01The nightmare input
Real-world handwritten haulage tickets — smudged, half-legible, scrawled in a cab. A nightmare for AI and humans alike. Click any docket to enlarge.
02The invoices
The supplier's side is pristine: three clean Vale Aggregates invoices, neatly typed. Everything looks correct — until you check it against the dockets. Click to enlarge.
03Reconciliation result
Every invoice line read off the dockets and checked against the record. Rate £214/load. Here is what doesn't add up.
| Invoice | Ticket no | Material | Loads | £ | Status | Why |
|---|
04The fail-safe
The trust moment — what the system does when it genuinely cannot read a field.
Two dockets are on record — but the vehicle reg, material or date is genuinely illegible. The system refuses to guess and routes both to a human as requires‑review. It never invents a number it can't read — the same discipline your yard team uses when the writing won't come clear.
Every governed decision — approved, flagged or held — leaves a tamper‑evident receipt: what was read, what couldn't be, and who it went to.
05Ask the data
Ask in plain language; the answer is grounded in the reconciled record. Pick a question.
In the live system this answers any free-text question against your real data.