Hiyaku builds one live model of a business and runs governed agents on it, so AI spend can be priced against what it actually produced.
AI spending reaches $2.59 trillion in 2026. Per-token prices have been falling the whole time and no enterprise's bill has followed them down.
An agentic task can consume a thousand times the tokens of a chat turn, and the same task run twice can differ by thirty. Demand compounds faster than price falls.
of enterprises are already over their AI budget, and only a fifth can tie any of that spend to a business result.
FlySBS had all three in miniature. A charter enquiry is decided within hours, first come, first served. Pricing one took two days — elapsed time spent assembling the cost by hand, line by line, while that window closed.
They were not losing on price. They were arriving after the decision had already been taken.
Tokens are easy to count: they arrive on a vendor invoice, itemised, every month. Outcomes are hard to count, because in most companies they are scattered across spreadsheets, inboxes and the ledger, and nothing knows that a quote, a trip and an invoice are one piece of work.
You cannot divide by an outcome you have never modelled.
So Hiyaku inverts the usual order. Model the business first, give it one governed way to be written to, and only then let agents run on it. Attribution stops being a tagging programme and becomes a property of the architecture.
Every token above is already charged to the enquiry it served.
| Agent | Reads | Writes | Routed to | Tokens in / out |
|---|---|---|---|---|
| 1 · Intake | The WhatsApp message | Enquiry record, confirmed by Operations | Small model | 6.4k / 0.4k |
| 2 · Journey | Leg, Aircraft | Air time, ground time, block time | Small model | 5.1k / 0.3k |
| 3 · Cost | Rate Card, Crew | The cost lines of the quote | Small model | 11.3k / 0.5k |
| 4 · Forecast | Past quotes and invoices | Predicted costs and the price band | Forecast model, one frontier call to explain it | 18.4k / 0.9k |
| Charged to ENQ‑4471 | 68% of input served from cache | Small ×3 · frontier ×1 | 41.2k / 2.1k | |
Customers, enquiries, quotes, trips, legs, aircraft, crew and invoices became one thing the whole company reads. Nothing was ripped out; the ledger stays the ledger.
People and agents go through the same governed door. Every AI call lands against a customer, a trip and a cost, not a line on a bill nobody can allocate.
The assistant proposes and a person confirms. Full autonomy stays a decision the business takes deliberately. An unbounded agent loop is an unbounded bill.
Leg and ground costs are forecast from FlySBS's own history, with a price band drawn off quotes won and lost, and the model retrains on what each trip actually burned. Trained on their data, not guessed at by a frontier model.
Two days of elapsed time became under fifteen minutes, with no estimator tied up for any of it.
Back-tested on FlySBS's own history before it priced a live one.
A decision window measured in hours, answered by a quote measured in minutes.
Every trip closes on real fuel and handling, so revenue per block hour is measured.
About these figures. The four above are the engagement's targets, back-tested against FlySBS's own history — not results claimed from a finished deployment.