on top of the M3 API. ### Write the OpenAPI definition **Detail:** Describe each endpoint clearly, since the model chooses actions from these descriptions. Return units and statuses explicitly. ### Apply enterprise controls **Detail:** Use an enterprise or API plan with the data-use and retention terms your legal team has reviewed, and restrict who can use the GPT. ### Test for wrong answers **Detail:** Ask questions with known answers from M3 and check totals. Add instructions requiring the model to cite the endpoint result and say when data is missing. ## Example Flows ### Planner asks about item availability **Trigger:** A planner asks a custom GPT about stock for item numbers. **Steps:** - The GPT calls the wrapper's balance endpoint with the item list. - The wrapper queries MMS200MI-style balance transactions under a read-only account. - JSON with warehouse, on-hand, allocated and available returns. - The GPT summarises and flags items below safety stock. **Result:** A quick answer with numbers traceable to the wrapper response. ### Export-based weekly analysis **Trigger:** A finance analyst exports open customer order lines each Monday. **Steps:** - Remove customer names or pricing columns that policy forbids. - Upload the CSV to an approved enterprise workspace. - Ask for backlog by week and the largest delayed lines. - Verify the totals against M3 before sharing. **Result:** A summary in minutes, with a manual check step to guard against errors. ## On Prem ChatGPT is a cloud service, so with on-prem M3 you either send selected data out through a controlled wrapper or you do not use it for that data. If contracts, export rules or customer classifications prevent sending data to a third party, keep the same pattern but swap the model: an OpenAI-compatible endpoint served locally with vLLM or Ollama using an open-weight model lets your wrapper and prompts work unchanged, with everything staying in your network. Many teams use ChatGPT for non-sensitive drafting and a private model for anything touching live order, price or technical data. ## Security - Never put M3 credentials or .ionapi files inside a custom GPT configuration or prompt. - Expose only read-only, allow-listed endpoints and enforce row and field limits server side. - Use enterprise or API terms that exclude training on your data and set retention you have approved. - Respect customer and export-control restrictions before any data leaves your environment. - Log requests at the wrapper so you can audit what was sent and returned. ## Options ### Manual exports into an approved ChatGPT workspace **Fit:** Ad hoc analysis on non-sensitive data. **Effort:** Very low; relies on user discipline. ### Custom GPT with actions over a wrapper API **Fit:** Repeatable live lookups where policy allows cloud processing. **Effort:** A few weeks including security review. ### Private model behind the same wrapper, with a partner such as Netray **Fit:** Regulated, defense or contractually restricted data. **Effort:** Medium; needs GPU capacity and an evaluation set. ## Pitfalls - Staff pasting customer-specific prices or controlled drawings into a personal account. - Trusting arithmetic from the model on pasted tables instead of computing totals in code. - Granting the wrapper broad API access because it is faster to build. - Assuming the GPT can see live M3 when it only has an old export. - Forgetting that an action endpoint must be internet reachable, which opens a new attack surface. ## Faq ### Can ChatGPT connect directly to Infor M3? **Answer:** Not natively. You connect through your own wrapper service that calls the M3 API or ION API Gateway and exposes simple HTTPS endpoints. The wrapper holds the credentials, enforces read-only scope, and shapes the data into compact JSON the model can use reliably. ### Is it safe to upload M3 exports to ChatGPT? **Answer:** It depends on the data and your plan. Enterprise and API terms differ from consumer accounts, and customer contracts or export rules may forbid external processing. Classify the data first and remove restricted columns before any upload. ### Will ChatGPT give accurate numbers from M3 data? **Answer:** Only if the numbers come from tool results and the model is told to use them. Models can miscalculate on large pasted tables, so compute aggregates in the wrapper or in code, and let the model explain rather than sum. ### What is the alternative if we cannot send data out? **Answer:** Run an open-weight model locally and keep the same wrapper and prompts. The user experience is similar for lookup and drafting tasks, while hard reasoning may need a stronger model, which you can test against your own questions. ### Do we need Infor's permission to do this? **Answer:** Check your Infor agreement and API usage terms, especially for cloud tenants where API calls may be governed or metered. Using documented M3 API and ION interfaces under your own service account is the normal route. ## Search terms infor m3 chatgpt, chatgpt for infor m3, chatgpt infor m3 integration, custom gpt infor m3, use chatgpt with infor m3 data, infor m3 openai, infor m3 chatgpt api, m3 erp chatgpt connector, private chatgpt alternative infor m3 --- Talk to an engineer: https://netray.co/contact?intent=erp_ai | Machine catalog: https://netray.co/agent/catalog.json