The review starts before the meeting.
A lean e-commerce team prepares its weekly review by exporting advertising data, combining store reports and copying figures into a brief. Several people repeat the same checks, yet a missing day or inconsistent currency can still reach the meeting.
This working example tackles one bounded part of that process: transform two stores’ daily aggregates into a validated set of metrics and a packet suitable for an AI-assisted draft. It deliberately keeps data validation, calculation, drafting and commercial approval separate.
A brief starts with reliable inputs.
12–18 October 2026 · EUR · UTC · synthetic data · runs locally
Choose an input and run the checks to inspect the result.
Inspect the input dataset
The complete fixture produces €42,000 in net revenue, €12,500 in ad spend and €5,990 in contribution after advertising and the illustrative fee, before overhead. The invalid fixtures stop before metrics are published. The same functions power the downloadable local version.
Stop errors before they become commentary.
| Control | What it checks | What it cannot establish |
|---|---|---|
| Coverage | One row for each declared store and each day in the reporting window | That the source system exported every underlying order |
| Uniqueness | No duplicate store-day keys | That two different source identifiers do not represent the same business event |
| Dates & currency | Valid dates inside the window and EUR throughout | That upstream timezone or currency conversion was performed correctly |
| Amounts | Finite non-negative amounts and explicit fee assumptions | That a cost input is complete or economically appropriate |
| Calculation | Versioned formulas, integer-cent aggregation and defined zero-denominator handling | A causal explanation for a change in commercial performance |
The input contract uses net sales after discounts and refunds recorded in the period. A negative net-sales day is held for a separately reviewed schema rather than silently coerced to zero. Later returns, tax treatment, payment fees and cost coverage require business validation before relying on the result.
In an implementation, data checks are paired with source reconciliation and exception ownership. A green validation result means the declared checks passed; it does not establish that the underlying reports are correct.
Give the model facts, boundaries and a reviewer.
Once the figures pass validation, the prototype prepares an AI handoff containing the verified metrics, source references and drafting constraints. This packet is visible in the demo output and included in the export. It is prepared input for a model, not a generated model response.
Only the approved evidence.
Pass the aggregate metrics and definitions required for the brief. Exclude credentials, customer-level data and unrelated operational records.
Observations before explanations.
Ask for observations tied to metric keys, separate hypotheses and open questions. Do not permit invented causes, client results or unsupported forecasts.
A person owns the decision.
Check the draft against the figures and source material. Campaign changes remain outside the workflow until an accountable person approves them.
The public example contains no inference step. Connecting a model in production requires a chosen provider, reviewed data-processing terms, output evaluation and handling for unsupported claims or malicious source instructions. Drafting constraints reduce ambiguity; they do not guarantee model behaviour.
What a real implementation adds.
A delivered project would connect to the agreed store and advertising sources using appropriately restricted access. It would define source mappings, currency conversion, reporting windows, schedules and how to reconcile changes after a reporting cutoff.
It would also cover retries, failure notifications, access controls, provider costs, approved retention and a named functional owner on the client side. The handover would include operating instructions and tests the team can rerun when a connector or metric changes.
| Test | Expected behaviour |
|---|---|
| A complete approved period | Metrics reconcile to the agreed sources within the documented tolerance |
| Missing, duplicated or mismatched input | Stop the report and route an actionable exception to the operator |
| An AI draft includes an unsupported figure or explanation | Reject or correct the draft; no publication or budget change |
| A source or provider is unavailable | Keep the last approved report distinct from the failed run and notify the owner |
| A team member takes over | Run, inspect, rerun and recover the workflow using the handover documents |
The assignment is an integrated process with an operating owner. Its usefulness should be measured against the actual preparation time, error rate and review quality of the existing process; this prototype claims no achieved time saving.
Take the example apart.
Download the dependency-free engine, command-line runner, four synthetic fixtures, expected output and automated tests. These files contain only illustrative material.
Version 1.0 · Prepared 29 October 2026 · Reuse the method with your own reviewed definitions. Examples are provided for evaluation, without a guarantee of commercial results.