Automated reporting & dashboard narratives

Every month, analysts and controllers copy numbers from dashboards into slides and write "revenue rose 4 % because..." by hand. AI turns governed metrics into draft narratives, flags what changed and why, and ships the report on a schedule. Humans review exceptions. Typical benefit: reporting cycles shrink from days to hours and analysts spend time on interpretation instead of formatting.

Problem & value

• Today: recurring reports (management pack, sales review, ops KPI deck, board paper) are assembled manually from BI tools and spreadsheets. Commentary is written under time pressure, varies by author and is often late. Most dashboard viewers never open the dashboard; they want a short "what changed and what needs attention".
• AI value: consistent first-draft commentary, automatic variance and driver explanation, proactive push of insights (email, Teams, Slack), personalised views per role, multilingual versions.
• Evidence: vendors document these features (Power BI Copilot narrative visual and report summaries, Tableau Pulse insights, Looker/Gemini) [1][2][3]; published ROI figures are mostly vendor claims, so measure your own baseline (hours per reporting cycle).
• Key constraint: numbers must come from a governed query, never from the LLM's memory.

Variants

• Management/FP&A commentary: variance narratives vs. budget/forecast (see DATA-024).
• Dashboard narratives: auto-summary text inside the BI tool, refreshed with the data.
• Operational digests: daily plant, store or support KPI briefs pushed to chat.
• Public sector: statistics offices and municipalities generating text for open-data dashboards.
• Regulated disclosures: sustainability and supervisory reports follow stricter paths, see DATA-012, DATA-020.
• Private use: household budget or fitness summaries from exported data, usually a generic LLM on a CSV (DATA-006).

How it works

1. scheduled trigger / data refresh
2. governed metric query (semantic layer / SQL, fixed definitions)
3. deterministic analytics (period deltas, variance, trend, anomaly, driver/contribution analysis)
4. fact pack (JSON: numbers, comparisons, flags)
5. LLM writes narrative strictly from the fact pack (template + tone + language)
6. verification (every number in text matched to fact pack; banned claims; length)
7. human review of draft (exceptions, sign-off for external/board reports)
8. distribution (PDF, slide, email, chat, BI tile) + audit log

• Deterministic first: statistics, anomaly detection and key-driver analysis compute what changed; the LLM only phrases it. This avoids arithmetic hallucinations.
• Semantic layer: metric definitions (e.g. dbt Semantic Layer / MetricFlow, Cube, LookML, Power BI semantic model) keep "revenue" identical across reports and give the LLM a safe interface.
• Human-in-the-loop: reviewer approves narratives for board, investor, customer or regulator audiences; internal daily digests may go out automatically with a visible "AI-generated" label.

Data preparation

Automated narratives need trustworthy numbers. The data lives in accounting or ERP exports (DATEV, SAP, Lexware), the BI tool, and above all in Excel files and slide decks on file shares and in mail. Typical gaps are inconsistent KPI definitions between departments, manual overrides in spreadsheets, missing prior-period and budget values, and past commentary scattered over old PowerPoint files.

Preparation steps: (1) agree one definition per KPI and write it down, see PREP-012; (2) move report inputs from copied spreadsheets into tidy tables or a BI model, see PREP-011 and PREP-005; (3) collect three to six past reports as gold-standard examples and templates, see PREP-016; (4) build a list of company terms and abbreviations, see PREP-015; (5) add data-quality checks for the source tables with DATA-008 and deduplicate master data with DATA-003; (6) define a calculated-numbers-only rule: the model writes text, code computes all figures.

Tools: BI semantic model or dbt, Power Query, a small Python or n8n pipeline. Rough share of effort: about 30 to 40 percent, mainly in aligning definitions and removing manual steps. Check readiness with GUIDE-003.

Implementation paths

Which path for whom

• smb: Path A, because existing BI and accounting tools offer mature narrative and report-summary features at SMB prices for standard KPIs.
• enterprise: Path B, because reporting is a core process with own data models, and API costs per report are low compared with analyst time.
• developer: Path B, because a few lines of orchestration over a metric layer and an LLM API cover most needs, with full control of the reconciliation checks.
• public: Path C on sovereign EU/German infrastructure, because budget and statistical data fall under procurement and sovereignty requirements.

Path A — Buy (recommended for most SMBs)

• Inside the BI tool you already have: Microsoft Power BI with Copilot (narrative visual, report summaries), Tableau Pulse (Salesforce), Google Looker with Gemini, Qlik Sense / Qlik Answers, ThoughtSpot Spotter and SpotIQ, Domo, Zoho Analytics (Zia), Sisense.
• Planning and finance reporting: Anaplan, Workday Adaptive, Pigment (EU, France), Jedox (EU, Germany), Lucanet (EU, Germany) and board-reporting tools offering AI commentary.
• Dedicated narrative generation (NLG): Arria NLG, Narrative BI, Yseop (EU, France) and Automated Insights-style engines.
• Choose A when: a BI tool is already in place and reporting is standardised. Check EU data residency, whether prompts/data are used for training, and capacity licences (Copilot features often require paid Fabric/Power BI capacity).

Path B — Build on cloud APIs

• Models: frontier LLMs with structured output via EU regions (Claude on Bedrock/Vertex EU, Azure OpenAI EU Data Zone, Gemini EU, Mistral La Plateforme as EU provider).
• Data access: read-only service account on the warehouse or semantic layer (Snowflake Cortex, Databricks AI/BI Genie, BigQuery, dbt Semantic Layer); the application computes the fact pack, the model only writes text.
• Orchestration: scheduled job (Airflow, dbt Cloud, Power Automate, n8n), template engine (Jinja), PDF/PPTX generation, MS Teams/Slack delivery.
• Choose B when: you need custom report formats, several data sources or branded multi-language packs.

Path C — Self-host (open source)

• Models: open-weight instruct models such as Qwen3, Llama 3.x, Mistral Small/Medium-class, Gemma 3 served with vLLM or Ollama; JSON-schema constrained decoding.
• Stack: Apache Superset or Metabase for dashboards (with scripted API calls), dbt + MetricFlow or Cube for semantic layer, Evidently / Great Expectations for data checks, Apache Airflow or Dagster for scheduling, Quarto/Jinja for rendering reports.
• Hardware: HW2 is enough for a 14-32B quantised model producing a few hundred reports a day; narratives are short.
• Choose C when: finance or personal data may not leave the premises.

Path D — Custom training

Rarely needed. Consider light fine-tuning (LoRA) for house style or domain terminology if hundreds of approved past commentaries exist; retrieval of past examples usually gets most of the benefit.

State of the art (2026)

• Narratives from fact packs with schema-constrained output are standard; "chat with dashboard" and "summarise this report" ship in all major BI suites.
• Agentic digests: agents watch metrics, run root-cause analysis and post alerts with explanation; combined with data-quality signals (DATA-008).
• Governed semantic layers are the main enabler for trustworthy answers; Model Context Protocol (MCP) servers expose them to agents.
• Limits: causal claims ("sales fell because of the campaign") are often unsupported; LLMs over-interpret noise, misstate percentages vs. percentage points, mix up periods and sound equally confident when wrong. Stale or incorrect source data produces fluent but wrong stories. Narrative quality drifts with silent model updates, so keep regression tests (golden reports).
• Mitigations: number-matching validators, separation of "facts" and "hypotheses" in the prompt, show source tile links, evaluate with a sample of past reports.

Shared resources & combinations

• Same stack as DATA-001 (NL-to-SQL, semantic layer), DATA-006 (spreadsheet assistant), DATA-024 (FP&A variance commentary), LANG-003 (report writing).
• Upstream: DATA-008 (data quality gates before publishing), DATA-007 (catalogue and metric definitions).
• Specialised reporting: DATA-012 (CSRD), DATA-020 (supervisory reporting), DATA-025 (pay-gap reports) add regulatory structure and stricter review.
• Shared infrastructure: one LLM gateway, one semantic layer and one scheduler serve all these use cases.

Effort, cost & KPIs

• Path A: days to weeks to enable and test; cost is mostly licence/capacity uplift.
• Path B: 1-3 months for two or three production reports incl. validators and distribution.
• Path C: 2-4 months; GPU server plus operations. LLM token cost per report is small relative to analyst time.
• KPIs: analyst hours per reporting cycle, time from period close to report delivery, share of reports sent without edits, numeric error rate found in review (target zero), reader engagement (opens, actions), stakeholder satisfaction.

People & change

Start with the controller or department lead who owns the report; they define what a good commentary looks like and sign off on it. Management decides which reports may go out with minimal review. IT connects the data sources and scheduling, data owners confirm KPI definitions, and end users (report readers and authors) are involved in tone and format.

Works council: depends. Management reports rarely monitor individuals, but if narratives cover per-person or per-team performance or the tool logs authors' behaviour, §87(1) no. 6 BetrVG applies. Always inform the works council in good time under §90(1) no. 3, since the way report work is done changes. Public bodies involve the Personalrat. Data protection: a DPIA is normally not needed for aggregated figures; the DPO should check personnel or customer-level content, and an AVV is required with any cloud provider.

AI literacy: support staff with a short session on hallucinated explanations ("why did revenue rise?" is causal speculation), see GUIDE-006. Run a pilot with one monthly report for two cycles in parallel with the manual version, measure time saved and corrections, and keep a named human approver for every report. Address the concern "my analysis job disappears" by shifting time to interpretation. Guides: GUIDE-008, GUIDE-009, GUIDE-011.

Risks, regulation & compliance

• EU AI Act: typically minimal risk; generated text shown to people may need transparency (Art. 50) where published to the public as information on matters of public interest unless human-reviewed with editorial responsibility. AI literacy (Art. 4) applies. Becomes sensitive if narratives feed employee evaluation or credit decisions (Annex III), then reassess.
• GDPR: reports with customer, employee or patient-level data are personal data; aggregate with minimum group sizes to avoid re-identification, apply role-based access (row-level security) so the narrative never exposes data the reader cannot see, data processing agreements, EU hosting.
• Financial and disclosure duties: for external, listed-company or supervisory reporting, management remains responsible for content; keep audit trail of AI draft vs. final text and four-eyes approval.
• Typical risks: hallucinated figures, leaked confidential metrics through prompts, prompt injection via free-text fields in source data, over-reliance. Mitigate with validators, least-privilege read-only access, logging and human sign-off.

Getting started

1. Pick one recurring report with a clear owner; record time spent and the current commentary style.
2. Make sure metrics are defined in one place (semantic layer or BI model) and basic data-quality checks run first.
3. Enable the built-in narrative feature of your BI tool (Path A) on a pilot report; compare with human commentary over 2-3 cycles.
4. If more control is needed, build a fact-pack + LLM pipeline with a number-matching validator and a review step.
5. Track edit rate and errors; widen to more reports only when the edit rate drops.

Sources

1. Microsoft Learn - Copilot in Power BI: create narrative summaries and report summaries - https://learn.microsoft.com/power-bi/create-reports/copilot-introduction
2. Tableau Help - Tableau Pulse - https://help.tableau.com/current/online/en-us/pulse_intro.htm
3. Google Cloud - Conversational Analytics / Gemini in Looker - https://cloud.google.com/looker/docs/studio/conversational-analytics
4. dbt Labs - dbt Semantic Layer and MetricFlow - https://docs.getdbt.com/docs/use-dbt-semantic-layer/dbt-sl
5. Databricks - AI/BI Genie and dashboards - https://docs.databricks.com/aws/en/genie/
6. Regulation (EU) 2024/1689 (AI Act), Art. 4, 50 - https://eur-lex.europa.eu/eli/reg/2024/1689/oj
7. Regulation (EU) 2016/679 (GDPR) - https://eur-lex.europa.eu/eli/reg/2016/679/oj
8. Apache Superset - https://github.com/apache/superset