How AI mapping suggestions work
ProcuLink uses OpenAI to help close the last mile of mapping — only the lines that deterministic rules could not resolve on their own.
When AI runs
Every incoming order is first put through your deterministic mappings and validation rules. AI suggestions only run for lines that the deterministic pass left unresolved. AI never overrides a successful deterministic match.
Anatomy of a suggestion
Each AI suggestion carries three pieces of metadata:
confidence— a percentage; the review screen reads 85%+ as high confidence, 70–84% as a good match, and anything lower as low confidence.reason— a short, human-readable explanation, e.g. "matched on description plus unit size".provenance— the source data the model was given, so you can audit what it actually saw.
Grounded in the supplier's catalog
If the supplier has a product catalog uploaded, suggestions are catalog-grounded: the model must pick from the supplier's real codes, and any answer outside the catalog is rejected after the fact — a hallucinated code can never reach your screen. Without a catalog, suggestions are best-effort from the evidence on the line, which is exactly why uploading the catalog is the single best thing you can do for suggestion quality.
Accept or enter manually
Nothing is written to the order automatically. Each suggestion appears as a card on the Review screen with two actions:
- Accept writes the suggested code to the line and marks it resolved — and the resolution is remembered as a mapping for next time.
- Enter manually opens the normal code-entry field (pre-filled with the suggestion) so you can type the right code instead. Both paths save to the server the same way.
The monthly AI budget
Every workspace has a monthly AI token budget shared across all AI features — PDF extraction, line suggestions, email-body extraction, and mapping inference. The budget scales with plan, from 200,000 tokens on the Pilot trial to 10 million on Enterprise (roughly 50 vs 2,500 PDF extractions a month).
When the budget runs out, nothing breaks and nothing blocks: AI quietly steps aside for the rest of the month, and unresolved lines simply wait for manual entry on the Review screen. The budget resets monthly.
No-egress workspaces
Workspaces configured for no-egress document processing (an enterprise option an operator enables) never send order data to OpenAI: suggestions, PDF AI extraction, email-body extraction, and mapping inference are all switched off, and unresolved lines always go to human review.
Need help? Email support@proculink.eu or see PO field mapping basics.