Process Orchestration: Error Cost Drives the Model

RFQ-to-quote pipeline: a request PDF enters as input, fields are extracted, values are checked against a reference catalog, questionable matches go for manual review, and the output is a draft quote in Excel.

Every step is named, and the checkpoint is explicit.

The catalog case is even stricter: the document passes through two independent parsers, and a value enters the result only if both agree.

When there is a discrepancy, the row is marked UNMATCHED with the source page indicated - the operator sees the exact location of the mismatch and corrects it manually.

The author priced the first 25 items at a fixed $150 - the price is for a verified row, not the volume of output.

The API-limit case is handled the same way: a queue, controlled concurrency, a request, and on a 429 response, reading the Retry-After header, pausing, then retrying with exponential backoff and jitter so workers do not wake up simultaneously.

In all three, the data chooses the path at runtime, but an engineer defined which paths were possible in the first place.

TTU - time to usable result - measures the interval from input to a value you can trust.

The elegance of the diagram does not factor into this calculation.

A guessed value that looks correct costs more than an empty cell because someone further down the chain may trust it, send the client the wrong price, or enter the wrong SKU into ERP.

The catalog case author puts it plainly: a guessed price that looks plausible is worse than a visible blank field.

The same applies to the API limit: if the model decided at its discretion whether to retry instead of using fixed backoff, one provider failure could become a cascading queue problem, requiring a real-time fix to the model's decision rather than to a specific graph step.

By default - a deterministic graph.

A dynamic branch is added where the data must choose the path at runtime.

The agentic step is reserved for narrow judgment within a graph that remains deterministic overall.

Autonomy as a sellable feature often means removing the person who was previously accountable for the outcome. RFQ-to-quote and catalog-to-CSV look simple precisely because most decision points were taken away from the model and handed to verifiable gates.