How to add an AI chatbot to Katana
Short answer
Katana does not embed a chat assistant in your customer portal, so you build one outside it: a chat front end, a backend that calls the Katana REST API for facts, and a model that phrases the answer. Scope it to order status and stock, and hand over to people for anything else.
Small manufacturers on Katana get the same questions all day. Where is my order? Do you have this in stock? When will the next batch be ready? Someone with Katana open answers each one, and that someone is often the owner.
A chatbot can take those questions, but only if its answers come from Katana rather than from a model's imagination. The architecture is simple: authenticate the asker, fetch the facts, let the model write a plain reply, and refuse anything outside the allowed scope.
What teams use it for
Customer order status
A customer enters an order number and email, the bot checks them against Katana and reports status and expected dispatch.
Availability for wholesale buyers
Trade customers ask if a variant is available in a quantity; the bot reads stock and open production to answer with a realistic date.
Internal stock questions on the shop floor
Staff ask where a material is and how much is left, instead of walking to a terminal.
Product specification answers
Using product descriptions, units and your documentation, the bot answers simple spec questions and links to data sheets.
After-hours triage
Outside business hours, the bot collects a request, checks status and creates a ticket for a human in the morning.
Katana integration surface
- Katana REST API for sales orders, manufacturing orders, stock and product data, accessed with a bearer API key held only by your chatbot backend.
- Customer identification through fields on the sales order or customer record such as email; confirm which fields are searchable via the API for your plan.
- Webhooks to refresh a small status cache when orders or manufacturing orders change, reducing live API calls.
- Storefront integrations (Shopify and others) that may hold customer accounts; you can use their login rather than inventing your own.
- No built-in chat widget in Katana, so the front end is a website widget, a messaging channel or an internal tool of your choosing.
How it works
- 1
Define scope and refusals
List exactly what the bot may answer: order status, stock, lead time, basic specs. Everything else routes to a person.
- 2
Choose the front end
A website widget, your customer portal, WhatsApp, or an internal chat. The backend is the same for each.
- 3
Build a lookup layer
Implement a few functions that call Katana with strict inputs and return minimal fields: status, dates, quantity available.
- 4
Verify identity
Require order number plus email, or a logged in account, before returning order data. Never expose data on an order number alone.
- 5
Ground the model
The model gets the function result and your tone rules, and must answer only from that data, saying so when it has none.
- 6
Add handoff and feedback
A button or trigger sends the conversation and context to a person by email or helpdesk. Review transcripts weekly to find gaps.
Example flows
Where is my order
Trigger: Customer opens the chat widget and asks about an order
- Bot asks for order number and email.
- Backend finds the sales order and checks the email matches.
- It reads order status and linked manufacturing order progress.
- Bot replies with status, expected dispatch and a tracking link if one is stored.
Result: A routine question answered in seconds, with no staff time.
Trade availability enquiry
Trigger: Wholesale buyer asks for 400 units of a variant
- Bot resolves the variant and calls the stock function.
- If stock is short, it checks open manufacturing orders for that variant.
- It replies with what is available now and the next expected quantity and date.
- It offers to send the request to sales for a firm quote.
Result: The buyer gets a realistic answer, and sales receives a qualified lead.
Keeping data on-prem
Katana is hosted, so the chatbot sits beside it rather than inside it. If you want the language model private, host an open-weight model on your own server and run the whole backend there; it calls the Katana API outbound and your chat front end calls the backend. Transcripts, prompts and cached order status then stay in your environment. Small shops often find a hosted model acceptable for order status, so decide based on contract and customer expectations rather than assumption.
Security checklist
- Require identity verification before returning any order-specific data.
- Return the minimum fields needed and never include costs, margins or other customers' details.
- Rate limit the public endpoint and add bot protection to stop scraping of order numbers.
- Keep the Katana API key on the backend only and use a key dedicated to the chatbot.
- Retain transcripts for a defined period and tell users they are talking to an AI.
Ways to get it done
Helpdesk chatbot product with a custom API action
Effort: Days
You already use a helpdesk tool and want quick coverage of status questions.
Build a custom backend and widget
Effort: Two to four weeks
You want control over identity checks, tone and cost.
Netray builds with an optional private model
Effort: A scoped pilot
Data handling or integration with several systems makes a bespoke solution worthwhile.
Pitfalls
- !Letting the model answer from memory when the lookup fails; force an I do not know path.
- !Looking up orders by number only, exposing other customers' information.
- !Calling the live API on every message and hitting rate limits during busy periods.
- !Quoting delivery dates the production schedule cannot honour; use cautious wording.
- !No human handoff, which frustrates customers whose issue is outside scope.
Frequently asked questions
Does Katana include a customer chatbot?
Not as a customer-facing assistant grounded in your orders, as far as we know, though features change. Check Katana's release notes. Most teams build a small layer using the API and a chat widget, or use a helpdesk bot with an API action.
How does the chatbot know my customer's order?
The backend searches Katana through its API using order number and a matching identifier like email, then reads status and manufacturing progress. The model only receives that result and writes the reply, which keeps answers factual.
Can the chatbot take new orders?
It can collect order details and create a draft or send them to sales, but we advise a person confirm before anything lands in Katana. Pricing, stock and customer terms are easy to get wrong without review.
Can it run without sending data to a cloud model?
Yes, with an open-weight model on your hardware. The Katana data still travels from Katana's cloud to your backend, but prompts, transcripts and cached data stay with you. Quality is fine for status and stock answers.
How much does it cost to run?
For status and stock questions, model cost per conversation is typically small, and the larger cost is build and maintenance time. Caching lookups and keeping prompts short reduce both API calls and model spend, so measure real usage during a pilot.
Want AI chatbot 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