Build an AI Reporting Agent That Explains KPI Changes—with Evidence

An AI reporting agent can turn approved metrics into a useful account of what changed. It should calculate differences through deterministic operations, inspect the relevant segments, and support its explanation with retrievable evidence.

Begin with one recurring report. Define the audience, metric, comparison period, delivery schedule, and conditions that should prevent publication. A scheduled query plus a summarization step may be sufficient; add agent-driven investigation only where follow-up questions improve the report.

Work through a small example

This fictional sales report compares two complete periods using the same definition and USD amounts:

Region	Previous period	Current period	Change
North	6,000	5,000	-1,000
South	4,000	3,500	-500
Total	10,000	8,500	-1,500

Total sales declined 15%. North contributed 1,000 of the 1,500 decline, or two-thirds in this example. North's own decline was about 16.7%; South's was 12.5%. Contribution to the overall change and a segment's growth rate are different calculations.

A useful explanation is: “Sales decreased 15% across the two complete periods. North contributed most of the absolute decline. The report does not establish the cause.”

An unsupported explanation is: “The new pricing policy caused customers in North to leave.” The table contains no evidence about pricing, churn, or causality.

Build the report in stages

1. Query the approved metric for explicit periods and permitted segments.
2. Check freshness, coverage, and definition consistency.
3. Calculate absolute and relative changes using application code or SQL.
4. Identify segments that explain the arithmetic change.
5. Retrieve permitted supporting context where useful.
6. Draft a narrative that distinguishes observations from hypotheses.
7. Validate the report and deliver it only to authorized recipients.

Store the query reference, metric version, data refresh time, period boundaries, and calculation inputs. A reader should be able to inspect the evidence behind a material claim.

Do not turn a data outage into a business alert

Suppose the last day of the current period is missing because ingestion failed. A comparison against a complete previous period would exaggerate a decline.

The agent should check the data contract before interpreting the result. Depending on the report, it can wait, notify the owner that data is incomplete, or compare matched partial periods with a clear label. Do not silently change the comparison just to produce a report on schedule.

Also handle a zero comparison value. Percentage growth from zero is not a normal finite percentage. State the absolute movement and explain the baseline instead of inventing a growth rate.

Keep the investigation bounded

If the agent can choose follow-up queries, restrict supported dimensions and set limits on time, query cost, and number of investigations. Repeated slicing can uncover coincidental patterns, especially with small groups.

Ask it to distinguish a descriptive contributor from a causal explanation. Region can explain where the arithmetic decline occurred; understanding why may require additional data, a controlled experiment, or an analyst's investigation.

Use semantic definitions for repeated calculations and access controls for both querying and delivery. A report run under an internal service identity must not be distributed more broadly than its contents permit.

Evaluate the report before automating delivery

Include fixtures for the table above, no change, a zero baseline, a missing period, and a segment whose increase offsets another segment's decline. Verify that segment contributions reconcile to the total and that the explanation does not invent causes.

Score numbers, periods, evidence, and narrative separately. Initially review the report before sending it, and assign an owner to unresolved data-quality issues. Measure review time as part of the operating cost.

Build the underlying interface with our AI data-agent guide. Labs4Change can help implement an evidence-based reporting workflow around the KPIs your team already uses.