Measured before it goes live

Anyone can demo an agent on a good example. The question is what it does on fifty of yours, including the ugly ones. So that is what happens before anything is written into your systems.

The contract

First, what correct means is written down: which fields the job must produce, and what counts as a match for each one. Dates, amounts and names can be written in several ways that are all correct, so the rules of comparison are part of the agreement. A person approves the contract before a batch is paid for.

The truth

You confirm the correct answer on a batch of your own files. Where a value is genuinely not on the document, you say so explicitly. That sounds like a detail and it is the difference between a measurement and a guess: an empty field nobody looked at is not evidence of anything.

The score

The platform scores the job against your answers, field by field, and splits the misses into three: a mistake of the program, something nobody could tell from the document, and a difference inside the agreed tolerance. You see the misses with the file next to them.

The gate

A number is agreed for the job. Below it the job does not get switched on. That is a lock in the platform, not a promise on a call. When the number is not reached, the job is tuned and measured again on the same batch.

The sample

A batch of only clean files proves nothing. The platform checks what kind of sample it was given and says so, so a comfortable score cannot be produced by choosing comfortable documents.

After it is live

The job keeps being measured. A shadow copy of what the agent wrote can be compared with what people later corrected, which turns human corrections into the honest weekly answer to “how often was it wrong”.