How to use ChatGPT with Epicor Prophet 21
Short answer
There are three ways to connect ChatGPT to Prophet 21: paste or upload exported data, give a custom GPT an action that calls a small API in front of P21, or connect through a tool server. The first is quick but static, the other two give live answers but need a read-only service you build and secure.
Many P21 users already have ChatGPT open in another tab. A counter salesperson pastes a customer email to get a cleaner reply, a buyer asks it to reformat a supplier spreadsheet, an analyst asks it to explain a formula. The natural next question is whether ChatGPT can just answer from P21 itself: what is on hand for this item, when does this order ship, who bought this part last year.
It can, but only through something you build. ChatGPT has no built-in knowledge of your Prophet 21 database, and Epicor does not ship a ChatGPT connector that you can assume exists for your version. What you get depends on how you bridge the two, and each bridge has different cost, freshness and risk.
This page lays out the realistic options, what data you should and should not send, and the cases where keeping the model on your own servers is the better call.
What teams use it for
Explain and summarize exported reports
Export an aging report, open order report or inventory valuation to CSV and have ChatGPT summarize exceptions, group by customer or draft the commentary for a month-end pack. No live connection required.
Live item and availability lookup
A custom GPT with an action calls a read-only endpoint that returns on-hand, allocated and on-order by location for an item, so staff can ask in plain language instead of opening the inventory window.
Customer reply drafting with order context
The GPT fetches order status and ship dates for a named sales order, then drafts the reply in the company's tone. The salesperson edits and sends from their normal email.
Cross-reference and substitution suggestions
Given a competitor part number or a vague description, the GPT queries item, alternate and cross-reference data to suggest stocked equivalents with availability and price class.
Data cleanup assistance
Feed it a sample of messy item descriptions, units or vendor names and get normalized suggestions to review before an import. This works on exports and never needs write access.
Epicor Prophet 21 integration surface
- Exports and reports: Crystal Reports, the P21 export to Excel or CSV, and scheduled report distribution. The simplest feed for ChatGPT, and static the moment it is created.
- SQL Server read-only views: the most flexible source for a lookup service. Expose only what a user would be allowed to see in P21.
- Epicor Prophet 21 APIs (Interactive, Transaction, and OData style data services in recent versions, depending on build and licensing): a supported path for reads that respect application logic. Confirm availability and naming for your release.
- A thin web API you host: a small service that exposes a few endpoints with an OpenAPI description. This is what a custom GPT action calls, and it is where authentication, row limits and logging live.
- ChatGPT side: custom GPT actions (OpenAPI plus OAuth or API key) on plan tiers that support them, and workspace connectors or apps where your plan allows. Features change often, so check current plan documentation.
How it works
- 1
Decide what the first question is
Pick the three questions staff ask most, for example item availability, order status and last price paid. Everything else can wait.
- 2
Build the read service
Create a small API with one endpoint per question, backed by SQL views or the P21 API. Return compact JSON with only needed fields, and cap result sizes.
- 3
Describe it with OpenAPI
Write an OpenAPI spec with clear operation descriptions. The quality of those descriptions decides whether the GPT picks the right endpoint.
- 4
Secure access
Use OAuth tied to your identity provider where possible so each user sees only their data. A shared API key is simpler but loses per-user scoping and auditability.
- 5
Configure the custom GPT
Add instructions that tell it to answer only from tool results, quote item numbers and dates exactly, and say when data is missing. Restrict sharing to your workspace.
- 6
Pilot with a small group
Have a few counter and purchasing users compare answers to P21 for two weeks, log mismatches and tighten the endpoint or instructions.
Example flows
Where is my order?
Trigger: A salesperson types: status of order 1048223 for Acme Supply.
- The GPT calls the order status action with the order number.
- The service reads header, lines, shipped quantities and expected dates.
- The GPT summarizes what shipped, what is backordered and the next promised date.
- It offers a draft customer email.
Result: The salesperson answers the customer in under a minute without navigating windows.
Month-end exceptions from an export
Trigger: A finance user uploads an AR aging CSV.
- ChatGPT reads the file in the conversation.
- It lists accounts over a threshold by bucket and flags ones with recent promises to pay in the notes column.
- It drafts a short commentary for the controller.
Result: A static analysis with no live access, ready to paste into the close pack after a human check.
Substitute for an out-of-stock item
Trigger: A counter user asks for an equivalent to a discontinued part.
- The GPT calls the alternates and cross-reference endpoint.
- The service returns stocked candidates with availability by location and price.
- The GPT ranks them and states any spec differences recorded in item notes.
Result: The user offers a stocked substitute with price and branch availability in one reply.
Keeping data on-prem
ChatGPT is a hosted service, so anything the GPT action returns leaves your network and is processed by the provider under the terms of your plan. Business and enterprise plans typically state that customer data is not used to train models by default, and offer admin controls, but confirm current terms yourself. For a self-hosted P21 estate the read service has to be reachable from the internet or through a gateway for a custom GPT to call it, which many IT teams dislike. If cost files, contract pricing or customer lists must stay inside, the alternative is a local open-weight model that talks to the same read service over the internal network, giving the same chat experience without the exposure.
Security checklist
- Expose only read endpoints, with a database login limited to named views and no access to credit card, banking or HR data.
- Use per-user OAuth where possible so P21 role restrictions carry through, rather than one all-seeing key.
- Decide what is allowed in ChatGPT: block supplier cost and customer-specific pricing from the response unless the user is entitled to it.
- Log every action call with user, endpoint and parameters, and review them periodically.
- Check your plan's data retention and training settings, and apply workspace-level controls rather than relying on staff using personal accounts.
Ways to get it done
Export and upload only
Effort: None to build, but needs staff discipline about what they upload.
Best for analysis, summaries and cleanup where freshness does not matter and the data is not highly sensitive.
Custom GPT with a read-only action
Effort: A couple of weeks for a few endpoints, plus ongoing upkeep.
Good for live lookups for a defined group, when you can host a secured endpoint and accept hosted processing.
Private model with the same read service
Effort: Weeks for a pilot, plus GPU hosting. Netray and similar partners do this.
Right when pricing, cost or customer data must stay on your network, or auditors require full control of logs.
Pitfalls
- !Pasting live customer or cost data into personal ChatGPT accounts. Provide an approved workspace and say what is permitted.
- !Trusting free-text answers about numbers. Make the GPT quote tool output and refuse to estimate stock or dates.
- !Returning entire tables. Large payloads waste context and increase the chance of leaking data the question did not need.
- !Shared API keys with no user identity, which makes it impossible to apply P21 security or trace who asked what.
- !Assuming a connector exists for your version. Verify before planning around one.
Frequently asked questions
Does Epicor offer an official ChatGPT connector for Prophet 21?
Do not assume so. Epicor has been adding its own AI features, but availability depends on product and version. Ask your Epicor representative what exists for your release. If nothing fits, a custom GPT action over a read-only service you control is the practical route.
Can ChatGPT write to Prophet 21?
Technically a custom action could call a write API, but this is rarely wise. Business logic, approvals and audit matter in an ERP. A safer pattern is for ChatGPT to prepare a draft that a user reviews and posts, or for a separate agent with an approval queue to handle writes.
Is my P21 data used to train ChatGPT?
Under business and enterprise offerings the provider generally states that your data is not used for training by default, but terms and settings vary by plan and change over time. Read the current agreement and configure workspace controls. Personal accounts follow different rules.
Do I need a data warehouse first?
Not for a small pilot. SQL views over the P21 database, ideally on a replica, are enough for item, order and customer lookups. A warehouse becomes useful when you want history, cross-system joins or heavy analytics without loading the production server.
What if our P21 is hosted by Epicor?
Then direct SQL access is usually not available, so build the read service on the supported APIs or on scheduled data exports. Confirm with Epicor what interfaces and limits apply to your hosted tenant before you design the action.
Want ChatGPT working on your Epicor Prophet 21?
A free 30 minute assessment: we map the integration path for your version and setup, and where it should run.
Get a free assessment