If you’ve ever onboarded a new supplier into a procurement tool, you know the friction. Their price list arrives as a PDF (or a CSV with mystery columns). Someone in your operations team spends a half-day mapping their SKUs to your standing-order list, cross-checking units, normalising pack sizes, and then it’s “done” — until they send a new list next month and the process restarts.
Most procurement platforms make this faster than email-and-spreadsheet. None of them have made it autonomous. So we built that.
The job to be done
A new supplier sends you their catalogue. You want, within a minute:
- A clean, structured list of SKUs in the platform.
- Units normalised so they’re comparable to your existing supplier prices.
- Allergens + nutrition tags inferred where possible.
- Matches to your existing standing-list items flagged for confirmation.
- Live in the Marketplace, side-by-side with their competitors, on the same screen.
The architecture
Five layers, executed in parallel where possible:
- Inbound — supplier emails the catalogue to
suppliers@yourvenue.inntally.com(or drag-drops into the buyer dashboard). - Extraction — AWS Textract reads PDF tables; CSV / Excel parsed natively; line items pulled into a normalised intermediate schema.
- Normalisation — GPT-4o-mini handles unit conversion (case to each, kg to per-100g, ml to per-litre), spelling tolerance (“Tomatos” → “Tomatoes”), and pack-size inference.
- Matching — embedding-based similarity match against your existing standing list; confidence-scored suggestions presented to the buyer.
- Catalogue write — new SKUs added to the supplier’s Marketplace catalogue; price feed scheduled for daily refresh.
Why the buyer is still in the loop
The system is “autonomous” but not unsupervised. The buyer sees:
- The proposed catalogue (every SKU pre-filled).
- Suggested matches against existing standing-list items (e.g. “This looks like your existing chicken-thigh SKU from BWG — confirm?”).
- Confidence indicators on unit conversions + allergen inference.
- An “approve all” for high-confidence sections; line-level review for low-confidence ones.
That keeps a human in the audit chain, which matters for procurement governance + dispute resolution.
What we don’t do (yet)
The areas the buyer still does manually:
- Negotiating the contract pricing tier with the supplier (the system reads the price list given; doesn’t negotiate the price).
- Confirming delivery schedule + minimum-order rules (we read what’s on the catalogue; the supplier still tells you their terms).
- Approving the supplier’s payment terms + KYC inside the payment processor (deliberate — this is a money-movement decision, not an automation decision).
The result
A new supplier’s catalogue is in the Marketplace within minutes of arriving in the inbox. The buyer’s job moves from data entry to approving + negotiating — the part that actually requires judgment.
This is the “agentic AI” pattern that’s genuinely useful in procurement: do the mechanical work; show the buyer the consequential decisions; keep an audit trail so the operator can review or roll back.
Read next
- Marketplace product page — how price-compare works once your suppliers are in.
- AI invoice extraction — the same pattern, applied to invoices.
- IntelliFlow — the technical engine behind the AI extraction.
“Illustrative scenarios based on industry benchmarks and our pilot rollouts. Named case studies available under NDA on request.”