How to use ChatGPT with Katana
Short answer
You can use ChatGPT with Katana by building a custom GPT whose actions call the Katana REST API, or by exporting data to a file and uploading it. Actions give live answers, exports are simpler and safer for one-off analysis. Review what leaves your account before connecting real order and recipe data.
Many Katana users already paste a stock export into ChatGPT to ask what to reorder. It works, but it is a snapshot, and it is easy to forget which version you uploaded. The more useful setup is a custom GPT with an action pointed at the Katana API, so questions like which sales orders are blocked by material return fresh facts.
The catch is that Katana is a system of record for your recipes, prices and customers. ChatGPT is a general tool sitting outside it. This page covers the three ways to connect them, what each costs in effort and risk, and the point where a private model is the more sensible answer.
What teams use it for
Ask about stock in plain language
A custom GPT answers questions like how many units of a variant are available at each location, calling a stock endpoint rather than guessing from an old export.
Explain a delayed sales order
Given a sales order number, the GPT fetches the order, its rows and linked manufacturing orders and explains which material or operation is holding it up.
Write supplier chasers
Using open purchase order data, it drafts a polite email per supplier listing overdue lines and requested dates.
Spreadsheet analysis of exports
Upload a sales order and cost export and ask for margin by product family, slow movers or ingredient price drift.
Quote and recipe sanity checks
Paste a proposed recipe change and ask the GPT to compare cost impact against current material costs you supply.
Training new planners
A GPT with your internal planning rules plus read-only stock tools acts as a coach for new staff.
Katana integration surface
- Katana REST API with bearer API key authentication, suitable as the backend for a custom GPT action defined by an OpenAPI schema. Katana may publish a spec; if not, you write a trimmed one covering only the endpoints you need.
- Custom GPT actions require a publicly reachable HTTPS endpoint. In practice you put a small proxy in front of Katana so the API key stays on your side and responses are trimmed.
- Katana data export options (CSV and similar exports from list views, depending on plan) for the upload route, which needs no API at all.
- Rate limits on the API key, which a chatty GPT can hit quickly; the proxy should cache and throttle.
- No on-prem endpoint exists, so there is no private network path for ChatGPT to reach; everything flows over the public internet.
How it works
- 1
Decide between upload and live action
Use exports for occasional analysis with no engineering. Use an action when people ask the same operational questions daily.
- 2
Build a proxy with read-only endpoints
Create a small service that exposes a handful of GET routes (stock, orders, purchase orders) and holds the Katana key. Never give the GPT the raw key.
- 3
Write the OpenAPI schema
Describe each route with clear parameter names and short descriptions. Good descriptions are what make the GPT pick the right call.
- 4
Authenticate the action
Use an API key or OAuth between ChatGPT and your proxy so only your workspace's GPT can call it.
- 5
Write instructions that constrain behaviour
Tell the GPT to cite which call produced each number, to say when data is missing and to never invent SKUs.
- 6
Pilot with two or three users
Compare answers against the Katana screens for a week and fix schema or trimming issues before wider rollout.
Example flows
Why is order SO-1042 late
Trigger: Planner asks the custom GPT in chat
- GPT calls the order route with the sales order number.
- It reads linked manufacturing orders and statuses.
- It calls the stock route for the blocking material.
- It answers with the shortage and the open PO expected date.
Result: A two sentence answer with the numbers and where they came from.
Monthly margin review from exports
Trigger: Owner uploads two CSV exports
- Upload sales order lines and material cost exports.
- Ask for margin by product family and the five worst variants.
- Ask for suspected causes, such as a material price change.
Result: A quick analysis without exposing a live API connection, accepting that the data is a snapshot.
Keeping data on-prem
Katana has no on-prem edition, but you can still keep the AI side private. If recipes or customer pricing should not go to a third party model, replace ChatGPT with an open-weight model hosted on your own hardware and keep the same proxy pattern: a read-only service in front of the Katana API feeding a chat interface you control. You lose ChatGPT's polish and file tooling, and you gain control of prompts, logs and retention. For many small manufacturers the hosted route with enterprise controls is enough; use the private route when a customer contract or export rule demands it.
Security checklist
- Never paste the Katana API key into a GPT; keep it in your proxy.
- Check your ChatGPT plan's data controls: business and enterprise tiers state that data is not used for training by default, but confirm the current terms yourself.
- Expose only the fields needed; strip customer contact details and unit costs unless the task requires them.
- Restrict who can use the GPT and review the action's call logs from your proxy.
- Treat uploaded exports as copies of regulated data and delete them when finished.
Ways to get it done
Upload exports to ChatGPT
Effort: Minutes
Ad hoc analysis, no setup, accepts stale data.
Custom GPT with a thin Katana proxy
Effort: Several days to two weeks
Daily operational questions with live data, one developer.
Private model with Netray or in-house
Effort: A scoped pilot
Sensitive recipes or contract limits on third party processing.
Pitfalls
- !Trusting numbers from an old export as if they were live.
- !Pointing the GPT directly at Katana and leaking the API key through the schema or instructions.
- !Returning huge payloads so the GPT truncates or misreads rows.
- !Assuming ChatGPT knows Katana terminology; define variant, material and recipe in the instructions.
- !Skipping pagination, so answers cover only the first page of orders.
Frequently asked questions
Is there an official Katana ChatGPT connector?
Check Katana's integrations page and the ChatGPT app directory for the current state, as this changes quickly. If none fits, a custom GPT calling the REST API through your own proxy is a reliable and well understood route.
Does ChatGPT train on my Katana data?
That depends on your plan and settings. Business and enterprise offerings state that customer data is not used for training by default, while consumer settings differ. Read the current terms and apply the strictest setting that your use needs.
Can ChatGPT write back to Katana?
It can if you expose write routes, but we advise against it at first. Let it draft, and have a person confirm in Katana or through an approval step that you control, so mistakes do not change inventory.
Which is better, upload or action?
Upload is better for occasional analysis and has the lowest risk. Actions are better when the same question is asked daily and freshness matters. Many teams use both, with actions limited to read-only stock and order status.
When should I use a private model instead?
When recipes or customer pricing are trade secrets, when a contract limits where data is processed, or when you want full logs and retention control. A private model costs hardware and setup but removes the third party from the data path.
Want ChatGPT working on your Katana?
A free 30 minute assessment: we map the integration path for your version and setup, and where it should run.
Get a free assessment