Field guide
An AI workflow needs a way to check its work
Start with a real task, an answer you can verify and a clear boundary for the assistant.
Start with an observable task
“An assistant for our documents” is a direction. A task is more concrete: find the applicable policy, identify the supporting passage, and prepare a draft a person can verify. The task determines what data, permissions and evaluation examples are needed.
Write down the input, expected output, who checks it and what happens when the evidence is missing. That creates a useful boundary for a first version.
Check three kinds of work
| Check | Useful evidence |
|---|---|
| Answer quality | A small reviewed set of representative tasks, including questions the documents cannot answer. |
| Source support | The passage or record that supports each consequential claim. |
| Operational outcome | Time, corrections, cost and failure handling measured in the actual workflow. |
A fluent response can still require substantial correction. Measure the full path to an accepted result, including review time.
Separate a draft from an action
An assistant can prepare a response or a proposed update before it is allowed to send or apply one. Human review is especially useful when the action affects a customer, money or an external record. Preserve the evidence and a way to recover from a mistake.
The NIST AI Risk Management Framework ↗ is a useful reference for thinking about evaluation and the setting in which an AI system will operate.
Publish what was measured
A case study should state the task, baseline, sample, measurement period and remaining review burden. Until those exist, a demonstration establishes that a workflow can run; it does not establish a customer’s savings or revenue gain.
This series will use that distinction when discussing our own products and experiments.