Marketing Dash — Daily channel reports without the morning slog

The Problem

The growth team's first four hours every morning went to pulling numbers from six dashboards and writing the same shape of paragraph six times. The bottleneck was not the data — it was the rewriting. By the time the report went out, the day's decisions had already been made on yesterday's gut feel.

The System

A nightly ingestion job pulls from Google Analytics, Plausible, Meta Ads, Google Ads, the CRM, and the email platform into one normalized BigQuery schema. A summarizer pass runs over the normalized rows with the prior seven days as context and produces a structured draft — headline, movements, anomalies, next actions — that lands in an operator console at 6am. The operator edits, approves, and Resend ships it to the distribution list.

The decision that mattered most: the AI does not generate the report, it drafts the review surface. The draft is intentionally a step short of send-ready — every anomaly links back to the BigQuery row that triggered it, and every claim shows its source channel. The operator spends 4 minutes confirming and tightening, not 4 hours writing.

What I built

• Nightly ingest job normalizing 6 channels into one BigQuery schema.
• Anomaly pass using rolling 7-day baselines, not single-day deltas.
• Structured drafter on Claude via AI SDK with a strict JSON schema.
• Operator console showing draft, source rows, and one-click edits.
• Resend distribution with per-recipient channel filtering.
• Audit log of every draft, every edit, every send — queryable by date.

Outcome

"I used to write the same paragraph six times before coffee. Now I edit one draft and send. The numbers I argue with are already on the page." Eighty-two percent of drafts ship with no structural rewrite; the team moved its planning standup from 11am to 9:15am because the report is ready first.

What I'd do differently

I underweighted the operator console in the first two weeks and spent that time tightening the prompt. The model was 90% there on day five, but the team didn't trust an unaudited paragraph — they wanted to see which BigQuery row triggered each anomaly call before they would put their name on the send. If I started over I would ship the source-row drill-down on day three, before the second prompt iteration, and let the model quality catch up alongside it. Legibility built adoption; prompt polish only built my own comfort.