PIM Implementation with HITL AI for Sanitas Troesch

Sanitas Troesch onboards more than 200,000 supplier SKUs into its PIM. Suppliers ship product data as 40-page PDF datasheets, multi-tab Excel workbooks, Jira tickets and email attachments — every supplier on a different schema. Internal product managers were retyping fields, normalising units and reconciling category trees by hand, while the same SKU often bounced between supplier and admin for weeks before it was clean enough to publish.

The brief was clear: replace the manual entry workload with an AI-assisted pipeline that suppliers and internal staff could share, without giving up the audit trail this kind of master-data work needs. This is the pim implementation problem we were hired to solve.

The platform splits into three separately deployable services so each side can be operated, scaled and replaced on its own clock:

• Laravel 12 admin API — source of truth for products, workflow state and role-scoped editing rights. MySQL is the persistent store, Redis carries queues and locks.
• React 19 SPA — the working surface for suppliers and internal admins, with live progress updates over Pusher.
• Node Mastra AI pipeline — the multi-agent ai brain runs out of process, so a slow OpenAI or Mistral call never blocks the web app or the editor.

Everything ships in Docker; the AI pipeline can be redeployed independently of the admin app, which matters when prompt or model providers change. The architecture sits squarely inside our Master Data Management practice.

One prompt cannot reliably turn a 40-page supplier PDF into a clean PIM row. The Mastra pipeline splits the job across specialist agents that hand work to each other:

• Analysis agent — classifies the document, locates product blocks and decides which downstream agents to call.
• Extraction agent — the ai document extraction step itself: pulls raw fields out of PDF, XLSX or text into a typed payload.
• Preparation agent — normalises units, cleans formatting and resolves obvious duplicates.
• Attribute-assignment agent — maps the cleaned payload onto Sanitas Troesch's PIM attribute schema.

Each agent is given a tight toolbox of deterministic helpers — a PDF reader, an XLSX reader, a unit converter, a calculator — so model output is grounded in real bytes, not improvised. OpenAI and Mistral are wired as interchangeable providers behind the same agent interface, and this is the heart of the ai product information management work we ship for Swiss wholesalers.

Sanitas Troesch needed suppliers and internal admins editing the same product records side by side, with strict boundaries: a supplier portal where each supplier sees only their own catalogue and can only edit scoped fields, and an internal admin surface that sees everything and owns category, pricing and publication decisions.

Laravel enforces the policy on every write. A workflow state machine tracks each SKU through draft, review, ready and published states; every transition writes a status-history record with the actor, role, event payload and timestamp. When auditors or compliance ask how a value got into the PIM, the answer is a query, not a Slack archaeology session — the kind of governance Swiss b2b automation projects actually have to deliver.

Suppliers do not all want to learn a new portal. The platform meets them where they already work:

• Direct upload in the supplier portal — drag a PDF or Excel file, watch the pipeline status stream back over Pusher.
• Jira integration — supplier tickets attached to a Jira workflow flow into the same intake.
• Email intake — attachments sent to a monitored address are routed to the analysis agent automatically.

All paths converge on the same multi-agent extraction step and the same human-in-the-loop ai workflow review queue, so an SKU that arrived as an email attachment is governed the same way as one entered in the portal. The document side of this work is documented in detail under our Document Automation hub.

What Sanitas Troesch now has in production:

• Manual product data entry replaced by an AI-assisted HITL pipeline across the full 200k+ SKU base.
• A faster, more predictable supplier onboarding cycle, with status visible to both sides at every step.
• Consistent attribute formats across the catalogue — units, dimensions and option sets normalised by deterministic tools, not human reformatting.
• A clean supplier-vs-internal ownership boundary on every edit, recorded in the audit trail.

If you run a Swiss or DACH wholesale, building-materials or industrial-supply business with a large SKU base and a PIM, the same pattern transfers. Read the FAQ below for the questions buyers usually ask, and when you are ready to scope yours, book a discovery call to plan your PIM AI onboarding.