Every AI demo works. The examples are chosen, the inputs are clean, and the person running it knows what to avoid. The question a business needs answered is different. Will this work on our documents, on a busy Monday, with the messy cases included?
The way to answer it is an evaluation, a test set built from your real work.
Build the test set first
Before building an automation, we collect real examples, typically dozens to a few hundred depending on the task. For document extraction, that means actual invoices, forms, or tickets. For drafting, it means real requests and the replies a person would consider correct. We deliberately include:
- Typical cases, the bulk of the work.
- Messy cases, such as faded scans, handwritten notes, and unusual layouts.
- Edge cases, the ones that require judgment.
- Adversarial cases, inputs that try to trick the system, where relevant.
Define what “correct” means
For each example, we write down what a good result looks like. For extraction, that’s the correct field values. For classification, it’s the right category. For drafting, it might be a checklist covering the facts included, the tone, and the format. Vague criteria produce vague evaluations.
Measure, then decide
Running the system against the test set gives real numbers, such as accuracy by field, the share of cases routed to a person, cost per item, and time per item. Those numbers drive decisions:
- Is it accurate enough to save time after review?
- Which cases should always go to a person?
- Would a different model or approach do better at acceptable cost?
Keep running it
An evaluation isn’t a one-time gate. Prompts change, models are updated, and document formats drift. We rerun evaluations whenever something changes, which catches regressions before users do.
The payoff
Evaluation turns “the AI seems good” into “the AI extracts these six fields correctly in 94 out of 100 real cases, and flags the rest.” That’s a statement a business can make decisions on, and it’s the standard we hold our automations and agents to.
Written by Syntonyl team. This article is general information, not legal, tax, or financial advice.