How to use ChatGPT with Dynamics 365 Business Central
Short answer
You can connect ChatGPT to Business Central through a custom GPT or app that calls a thin proxy in front of the API v2.0, or you can export data into chats manually. The proxy matters: it holds credentials, filters fields and enforces read-only access.
Business Central users already live in Microsoft 365, so the first thing most teams do is paste a list of customer ledger entries into ChatGPT and ask for a summary. It works, and it is also how salary data and customer bank details end up in places nobody approved.
A deliberate setup replaces copy and paste with a governed connection: ChatGPT calls a small service you control, the service calls Business Central with a scoped identity, and only the fields you decided on come back.
This page covers the three levels of that: manual export, a custom GPT with actions, and a managed company-wide setup, along with where ChatGPT is the wrong tool for regulated manufacturing data.
What teams use it for
Explain a customer's overdue position
A credit controller asks about one customer. The action returns open customer ledger entries and aging, and ChatGPT writes the collections email with the correct invoice numbers.
Draft vendor and customer emails from live records
Given a PO number, the GPT reads the header and lines and drafts a chase or confirmation message in the buyer's tone.
Explain variances in plain language
Export a month of item ledger entries or a production variance report and ask for the three biggest drivers. Good as a first read, not as an audit trail.
Write AL or Power Automate snippets
Developers use ChatGPT to draft AL extension code, table extensions and event subscribers. This needs no live data connection at all.
Translate and standardise item descriptions
Bulk-clean item descriptions and add translations, then import through a configuration package or the items API after review.
Microsoft Dynamics 365 Business Central integration surface
- API v2.0 REST endpoints under /api/v2.0/companies({id}) with OData query options, which a custom GPT action can describe in an OpenAPI schema. Keep the schema to a handful of read operations.
- OData V4 web services published from any page or query object, useful for filtered datasets that the standard API does not offer.
- Export paths for the manual route: Edit in Excel, page export to Excel or CSV, and report data export. These create files that people then upload to ChatGPT.
- Entra ID OAuth 2.0 for the service identity. A GPT action cannot safely hold a client secret for a tenant, so the secret lives in your proxy and the GPT authenticates to the proxy instead.
- Rate limits and throttling in SaaS: the service returns HTTP 429 under heavy load, so a proxy should cache and page results rather than let a chat issue wide queries.
How it works
- 1
Decide manual, GPT action or app
Manual exports suit one-off analysis. A custom GPT with actions suits a team that asks the same ten questions. A larger rollout suits a managed ChatGPT workspace with admin controls.
- 2
Stand up a proxy
A small service (Azure Function, container or on-prem host) that authenticates the caller, holds the Entra client credentials, calls the BC API, trims fields and returns JSON.
- 3
Write the OpenAPI schema for actions
Describe only the operations you want: getCustomerBalance, getPurchaseOrder, listLateLines. Use clear parameter descriptions so the model fills them correctly.
- 4
Create the custom GPT or app
Add instructions (answer only from returned data, show record numbers, never guess balances), attach the actions, and authenticate users via OAuth to your proxy rather than shared keys.
- 5
Review data controls
Check your ChatGPT plan's data retention and training settings, workspace admin controls and region options against your customer contracts. Decide which entities are never allowed.
- 6
Test with real questions and pin results
Ask 30 real questions, compare against Business Central pages, and record failures. Tighten the schema until answers are boringly correct.
Example flows
Collections email for one customer
Trigger: A credit controller types a customer number into the GPT.
- GPT calls getCustomerAging on the proxy.
- Proxy queries customer ledger entries filtered to open items and returns amounts, due dates and document numbers.
- GPT drafts a polite reminder referencing those documents.
- Controller copies the email into Outlook.
Result: A correct, specific reminder in under a minute with no free-typed invoice numbers.
Month-end exception review via export
Trigger: Finance exports the general ledger entries for one account to Excel.
- A person strips bank and personal data columns.
- File is uploaded to a ChatGPT workspace configured for no training.
- ChatGPT flags postings with odd descriptions or unusual amounts.
- Finance investigates each flagged item in Business Central.
Result: A faster first pass, with a human confirming each finding against the ledger.
Keeping data on-prem
ChatGPT is a hosted service, so any live action means your Business Central data crosses to OpenAI's infrastructure under your plan's terms. For on-premises Business Central the proxy has to be reachable from the public internet or through a gateway, which is often exactly what an on-prem policy forbids. If drawings, defense work, pricing or export-controlled data are involved, a locally served open-weight model (Llama, Qwen, Mistral, gpt-oss class) behind the same proxy pattern gives you the chat experience with no external call. The proxy and tool schema you build for ChatGPT can usually be reused unchanged when you swap the model, so the work is not wasted.
Security checklist
- Never put the Entra client secret inside the GPT configuration. Keep it in the proxy and authenticate users to the proxy.
- Return only whitelisted fields. Bank accounts, tax ids and employee data should not be in any action response.
- Use read-only operations first; if you add writes, make them create drafts only.
- Confirm retention, training and region settings on your ChatGPT plan and keep a record for audits.
- Log proxy requests with the user identity so you can answer who asked ChatGPT for what.
Ways to get it done
Manual exports into a managed ChatGPT workspace
Effort: Minutes to start, ongoing discipline needed.
Ad hoc analysis by finance with sanitised files.
Custom GPT with actions over your own proxy
Effort: One to three weeks including schema, auth and testing.
A team asking recurring questions, with SaaS Business Central and acceptable data terms.
Private model with the same tool layer (in-house or with Netray)
Effort: Several weeks, including GPU hosting and evaluation.
Regulated, export-controlled or on-premises environments.
Pitfalls
- !Letting users upload unfiltered exports that include bank details or personal data.
- !Exposing the whole API as an action. The model will happily call endpoints you did not mean it to.
- !Trusting arithmetic produced by the model. Have the proxy compute totals and return them.
- !Assuming a consumer ChatGPT account has the same controls as a business plan.
- !Forgetting multi-company setups: always pass the company explicitly through the proxy and show it in answers.
Frequently asked questions
Is there an official ChatGPT connector for Business Central?
Availability of prebuilt connectors changes quickly, so check the current ChatGPT apps directory and Microsoft AppSource. If none fits, a custom GPT calling your own proxy over the Business Central API is a dependable route that you fully control.
Can ChatGPT write to Business Central?
Technically yes through API actions, but you should not allow it. Have actions create drafts such as an unposted journal line or an open purchase order and require a user to review and post. That keeps accountability where it belongs.
Is it safe to paste Business Central data into ChatGPT?
It depends on your plan, your contracts and the data. Business plans offer retention and training controls that consumer accounts do not. Even so, strip personal, banking and export-controlled data, and write a short policy that says which exports are allowed.
Why not just use Microsoft Copilot?
Copilot inside Business Central is a good default for supported tasks in SaaS. ChatGPT is attractive when staff already use it, when you want a custom workflow, or when you want to compare models. The best answer is often both, with clear data rules.
Does this work with Business Central on-premises?
It can, if the proxy can reach your service tier and ChatGPT can reach the proxy, which usually means a DMZ or gateway. Many on-premises customers choose a private model instead so no request leaves the network.
More for Microsoft Dynamics 365 Business Central
Want ChatGPT working on your Microsoft Dynamics 365 Business Central?
A free 30 minute assessment: we map the integration path for your version and setup, and where it should run.
Get a free assessment