# Define a business metric before testing AI answers

> Agree on one metric, connect an approved data slice, and evaluate twenty questions against analyst-verified answers.

- Canonical: https://gptnavi.com/workflows/business-metric-to-evaluated-ai-answers
- Category: Operations
- Difficulty: Advanced
- Setup time: 4-8 hours
- Estimated time saved: Measure in a pilot; savings depend on the task
- Last materially updated: 2026-10-22
- Best for: Enterprise data teams
- Tools: Notion, AtScale for AI, Google Sheets

## Quick answer

AtScale for AI is aimed at governed enterprise metrics. This is a small evaluation, not a full platform rollout: choose one metric and use a supported data source. Data engineering and vendor access are prerequisites; teams without a warehouse can begin with a metric dictionary and manual answers.

## When to use it

- Revenue-definition consistency
- Internal metric question answering

## Steps

1. **Approve one metric definition** — Ask the metric owner to define formula, source, grain, currency, time zone, exclusions, and reporting period. Include examples such as refunded orders and late records, then sign off the definition. Tool: Notion. Expected output: An approved metric contract
2. **Model an approved data slice** — With the data team and AtScale, confirm supported connectors, access restrictions, and the appropriate modeling and computation capabilities. Implement the metric over a small approved slice and reconcile totals with the existing report. Tool: AtScale for AI. Expected output: A reconciled metric model
3. **Prepare independent questions** — Write twenty questions covering normal totals, comparisons, ambiguous dates, empty slices, and unauthorized dimensions. Have an analyst calculate expected answers, units, and permitted access independently of the model. Tool: Google Sheets. Expected output: A labelled evaluation set
4. **Evaluate and fix the source of errors** — Evaluate the selected AtScale answer path against the fixed questions. Compare exact figures, period interpretation, clarification behavior, and access. Assign failures to metric definition, data quality, permissions, or question interpretation; rerun after changes and retain the previous baseline. Tool: AtScale for AI. Expected output: An evaluation report with owners and a baseline

## Prompt templates

### Prompt to adapt

For [metric definition] and [question], identify the requested period, grain, units, and access scope. Ask for clarification when these are missing. Use only approved computed values [results]; cite the metric version and do not invent missing figures.

## Common mistakes

- Changing the metric formula to make an answer look correct
- Testing only easy aggregate questions

## Related workflows