Details for the curious
Open any section for the full detail — figures, design decisions and how it fits your business.
The challenge
Every order was keyed in by hand. Each one took about an hour of skilled effort before it even reached the queue.
- Slow: about 60 minutes of skilled effort per order.
- Lost margin: orders arrived with a net total, while the price breakdown sat in an old quote.
- Costly errors: a wrong billing entity or fee code came back later as an invoice dispute.
- Hard to scale: more volume meant more people, and the rules lived in operators' heads.
How it is designed
AI for judgement. Rules for money. Models read any document, rank the right customer account and explain their reasoning. Pricing, tax and fee logic is deterministic and testable, so a model never decides a price.
Humans approve everything. Every step offers Approve, Correct or Regenerate. There is no unattended mode. Each order posts once, and only a real ERP order number counts as success.
How it works
Four agents, eight steps, one approval.
Proven on a hard order
A 7-page, image-only scan worth $86,149.80: 229 units, 7 delivery batches over 13 months, 5 components per unit — handled end to end.
| In the order | What the agents did |
|---|---|
| Image-only scan, no text | OCR with confidence-based retry, every line read |
| European numbers such as 418,00 | Read by convention: 418.00, not 41,800 |
| Net price only, no breakdown | Matched to the quote, component sum rebuilt |
| 229 units, 7 delivery dates | Split into 7 lines, one per delivery date |
| Mandatory items not on the order | Gap detected, items added by rule |
Measured · 13 of 13 sample orders correct · 154 fields checked against the source.
What it is worth
Illustrative, on your volume. The 60 to 6 minutes is reported; volume and cost are planning assumptions (54 minutes saved per order at $45 an hour).
Labour only · the larger value is likely faster delivery and fewer disputes.
Where it applies
The same pattern fits any business that receives orders by email, Excel, PDF or portal and prices them against contracts — for example fuel and lubricant suppliers serving fleets, stations and industrial customers, with surcharges, exemptions, credit limits and scheduled multi-site delivery.
