From intent to invoice line

The problem

When buyers add a free-text line to a purchase requisition, they write whatever feels natural — "laptop for the new hire", "catering for offsite", "some cables". Those informal descriptions caused two recurring problems: approvers lacked enough context to approve confidently, and the line landed in the wrong category, mis-routing it to the wrong approval chain or GL account.

• Vague descriptions blocked approvals — approvers sent lines back for clarification.
• Wrong categories mis-routed requisitions and created reconciliation work for finance.
• Buyers didn't know how to categorise items correctly — the category tree had 400+ nodes.

The approach

The trigger for AI assistance is the moment a buyer finishes typing their free-text description and moves to the next field. At that point, the model has enough signal to both enrich the description and suggest the most likely category. The design challenge was making the AI feel helpful rather than intrusive — it suggests, the buyer confirms.

• AI enrichment fires on field blur, not on every keystroke — no jarring mid-typing rewrites.
• Suggested description shown inline as editable text, pre-accepted but overridable.
• Category suggestion surfaces with a confidence indicator and one-click alternative options.