Compare economic growth with category trade.
Poland’s GDP grew 3.03% while selected category imports fell 7.83%. Keep the conflicting evidence.
Inspect the country comparison ↗WHITEBOXX / RESEARCH & ENGINEERING
Seven analytical tasks
for enterprise AI agents.
Compare markets, monitor filings, reconcile revenue and investigate business performance. These examples show the inputs, operations and results behind seven execution-agent assignments.
Start with the analysis you need.
Give an execution agent a specific task, the data it can use and the output you need. It might compare observations, calculate a revenue movement or find what changed since a previous review.
A reactive task begins with a question or brief. Proactive work begins with a schedule or event, such as a new filing or an exception in a business record. In either case, the agent needs an analytical assignment and the inputs to carry it out.
These seven analytical jobs illustrate one set of Evidence OS capabilities. The broader platform also provides data and workflow services that can be extended into enterprise applications. Professional persona agents play a smaller, reactive role: exploring a proposal from a specified perspective.
WORKING REFERENCE
The task determines the output.
| Job | Inputs and operation | What the team receives |
|---|---|---|
| 1. Compare candidate markets | Align countries, measures, units and a common period; compare category imports alongside broader economic activity. | A comparison with conflicting signals exposed, rather than an unexplained country score. |
| 2. Track a company filing | Capture successive filing inventories; compare accession identifiers and metadata. | A dated change set distinguishing additions, revisions and removals. |
| 3. Reconcile revenue movement | Compute new, expanded, contracted and churned recurring amounts at account grain. | An exact bridge from opening to closing MRR, with a residual check. |
| 4. Locate concentration | Group a measure by a relevant dimension before ranking its contribution. | The accounts, categories or geographies that explain the measured total. |
| 5. Compare like-for-like periods | Select standalone quarters; exclude year-to-date totals and retain filing dates. | A time series with comparable periods and traceable revisions. |
| 6. Test a suspected explanation | Join account-level support aggregates to revenue changes; look for counterexamples. | A narrower investigation that separates association from a demonstrated cause. |
| 7. Turn a finding into follow-up work | Carry evidence identifiers, unresolved questions and proposed next checks into a workflow. | An investigation queue or reporting task with a defined owner and acceptance condition. |
WORKED EXAMPLE
Three examples with downloadable calculations.
One new filing, observed.
Two SEC captures, 7.66 minutes apart, found one added accession. Detection and interpretation are separate jobs.
Inspect the capture pair ↗Check the revenue calculation.
A naive support-ticket join turns a 10% decline into 12%. The account-level bridge exposes the error.
Inspect the bridge ↗| Job | Inspectable output | Meaning |
|---|---|---|
| Locate concentration | 65.89% | Germany’s share of gross import value summed over these three countries only; not a global market share. |
| Compare like-for-like periods | $11.345bn vs $10.236bn | Salesforce standalone revenue quarters ending July2026 and July2025, with actual comparative filing dates retained. |
| Test an explanation | Aster +$300; Cedar −$500; three tickets each | An observed counterexample to ticket count alone explaining revenue movement. Records are synthetic. |
| Route follow-up work | Read the new 8-K; review Elm; resolve product classification | Proposed next investigations with acceptance checks in the download. No tasks were created or workflows executed. |
The seven examples use reference calculations and proposed follow-up tasks. They do not represent production-agent runs.
Use the calculation to support the recommendation.
Whiteboxx’s statistical query code supports finding measures, ranking regions, examining distributions, comparing geographies and reading time series. These operations provide the data and calculations for an investigation. A recommendation also needs an interpretation that fits the business question.
The examples combine these analytical patterns with reproducible calculations. We collected public data directly from publisher APIs and used labelled synthetic records for the revenue diagnostic. These are reference analyses, rather than records of production-agent runs or client results.
Choose the first assignment around an existing review.
Start with something your team already needs: a weekly exception list, market comparison or reconciled reporting view. Agree the entities, periods and measures, then define what would cause the next report to change.
Keep the earlier inputs and results so each update can explain the difference. Additional sources, scheduling and system integrations can then be added as the workflow requires them.
Use the operator reference.
A populated reference with the seven jobs, their inputs, operations and acceptance checks.
Seven-job reference (CSV) ↗Seven worked outputs (JSON) ↗