Quick answer: Calculate the numbers with deterministic queries or spreadsheet formulas. Use AI to explain the already-calculated results and suggest questions—not to fabricate totals from incomplete data.
Example: a four-person service team
A team tracks enquiries received, appointments booked and jobs completed. Its weekly report should state the date range and time zone, define each metric and show where data came from. An AI summary can identify that conversion fell, but it should not assert a cause without supporting evidence.
| Metric | Definition | Quality check |
|---|---|---|
| New enquiries | Unique first inbound requests this week | De-duplicate by enquiry ID |
| Booked appointments | Confirmed appointments created this week | Exclude cancelled requests by rule |
| Completion rate | Completed jobs / scheduled jobs | State denominator and dates |
| Median response time | Median of first reply minus first enquiry | Exclude negative and missing timestamps |
A trustworthy workflow
Export or query data with an authorized read-only integration. Validate row counts, missing values and timestamp boundaries. Calculate metrics in code or a transparent formula. Pass a small, validated JSON object to the writing step with a prohibition on inventing figures. A manager then checks the summary against the chart and approves distribution.
Use a neutral instruction
Summarize only the supplied metrics. Clearly distinguish observed change from a hypothesis. If a comparison period is missing, state that the comparison is unavailable. Do not infer causes or invent benchmarks.
Failure testing
Test incomplete exports, duplicated rows, a week with zero appointments and date changes caused by local time zones. The system should show “not available” where division would be undefined, rather than presenting a dramatic percentage.
Cost control
Keep a short model input: aggregated metrics and a few anonymized examples, not a full customer database. Review whether the written summary is more useful than a basic dashboard before paying for an AI step.
FAQ
Can AI produce the calculations?
A model can suggest formulas, but the production calculations should be executed and tested by deterministic software.
References and further reading
These references provide context and official guidance. The sample workflows above are illustrative designs, not vendor performance tests.
Published by: FlowHarbor Editorial Desk. Read our sourcing and review policy.