Agent-ready workflow system: evidence to controlled decision
Alex Yang
What this actually is
A working local prototype, on synthetic data, that answers a design question rather than a delivery one: what does an operations workflow have to look like before it is safe to let an agent near it? No live model, no writes, no client data. The constraint is the point — the interesting part is the contract, not the automation.
The problem it takes seriously
Most AI-for-operations demos show a recommendation and hide the reasoning. That leaves the person who has to sign off with two bad options: accept it on faith, or redo the work by hand. Either way the automation has saved nobody anything.
From documents to a decision
The worked case is an invoice exception. Source records are ingested, supplier identity is resolved across its aliases, a three-way match locates exactly where invoice, order and receipt disagree, and the discrepancy becomes a case that carries its own evidence bundle rather than a pointer to somewhere else.
The recommendation is shown next to the trace that produced it, step by step, so a reviewer can disagree with one step without discarding the whole thing. Then a person decides, and the workflow ends there by design.
The decision trace, shown to the reviewer in order. Seven steps, and the last one is a person.
What the interface has to guarantee
Every source record stays attached to the exception — invoice, purchase order, receiving record, contract, and the unstructured note a person actually wrote
Every recommendation carries its decision trace, so a reviewer checks the reasoning rather than the conclusion
Nothing is written anywhere: there is no downstream write path, and no agent acting on the outcome
Every record is labelled synthetic in the interface itself, so a screenshot can never be mistaken for a live system
The review contract: every question a reviewer will actually ask, and the thing in the interface that answers it.
Honest scope
This is an internal, local-only concept on synthetic data. It is not client delivery, it does not represent client approval or production access, and no live model is wired into it. It ships with tests, a one-command verification step and a dependency audit, because a prototype that cannot be re-run is a screenshot. It is here because the design contract is the transferable part — the same one I apply when the data is real.
Like this project
Posted Aug 27, 2026
Independent AI-native system design: structured evidence, human approval, replay, audit, and fail-closed automation. Synthetic data; no live model or writes.
