Automatic narrative for dashboards

The AI agent reads dashboards and the data warehouse, identifies significant metric changes, and writes a text comment in business language. Instead of "what happened to this chart," management gets an explanation right away. Analysts free up time previously spent on repetitive report commentary and shift to deep root-cause analysis.

Step-by-step process:

1. The agent connects to a BI tool (Looker, Tableau, Power BI) or directly to the data warehouse (Snowflake, BigQuery, ClickHouse, Redshift)
2. Extracts fresh metric values for the reporting period and compares them with previous ones — week over week, month over month, or against plan
3. Identifies significant shifts: growth, decline, anomalies, deviations from forecast, breaches of pre-defined thresholds
4. Classifies changes according to business logic — routine seasonality, a real trend, a one-off anomaly, or a structural shift
5. Generates a narrative draft in Russian or English: what happened, how significant it is, which metrics and segments are involved
6. Publishes the result in the chosen format — a Slack post, email newsletter, a page in Confluence or Notion, a PowerPoint slide alongside the chart

Executive reporting turns from a week-long project into a one-day one. Analysts receive a ready draft to review and adjust, rather than writing from scratch for every Monday.

What automation does NOT do

• Does not replace deep root-cause analysis — the AI agent records the fact of a change, but investigating "why exactly" remains with the analyst and requires domain context
• Does not make management decisions — the narrative provides a description and hypotheses, while conclusions and actions are made by management
• Does not work without quality data — if metrics in the warehouse are calculated with errors or the schema is undocumented, the narrative will repeat and amplify those errors

Technical architecture of a custom-code solution: a Python or Node.js service with access to the data warehouse and LLM API. The AI agent runs on a schedule or trigger and converts numerical shifts into text in the team's business language.

Step-by-step implementation:

1. Data source connection — SQL connector to the warehouse (Snowflake, BigQuery, Redshift, ClickHouse) or to the BI tool API (Looker, Tableau, Power BI)
2. Semantics layer setup — a list of key metrics with clear business names, units of measurement, direction "growth good / bad" and acceptable thresholds
3. Change detector — an algorithm that compares current values against history, accounts for seasonality and statistical significance, and filters out noise
4. Narrative generator — an LLM (AI model or comparable model) receives a table of changes as input and writes text according to a prompt template, taking corporate tone into account
5. Format templates — a separate prompt for a Slack message, a separate one for an email newsletter, a separate one for a PowerPoint slide or Confluence page
6. Delivery — integration with the CMS/content system and BI tool: publishing the narrative alongside the dashboard or to the corporate messenger on a schedule

System components

Security and compliance

The agent requires read-only access to the warehouse and BI tools — write access and schema modifications are not needed. Secrets and tokens are stored in a manager such as AWS Secrets Manager or Doppler, not in code. If metrics contain PII, filtering and aggregation happen before the data is passed to the LLM. An audit log records which numbers were sent to the model and what narrative was produced — this helps investigate errors, verify factual accuracy, and respond to compliance requests.

Alternative approaches

• Ready-made BI add-ons with AI comments (Tableau Pulse, ThoughtSpot Sage) — easier to launch, but less flexibility for the company's internal language and reporting specifics
• No-code solutions on a workflow engine + LLM API — faster to build a first version, but limited when working with a large number of metrics and complex detector logic
• Pure custom-code in Python — maximum customization for reporting and integration specifics, but requires more time to build

Potential pitfalls

• Poorly labeled metric semantics cause the narrative to confuse causes and effects
• Lack of history — if the company has little data for past periods, the detector cannot distinguish a normal shift from an anomaly
• Without reviewing drafts in the first few weeks, prompts do not have time to calibrate to the actual tone of management

What you need to get started:

• A structured data source — a warehouse with regularly updated metrics (Snowflake, BigQuery, ClickHouse, Redshift) or a BI tool with an API
• A list of key metrics with business definitions — without this, the narrative will not be able to distinguish "revenue dropped" from "revenue in one segment returned to normal"
• Access to an LLM API — Anthropic, OpenAI, or a locally deployed model with sufficient context length
• A publishing channel — Slack, Confluence, an email template, or a Notion page where the agent places the finished narrative

Team readiness

• One engineer with Python or Node.js and SQL experience — to build connectors, a change detector, and integrations with delivery channels
• One analyst or data engineer — to define metric significance, write prompt templates, and calibrate the narrative tone
• A stakeholder from management — to validate the first narratives and adjust the style and level of detail

Timelines

For week-level complexity, the build takes 2-4 weeks. The first week goes to connecting sources and labeling key metrics. The second — to building the narrative pipeline and change detector. The third and fourth — to iterations with pilot reports, prompt calibration, and delivery setup in Slack, Confluence, or Notion.