How to use ChatGPT with ECI M1
Short answer
There are three routes: paste or upload exports into ChatGPT, give a custom GPT live read access through an API you host in front of M1, or run a private model instead. Because M1 usually sits on-prem, the API you host is the part that needs real design.
Most M1 users first meet ChatGPT as a personal tool: someone exports a report to Excel, uploads it, and asks for a summary. That works for one-off analysis and is also the riskiest pattern, because exports of costs, customers and part numbers end up in individual accounts.
The structured version is a custom GPT or a ChatGPT connector that calls a small API you control. The API reads from M1's SQL Server data, enforces who can ask what, and returns compact JSON. ChatGPT supplies the language, you supply the guardrails.
This page walks through that path and is candid about where ChatGPT is the wrong tool, mainly where export-controlled or customer-restricted data is involved.
What teams use it for
Explain a variance
Upload a job cost export and ask why actual material exceeded estimate, with the model comparing operations, scrap and substitutions.
Customer-ready status wording
Give ChatGPT a list of open order lines and promise dates and have it draft a polite delay notice for a customer service rep to send.
Custom GPT for order lookups
A custom GPT with an action calling your read-only API answers questions like what is open for customer X shipping this week.
Spreadsheet cleanup for item master work
Normalise descriptions, units of measure and vendor part numbers in an export before a controlled re-import.
SOP and work instruction drafting
Draft inspection steps or training notes from routing text, reviewed by a quality engineer before release.
ECI M1 integration surface
- SQL Server data behind M1: the source for any live-access API. Expose views, not raw tables.
- M1 report outputs to Excel or CSV: the simplest feed for the export route.
- Any web services or integration module available for your version, to be verified with ECI before relying on it.
- A small HTTPS API you host (for example a FastAPI or ASP.NET service) that adapts M1 data to an OpenAPI schema ChatGPT actions can call.
- Entra ID or your identity provider for authenticating the API's callers.
How it works
- 1
Decide export or live access
If questions are ad hoc and data is low sensitivity, governed exports may do. If you want repeatable lookups, build the API.
- 2
Classify your M1 data
Mark which fields are export-controlled, customer-restricted or commercially sensitive, and decide which may ever leave your network.
- 3
Build the read API
Create narrow endpoints such as order status, item availability and open PO by vendor, each backed by a SQL view and with row limits.
- 4
Describe it in OpenAPI
Write a precise schema with clear descriptions so a custom GPT knows when to call each action and what the fields mean.
- 5
Publish through your workspace
Use a ChatGPT Business or Enterprise workspace so sharing, retention and training settings are under admin control, and restrict the GPT to named users.
- 6
Test with real questions
Collect 50 questions from planners and customer service, record correct answers, and fix schema descriptions where the model picks the wrong action.
Example flows
Customer service order lookup
Trigger: A rep asks the GPT what is still open for a named customer.
- The GPT calls the order status action with the customer code.
- The API reads the open order view and returns lines with quantities and promise dates.
- The GPT summarises, flagging lines that are late.
- The rep copies a drafted update to the customer email.
Result: The rep answers in a minute instead of navigating several M1 screens.
Weekly cost review prep
Trigger: A controller uploads the job cost report.
- ChatGPT parses the file and groups jobs by overrun percentage.
- It lists the top overruns with likely drivers from the columns present.
- The controller asks follow-ups and exports the summary to a slide.
Result: Review preparation drops from hours to minutes, with the controller still verifying figures.
Keeping data on-prem
ChatGPT is a cloud service, so a fully on-prem M1 site must decide what is allowed across the boundary. For data that cannot leave, run an open-weight model (Llama, Qwen, Mistral or gpt-oss class) on your own GPUs with vLLM or Ollama and put the same read API in front of it. The OpenAPI design carries over: you swap the model, not the integration. Many shops run both, using ChatGPT for non-sensitive drafting and a private model for anything touching customer drawings, costs or controlled parts. A hosted API needs an inbound path from the internet to your on-prem service, which your network team should place in a DMZ with rate limiting.
Security checklist
- Use a business or enterprise workspace with training on your data disabled and admin-managed sharing; avoid personal accounts for anything M1-related.
- Expose only read endpoints first, authenticated per user rather than with one shared key.
- Strip or tokenise export-controlled and customer-restricted fields before they leave the network.
- Log every action call with user, parameters and row count.
- Review vendor data-retention and region terms against any contractual obligations to your customers.
Ways to get it done
Governed exports into ChatGPT Business or Enterprise
Effort: Days, mostly policy and training.
Reasonable for non-sensitive analysis and drafting.
Custom GPT with an in-house read API
Effort: Two to four weeks for a handful of endpoints.
When you have a developer and want repeatable lookups.
Partner-built private assistant with optional ChatGPT for safe tasks
Effort: A scoped pilot, usually weeks.
When regulated data is involved and you want a private model.
Pitfalls
- !Employees uploading full cost or customer exports to personal ChatGPT accounts.
- !A shared API key that makes every GPT user look like one super-user.
- !Letting the GPT guess at columns: vague OpenAPI descriptions cause wrong answers.
- !Trusting arithmetic on large exports without checking totals against M1.
- !Opening the API to the internet without rate limits or authentication.
Frequently asked questions
Can ChatGPT connect directly to ECI M1?
Not out of the box in a way you should rely on. The usual pattern is a small API you host that reads M1's data and exposes it to a custom GPT as an action. Confirm integration options for your version with ECI.
Is it safe to upload M1 exports?
Only under a business or enterprise agreement with data controls, and only for data your policies allow to leave your network. Export-controlled or customer-restricted data usually belongs on a private model instead.
Can ChatGPT write back to M1?
Technically yes through an API, but we recommend read-only until you have approval workflows. Writes should use a supported M1 import or API path, not direct SQL inserts.
Do I need a developer?
For exports, no. For live lookups you need someone who can host a small web service and write SQL views against your M1 schema, in-house or through a partner.
What is the alternative if data must stay on-prem?
Run an open-weight model locally and reuse the same read API. You keep the conversational experience while the data and the model both stay inside your network.
Want ChatGPT working on your ECI M1?
A free 30 minute assessment: we map the integration path for your version and setup, and where it should run.
Get a free assessment