CFO narrative from raw financial statements

The AI agent converts raw data from financial statements and the data warehouse into a ready-made management commentary draft — the narrative accompanying the close that the CFO sends to shareholders, the board of directors, or the owner. Instead of the CFO or financial analyst manually converting figures into a narrative, the system does this the moment the close data is ready. The draft contains figures, variance calculations, and initial explanations — enough for editing, not for raw publication. The CFO remains in the role of editor and interpreter but is freed from the routine of assembly.

What automation does

1. Retrieves current data from the data warehouse or BI layer at the period close mark.
2. Calculates variances across several axes: month over month, quarter over quarter, plan vs. actual, year over year.
3. Filters statistically significant changes by configurable materiality thresholds for each dimension.
4. Groups metrics into blocks — revenue, gross margin, operating costs, cash, unit economics, key operating metrics.
5. Generates text explanations for each significant line item with a reference to the data source in the DWH.
6. Assembles the final draft in the company's accepted format — email, Notion, Google Docs, Slack post, or internal portal section.
7. Returns the CFO a draft with attached links to the source figures for quick verification of each statement.
8. After CFO approval, automation sends the final commentary to the investor mailing list or investor update.

What automation does not do

• Does not interpret strategy or make decisions on capital allocation, cost cuts, or hiring — that is the work of the CFO and CEO; the AI agent only highlights variances.
• Does not replace the full management discussion in the board package. Prepares a separate, compact commentary for regular updates between board meetings.
• Does not communicate with auditors or justify the calculation methodology. The AI agent works with already closed and reconciled financial data, and does not conduct an audit.

Automation closes the bottleneck of 'data exists → no text.' The close is technically complete, the data engineer has already loaded everything into the DWH, but the commentary sits with the CFO and delays sending the update to shareholders. For SaaS companies with a regular investor updates cadence, this is critical — the regularity of investor communication matters more than the perfect wording of a single paragraph.

The automation is built as a custom-code pipeline of four layers: data, calculation, narrative generation, delivery. The layers are isolated — each can be replaced to fit a specific company's stack without rewriting the others.

Technical flow

The pipeline runs on a close-calendar schedule or is triggered by a data-readiness event. In the first step, the Data connector pulls current metrics from the data warehouse (Snowflake, BigQuery, ClickHouse) or BI layer (Looker, Metabase, Tableau). The Variance engine calculates deviations against configured rules and filters material changes — data below the threshold does not reach the commentary, to avoid inflating the text with immaterial detail. The Narrative generator on an AI model receives structured context — figures, deviations, period context — and drafts commentary according to the approved template. The Delivery layer packages the result into the required format and sends it via the designated channel.

Implementation steps

1. Intake session with the CFO and team: review the current close process, deadlines, commentary stakeholders, output format, and the approved tone of voice.
2. Data mapping: establish canonical sources for metrics and the chart of accounts. A single table in the DWH or a reconciled BI dashboard serves as the source of truth.
3. Materiality rules: define deviation thresholds for each dimension. Revenue materiality differs from operating costs materiality — thresholds are configured for the specific business and company size.
4. Prompt-engineering: build a commentary template to match the company's style. The template includes tone, structure, mandatory sections, number formatting, and rules for referencing stakeholders.
5. Integration: configure the connection to the DWH and delivery channels — corporate email, Slack, Notion, investor portal, or internal dashboard.
6. Dogfood on historical data: run the system against 2-3 past closes, compare with actual CFO commentary, calibrate the prompt and materiality thresholds.
7. Production: launch on the next close cycle with manual review of the first 1-2 runs, then move to routine operation with periodic QA.

Solution components

The pipeline is built so that the CFO can always trace from the text to the figure in the DWH in two clicks. This is a mandatory requirement: without traceability, automated commentary cannot be published externally. The Audit log records every run, prompt versions, and data sources — required for both internal controls and any external audit of reporting processes.

The Narrative generator does not operate on unclosed periods. If the close is not technically complete, the pipeline does not run or marks the commentary as preliminary. This protects against premature investor communications based on incomplete data, and against wording that would later need to be retracted upon restatement.

Implementing CFO narrative requires a structured financial DWH or BI layer, an aligned chart of accounts, and 5-6 historical close cycles for prompt calibration. Without these three conditions, automation will not assemble correctly.

Data and access

• Data warehouse or BI layer with financial metrics. Snowflake, BigQuery, ClickHouse, Looker, Metabase, or Tableau will work — any source with a stable data structure.
• An aligned chart of accounts and canonical sources of figures. If one metric is calculated differently in two places, we clean the mapping first, then automate.
• A minimum of 6 historical close cycles with real CFO commentary for prompt calibration.
• Read access to DWH/BI for the agent (service account) with a limited scope.
• Delivery channel: corporate email, Slack, Notion, or an internal portal.

Team readiness

• The CFO or responsible finance professional provides feedback on drafts during the calibration stage and validates the first production runs.
• A data engineer or BI analyst assists with metric mapping and configuring the connection to DWH.
• An approved template commentary is not required, but reduces the prompt-engineering workload. Without a template, Grow2.ai helps assemble one based on past commentary.

Timeline

For the basic version: 1-3 weeks with a ready DWH and an approved commentary template. This timeframe includes data mapping, variance engine setup, prompt assembly, and dogfooding on historical data.

For the extended version with multiple output formats (investor email, Slack update for the team, short summary for the board): 3-5 weeks. The main delay is aligning the tone and structure for different audiences.

If DWH is not set up or data is spread across multiple systems, we address this task separately first — it is a separate data infrastructure project with its own timeline.