# How AI Reports Work How a Cogny report is produced: the template and context that shape it, the bounded analysis loop over your warehouse and integrations, the data-quality checks, the output structure, scheduling and delivery. **Author:** Cogny Team **Published:** 2025-02-19 **Updated:** 2026-10-05 **Canonical:** https://cogny.com/docs/ai-report-generation-how-it-works A Cogny report is a scheduled analysis of your data that ends in recommended actions. This page explains what happens between the schedule firing and the email arriving, without the implementation detail. For which reports to turn on and how to configure them, see [Reports and Templates](/docs/reports-and-templates). ## What shapes a report Every run starts from three inputs: - **A template or your own instructions.** More than 50 role-based templates cover paid media audits, performance signals, creative and audiences, executive and strategy, and data and tracking. Each is a structured, multi-phase analysis plan with the thresholds and checks a specialist would apply. A custom report replaces the template with what you write. - **Your context.** The industry, company and competitor documents, the growth strategy, the ideal customer profile, your core metric definitions, any custom instructions on the report, and the organisational memory of what earlier runs found. The agent can also browse and search the context tree during the run. - **Your data.** The workspace's BigQuery warehouse (including the GA4 and Search Console exports and anything you loaded yourself) and every connected integration, through the same tool permissions you set in Settings. ## The analysis loop The agent works through the template's phases as a bounded loop of reasoning and tool calls. It inspects which datasets and tables exist, reads schemas, writes and runs queries, calls integrations for platform data the warehouse does not hold, and, where the template allows, searches the web for context. Each result feeds the next step, so the analysis adapts to what the data actually shows rather than following a fixed script. The loop has a hard cap on the number of steps, so a run cannot spiral. Two properties matter for trust: - **Queries are traceable.** The SQL and the tool calls behind every finding are kept with the report, so a sceptical analyst can check any number. - **Reads are automatic, writes are not.** A report can read anything it has permission to read. It does not change campaigns, budgets or pages; it recommends, and the [ticket system](/docs/growth-tickets-architecture) handles execution with approval. ## Data-quality checks Before drawing conclusions the agent verifies what it can: whether conversions are firing, whether tracking looks duplicated or broken, whether a source stopped updating, whether a number contradicts another source. What it could not verify is stated in the report as a data-quality note rather than silently assumed. Several templates exist specifically for this (Data & Tracking Quality, Connection & Setup Verification). ## What a report looks like Every insight block follows the same shape: a summary, the underlying data and charts, data-quality notes, and recommended actions for the week. Findings are expressed against your core metrics when you have defined them, so an SEO report talks about organic orders rather than clicks. Reports are stored in the workspace, can be exported, and are emailed to the recipients you select. ## Organisational memory At the end of a run the agent updates a running summary of what it has found and recommended. The next run reads it, which is how reports avoid repeating last week's insight and can say "the fix from two weeks ago is now visible in the data". You can read and edit this memory under Settings → AI Knowledge → State. ## Scheduling and isolation Reports run on a schedule you set (daily, weekly, monthly or a cron expression) or adaptively, where the agent chooses its own next run based on how much has changed. Each scheduled run executes in its own isolated environment, so a slow or failing report cannot affect another, and Run Now gives you an immediate run without changing the schedule. Runs are queued and retried on transient failures. ## Privacy With Cogny Shield enabled, personal data in tool results is masked before it reaches a model, and instruction-like text in data coming back from an integration is quarantined so it cannot steer the analysis. Your data stays in your workspace's own cloud project in the EU. ## Models Cogny does not depend on one model. It continuously tests and evaluates the best model for each purpose in the workflow, including report analysis, on its own benchmark of marketing-analytics problems built on synthetic data, and routes each step accordingly. Workspaces that need data to stay in Europe can run on EU-hosted options. ## Related - [Reports and Templates](/docs/reports-and-templates): which reports to enable and how to configure them - [How Growth Tickets Work](/docs/growth-tickets-architecture): what happens to the recommendations - [Report Builder API](/docs/report-builder-api-reference): running and reading reports programmatically --- In this section: - Webhooks: Send Cogny Events to Slack, Zapier or Your Own Server: https://cogny.com/docs/webhooks - How Growth Tickets Work: https://cogny.com/docs/growth-tickets-architecture - **How AI Reports Work** (this page): https://cogny.com/docs/ai-report-generation-how-it-works - How ICP Analysis Works: https://cogny.com/docs/icp-analysis-technical-overview Source page: https://cogny.com/docs/ai-report-generation-how-it-works All documentation: https://cogny.com/docs Cogny for agents: https://cogny.com/llms.txt