Export, accounting practice and F&B look nothing alike on the surface. Underneath they run on the same parts: read the input, draft the work, stop for a person, then record what happened. This is that shared layer.
A run moves through five steps
This is the contract every agent in the system follows. There is no shortcut past a step.
- 01
Read
Take the input and tie each field to its source
- 02
Compare
Check it against what you already hold and against your policy
- 03
Draft
Write the draft, with the reasoning behind each number
- 04
Self check
Look for the weak points in what it just wrote
- 05
Revise and queue
Fix them, then join the approval queue. The draft stops here
Needs approval
One runtime, three industries standing on it
This is the shared layer the page opened with, drawn side-on.
Export, accounting practice, F&B: each with its own vocabulary, forms and documents
differsSix stages, one or two of which stop for a person
same shapeRead, compare, draft, self check, revise. Every field keeps a path back to its source
sharedWorkspace boundaries, the approval gate, the trail, the spend ceiling
sharedFour things that run throughout
Fields keep their source
Every value holds a path back to the document or record behind it. No source, no value.
The approval gate
Anything customer-facing stops for a person. Checked at the API and again in the database.
Policy gates
Spend capped per workspace, and the interface says which mode it is in at the cap.
The trail
Which step the machine did, which a person did, how long, what it cost.
Why one runtime rather than three products
The expensive part of an AI system that survives real operations is not the model. It is data boundaries, approvals, evidence and honest states. Paying for that once and reusing it for the second and third industry is how a small team covers three.
See it on your own work
Pick the process eating the most hours. We build that one and run it alongside for a month.