Weekly client reports where every number is recomputed before it ships.
Raw exports and your brand kit in. A your-branded weekly report (PDF and web page), a monthly deck and a draft client email out. AI writes the insights; a separate checker proves every figure against the raw data.
- Inputs
- Meta, Google Ads, TikTok, GA4, Shopify exports (CSV)
- Outputs
- Weekly PDF + web page, monthly deck, client email
- Number check
- 1,026 facts recomputed, 0 mismatches
- Text check
- 1,069 numbers traced in HTML, email and PDF, 0 problems
Weekly report · 4 pages
Monthly deck · 16:9
How a number gets into a report
The AI never types a number.
Wrong numbers in front of a client cost more than a late report. So the system is built so the writing model cannot invent one.
- Ingest. Each platform export is read into one clean table. Odd columns and gaps are logged.
- Facts. Code computes every metric, change and target gap into one facts file.
- Insights. The AI writes the story using tokens that point to facts, not typed figures.
- Render. The facts are filled in and the report is drawn in the agency's brand: PDF, web page, deck and email.
- Check. An independent script recomputes every fact from the raw exports and traces each number in the final text. One mismatch and the report does not go out.
The demo also found real problems a person would want flagged: in this sample week, TikTok reported no purchases for two days despite spend, and the report says so.
About this sample. Harbor & Pine Digital and Kindred Roast Co. are invented. The exports are generated sample data. The pipeline, the checks and the outputs are real. The live API connectors for Meta and Google Ads exist but have not been tested on a live account yet, so a pilot starts with CSV exports.
Next step
Send me last week's exports for one client.
I'll turn them into a report in your brand, with the number check attached. Book 15 minutes and we'll pick the account.