Automation for marketing platforms: the Monday report is ready by Monday morning

In a marketing platform and in an agency most manual work hides in reconciliations and reports: export from the ad accounts, compare, build a table, send. We turn that into workflows that pull data on a schedule, compare by rules, post to the chat on a discrepancy and assemble the report with no human input — the way it runs at a performance marketing agency.

What we automate

• Metric reconciliation: impressions, clicks, spend, conversions from the ad accounts versus the platform for the same period and attribution — on a schedule, with an alert when a discrepancy exceeds the threshold.
• Single lead intake: forms, email, Telegram, website — into the CRM tagged with source, campaign and owner, with no manual entry.
• Reports: weekly and monthly assembled from ad accounts, the CRM and spreadsheets at a set time and sent to the lead and to clients.
• Alerts: spend above plan, campaign stopped, token expiring, conversions dropped — to the team chat at once.
• Token and integration control: scheduled checks of ad account access, notification before expiry.
• Integration regression: a request collection against the ad account APIs runs after their updates, the report goes to the chat.

What it looks like in practice

At a performance marketing agency a money and data control system was built: spend sync across all accounts several times an hour plus a deep nightly sync, reconciliation of spend against crypto and card top-ups for every account to the cent, a daily MMP export and a client bot with balances. In total, 16 jobs on systemd timers, and any failure goes to the chat as an alert. The defects found cost money: the panel lost part of the spend because of two rows with one key, and the MMP export doubled registrations by adding server and SDK events. Stack — Python, FastAPI, SQLite, pytest, Podman.

Reconciliation rules — the main thing to agree

1. Fix for every source: attribution window, time zone, update delay, rounding, status filters.
2. Set the discrepancy threshold above which the workflow raises an alert — separately per metric.
3. Run the reconciliation in parallel with the manual one for two weeks and compare what the workflow found with what a person found.
4. Switch over; review thresholds monthly against accumulated alerts.

Without agreed rules a reconciliation either stays silent or shouts about every one-percent discrepancy. So the rules are an annex to the contract, not a setting in the code.

Guard rails

Workflows read data and write reports but do not manage campaigns: stopping, changing bids and budgets stay with people. Access to ad accounts — separate read-only tokens. The language model is used for classifying leads and drafting reports, not for decisions. Every workflow logs its actions and reports errors to the chat.

Implementation

We map which reconciliations and reports the team does by hand and how often; we measure the time. We build the first workflow — usually lead intake or spend reconciliation — in one to two weeks, two weeks of parallel runs, switch-over. Then one workflow every two to three weeks. Automation is a separate contract as development.