How to use ChatGPT with proALPHA
Short answer
There are three routes: paste or upload exports for ad hoc analysis, build a custom GPT with actions that call a small API in front of proALPHA, or skip ChatGPT for sensitive data and use a private model. Pick by data sensitivity, not by convenience.
People running proALPHA usually start with ChatGPT in the simplest way: export a list, upload it, and ask for a summary. That works for a one-off question about supplier delivery performance or stock ageing, and it costs nothing to try. It stops working when the answer needs live data, when the export contains customer or pricing information, or when the same question is asked daily.
The next step is a custom GPT or a ChatGPT workspace connector that can call an endpoint. proALPHA does not ship a ready-made ChatGPT connector that we can point to, so the endpoint is something you provide: a small service that reads from the OpenEdge database or from proALPHA's exposed services and returns narrow, safe results.
The deciding question is where the data goes. Many proALPHA customers are machine builders and suppliers with drawings, customer prices and contractual terms in the ERP. For them the honest answer is often to use ChatGPT for non-sensitive tasks and a private model for ERP data.
What teams use it for
Ad hoc analysis of an export
Export open order lines to CSV, upload it, and ask which customers have the highest share of lines past due. Good for a one-time question with non-sensitive data.
Custom GPT for order status
A GPT with one action, get_order_status, that calls your API. Sales staff ask in plain German or English and receive status and dates.
Drafting supplier and customer correspondence
The GPT takes structured order data from the action and writes a polite delay notice, leaving the sending to a human.
Explaining ERP field meanings
New staff paste a screen description or field list and ask what it means in a process context. No live data is needed.
Formula and report help
Planners ask ChatGPT to build an Excel formula or Power Query step for an export they already hold.
proALPHA integration surface
- Exports: proALPHA list outputs, CSV and Excel exports are the lowest-risk way to hand data to ChatGPT. Strip personal data and prices before upload.
- OpenEdge SQL access: a read-only ODBC or JDBC account lets your own small API query selected tables. Names and structure vary by version and customisation.
- Application services: where your release exposes web service or REST style endpoints through the OpenEdge application server, your API can wrap them so authorisation logic stays in proALPHA.
- Custom GPT actions require an HTTPS endpoint described by an OpenAPI schema, so on-prem proALPHA needs a reverse proxy or gateway published safely, which is itself a security decision.
- Authentication: proALPHA users are managed in the ERP, so your API needs its own mapping to the asking person if row-level rules matter.
How it works
- 1
Classify the data
List which fields are public, internal or restricted. Customer prices, drawings and supplier terms usually stay out of any third-party cloud.
- 2
Start with exports for safe questions
Use ChatGPT Team or Enterprise rather than a personal account, and confirm the workspace data-use settings with your administrator.
- 3
Build a narrow API
Create two or three endpoints such as order status and stock by item, each returning a small JSON object, not table dumps.
- 4
Describe it in OpenAPI
Write the schema with clear parameter descriptions. The quality of descriptions largely decides whether the GPT calls it correctly.
- 5
Create the custom GPT
Add instructions, attach the action, and set authentication. Restrict sharing to your workspace.
- 6
Test against known orders
Compare the GPT answer with the proALPHA screen for twenty real orders before anyone relies on it.
- 7
Decide what stays private
Route restricted questions to a local model with the same API, so users get one habit and your data policy still holds.
Example flows
Where is my order
Trigger: Sales person asks the custom GPT about a customer order
- GPT extracts the order number.
- Action calls your API with the number.
- API reads status, promised date and last goods movement.
- GPT answers and offers a draft customer message.
Result: A fast answer without opening three screens, with the draft left for the person to send.
Supplier delivery review
Trigger: Quarterly purchasing meeting
- Buyer exports receipts with promised and actual dates, with supplier names pseudonymised.
- Uploads to ChatGPT and asks for on-time percentage by supplier.
- Reviews the chart and notes the outliers.
Result: A meeting pack in minutes, using data that was prepared to be safe to share.
Keeping data on-prem
ChatGPT itself is a hosted service, so a strictly on-prem or air-gapped proALPHA site cannot use it for ERP data. The usual compromise is to keep your API and gateway inside the network and expose only a minimal, filtered response outward, or to avoid exposure entirely and run a local open-weight model behind the same API contract. That gives users a ChatGPT-like experience while customer data, prices and drawings never leave the plant. Where a hybrid is acceptable, keep ChatGPT for general drafting and training questions and local models for anything touching ERP rows.
Security checklist
- Use a business workspace plan with administrator controls and review its data retention and training terms yourself.
- Return the minimum fields; never expose a generic query endpoint to a GPT.
- Authenticate every action call and rate-limit it; the endpoint is an internet-facing door into your ERP data.
- Pseudonymise or remove personal data in exports before upload.
- Log which user triggered which action so you can answer an audit question later.
Ways to get it done
Manual export plus ChatGPT
Effort: Minutes per use; no build.
One-off analysis on non-sensitive data.
Custom GPT with an in-house API
Effort: Days to a couple of weeks including security review.
Small teams with web developers and low-sensitivity order status needs.
Private model behind the same API (Netray or in-house)
Effort: Several weeks for a first pilot including GPU setup.
Customer prices, contracts or defense-adjacent work where data must stay local.
Pitfalls
- !Using personal ChatGPT accounts for ERP exports, which sits outside company controls.
- !Publishing an on-prem API to the internet without a gateway, authentication and rate limits.
- !Letting the GPT guess internal codes, instead of giving it lookup actions for items and customers.
- !Treating the answer as authoritative without sampling against proALPHA screens.
- !Forgetting that actions can also be abused through injected text in free-text fields.
Frequently asked questions
Is there an official ChatGPT connector for proALPHA?
We are not aware of a standard one, and availability could change, so check with your proALPHA partner. In practice you build a small API in front of the database or services and describe it to a custom GPT with an OpenAPI schema.
Is it safe to upload proALPHA exports to ChatGPT?
It depends on the data and your plan. Business plans offer administrative controls, but you must still decide whether customer prices, personal data or drawings may leave your environment. Remove or pseudonymise sensitive fields first, and keep restricted data local.
Can ChatGPT write back to proALPHA?
Technically an action can call a write endpoint, but we advise against it at first. Have ChatGPT draft and propose, then let a person confirm and execute through proALPHA's own services so validation and audit still run.
Why would I use a private model instead?
Data control, predictable cost at volume and no dependence on an outside service. For regulated or contract-sensitive manufacturing data, a local open-weight model behind the same API often meets policy where a hosted service does not.
Does this work with on-prem proALPHA?
Yes for exports. For live actions, the endpoint must be reachable from the service, which means a controlled gateway. If that is not acceptable, run the model locally instead of exposing the ERP outward.
Want ChatGPT working on your proALPHA?
A free 30 minute assessment: we map the integration path for your version and setup, and where it should run.
Get a free assessment