About KatanaPIM
KatanaPIM is an enterprise SaaS Product Information Management (PIM) platform for online sellers, based in Leiden. Its promise is “Publish once, sell everywhere.” Brands use it to centralise product data from suppliers, ERPs and spreadsheets, enrich it with translated and SEO-optimised content, validate it through custom workflows, and distribute it to sales channels such as Shopify, Magento, BigCommerce, WooCommerce, Amazon, bol and Zalando.
The challenge
As a SaaS product company, KatanaPIM had to keep developing its core platform while also delivering implementations for a growing list of e-commerce customers, each with its own catalogue structure, channels and integrations. That combination needs senior .NET engineers who can work independently on a complex multi-tenant product, backed by solid QA so releases don’t break live customer catalogues.
Our approach
Our way of working is trust first, scale second: we prove our value on real work, and only then grow the relationship. Amsby set up a dedicated team in India that worked as part of KatanaPIM’s own product and delivery organisation:
- 3 senior .NET developers for platform development, customer implementations, integrations and APIs.
- 1 software tester for functional and regression testing of platform releases and customer implementations.
The team delivered and supported KatanaPIM implementations for e-commerce brands across very different product categories, from baby products and Asian food to wallpaper, wine and guitar parts.
From team to AI programme
After years of working on the platform, KatanaPIM came back to Amsby for automation, and then for two AI agents. One client, a growing relationship: from development capacity to AI that works across product, marketing and support.
Product data onboarding automation
Every new PIM customer starts with the same hurdle: getting their existing product data into the PIM. For manufacturers this usually means a large ERP export with dozens of Dutch-labelled columns, covering article numbers, EAN codes, units, shelf-life rules, pallet configuration and more. None of it matches the PIM’s data model, and mapping it by hand is slow and error-prone.
- The customer’s ERP export (Excel) is uploaded through a simple web form.
- All 40+ source fields are stored in a master data sheet, keeping a complete, auditable copy of the original data.
- A transformation step maps the source fields to KatanaPIM’s data model (SKU, name, description, GTIN, weight, unit, brand, category, supplier).
- The mapped data is written to five import-ready tables: Products, Stock, Suppliers, Descriptions and Categories.
Built with n8n and Google Sheets, the same mapping runs for every export and every update.
AI release notes from Jira
Every SaaS release ends with the same task: going through dozens of Jira tickets, working out which changes customers will notice, translating developer language into customer benefits and building a branded email. That took hours of a product owner’s time every release.
- When a release is marked as released in Jira, a webhook starts the agent automatically; a simple form also lets the product owner generate notes on demand.
- The agent collects every ticket in the Jira fix version and converts it into clean text.
- Claude filters out internal work, groups related tickets into topics, classifies them as New feature, Improvement or Fix, and scores their impact.
- Claude drafts the email in KatanaPIM’s established release-email tone, with “Learn more” links chosen only from a whitelist of real help-centre pages, so the AI cannot invent URLs.
- The draft is rendered as a fully branded HTML email in KatanaPIM’s house style.
- The product owner receives it marked DRAFT FOR REVIEW, with a reviewer-only panel showing which tickets each topic came from and which were excluded and why. Nothing reaches customers without human approval.
AI support assistant (RAG)
Support teams spend much of their time answering questions that have already been answered, in the documentation or in earlier tickets. Quality depends on who picks up the ticket, and real bugs can sit in the queue before anyone recognises them.
- Every night, an indexer collects the Confluence documentation and up to 200 recently resolved tickets, and stores them as vector embeddings in Elasticsearch, so every resolved ticket becomes part of the knowledge base.
- A Jira Service Management webhook starts the assistant as soon as a customer raises a ticket.
- The ticket is matched via vector search against the knowledge base, retrieving the 8 most relevant passages.
- Claude drafts a first response in the customer’s own language, bases every claim on the retrieved sources, cites them by number, and says so openly when the sources don’t cover the question.
- Each draft gets a confidence level and a recommended route: draft ready, needs more info, or escalate to development.
- The result is posted as an internal note on the ticket, so the customer never sees an unreviewed AI answer; likely bugs automatically get an escalation label.