The number we were asked to move was simple: how long it takes to bring a new supplier live. Nine days, on a good week. Long enough that the merchandising team had quietly built a spreadsheet to track which suppliers were stuck, and with whom.

What made it slow was not the work — it was the shape of the work. A new supplier pack arrived as a bundle of PDFs and spreadsheets, and from there it travelled through five inboxes: intake, compliance, finance, category, and data. Each step waited on the one before it, and each re-keyed the same fields into a different system.

01 — The baselineMeasuring the ‘before’ before we touched anything.

We spent the first week doing nothing but watching. We timed forty real onboarding runs end to end, tagged where each one stalled, and priced the rework. Two findings reframed the brief:

  • Roughly 70% of a pack was complete and unambiguous on arrival — it never needed a human at all, it just waited in a queue for one.
  • The slow 30% was slow for a handful of repeating reasons: a missing bank-verification doc, a GST number that didn’t match, a product taxonomy the catalogue had never seen.

That split is the whole case. If most of the volume is routine, the win isn’t a smarter human — it’s getting the routine out of the human’s way and pointing their attention at the exceptions.

“We weren’t drowning in hard decisions. We were drowning in easy ones that had to wait their turn.”

Head of Merchandising Operations

02 — What we builtOne pipeline, not five inboxes.

We replaced the relay with a single agentic pipeline. A supplier pack lands once and is processed in one pass:

  1. Intake — the agent reads every document in the pack, extracts the fields each downstream system needs, and normalises them against the existing supplier master.
  2. Validation — bank details, GST and NZBN numbers, and insurance dates are checked against source registries, not just against each other.
  3. Enrichment — products are mapped to the catalogue taxonomy, with the agent proposing a category and a confidence score instead of leaving the field blank.

Anything that clears every check is staged for go-live automatically. Anything that doesn’t is routed to one reviewer with the specific problem highlighted and a suggested fix attached — no hunting through a 40-page pack to find the one stale date.

Pipeline diagram
Intake → validation → enrichment, with a single human gate on exceptions.

03 — Where the human stayedAgentic, deliberately not autonomous.

We never aimed for zero humans. We aimed for humans on the decisions that actually carry risk — a supplier whose insurance has lapsed, a bank account that doesn’t match the registered entity. The agent drafts; a person decides. Every approval and override is logged, which gave compliance something the old inbox relay never could: a complete, queryable audit trail.

04 — What it returnedFour hours, and a team that kept tuning it.

Median time to a live supplier fell from nine days to about four hours, and the long tail of stuck packs disappeared because nothing sat in a queue waiting for a free pair of hands. The reviewer who used to chase packs now spends that time on the genuinely ambiguous ones — and on adjusting the validation rules as new edge cases surface.

That last part matters more than the headline number. We left during week six, but the workflow didn’t freeze the day we did. The team owns the rules, can see why any pack was flagged, and changes them without us. The redesign was the easy part; the handover was the point.