From Data to Narrative: What AI Actually Does in Reporting

Marketing reporting is the single most time-consuming administrative task in most marketing teams. Surveys by HubSpot, Superside, and Gartner consistently find marketers spend 15-25% of their week on reporting and meta-work: pulling data, formatting slides, writing narrative, and rewriting the same insights for different audiences. The AI opportunity here is specific and measurable: compress the narrative-writing and formatting portions of reporting without sacrificing analytical rigor. AI handles four reporting tasks genuinely well: (1) transforming structured data into English prose, (2) identifying notable changes and outliers when pointed at them, (3) formatting the same information for different audiences (executive vs operational vs board), and (4) generating consistent week-over-week/month-over-month narrative with the same structural beats. What AI does NOT handle: (a) pulling the data itself (requires BI integration), (b) deciding which metrics matter this period (strategic prioritization), (c) verifying that the numbers are correct (hallucination risk), (d) supplying the strategic context that distinguishes 'traffic up 12%' from 'traffic up 12% because the pricing page ranked for a high-intent keyword after last week's update.' The pragmatic split: AI handles 70-80% of the report structure and narrative drafting; humans add context, interpret causation, and validate. A weekly report that took 3 hours manually drops to 35-45 minutes with AI; monthly reports drop from 1-2 days to 60-90 minutes; QBRs drop from 3-5 days to 2-3 hours. The fundamental prompt pattern: 'Here is structured data for the reporting period [paste numbers or CSV]. Here is the prior-period comparison [paste]. Here are the 3 strategic themes we want to emphasize [paste]. Write a [report type] for [audience] that covers [specific sections], follows [structural template], and uses [voice sample]. Do not invent numbers. Flag any anomalies. Output as structured markdown.' Pitfall: feeding raw dashboards without context. AI given '1,234 leads vs 1,100 prior period' produces a flat statement. AI given '1,234 leads (+12%) vs 1,100 prior period; team shipped a pricing page update on day 3; we ran a LinkedIn campaign starting day 7' produces a coherent narrative. Context is the input that converts noise into narrative. Tradeoff: more AI automation accelerates throughput but increases the risk of AI confidently misrepresenting numbers or missing context a human would catch. Build verification checks into every workflow.

Prompt Templates by Report Type: Weekly, Monthly, Campaign, Channel, QBR

Different report types require different prompt structures because the audience, cadence, and purpose differ. Five types cover most marketing reporting needs. Weekly performance report (15-30 min target). Audience: marketing team and direct manager. Focus: operational metrics (leads, MQLs, traffic, campaign activations, content published) and 'what changed since last week.' Prompt: 'Generate a weekly marketing report for the week ending [date]. Include sections: (1) Topline numbers vs last week vs 4-week average, (2) Top performers (3 items, each with specific number), (3) Items needing attention (3 items with specific issue), (4) Team activity highlights (non-numeric wins), (5) Next week's priorities. Keep to 400-500 words. Source data: [paste]. Do not invent metrics; flag missing data.' Monthly review (60-90 min target). Audience: CMO, functional peers, senior leadership. Focus: trend analysis, goal attainment, strategic takeaways. Prompt additions: explicit goal/actual comparison, trend assessment ('month 3 of a declining CAC trend'), and one risk/opportunity per channel. Length: 800-1,200 words. Campaign wrap report (30-45 min target). Audience: campaign stakeholders (marketing, sales, product). Focus: what happened vs the plan, what drove the results, what we learned. Prompt: 'Generate a campaign wrap report for [campaign]. Include: objective recap, results vs target (with % attainment), top 3 drivers, top 3 friction points, 2 recommendations for next campaign, and an attribution summary. Source data: [paste campaign metrics, attribution breakdown, and timeline].' Channel report (30-60 min, per channel). Audience: channel owner + manager. Focus: channel-specific metrics, cohort analysis, and channel-relative benchmarking. Common channels: paid search (CAC, CPC, conversion rate by keyword cluster), paid social (CAC by audience + creative, frequency, creative fatigue signals), email (open/CTR by segment, list health, sender reputation), organic (rankings, traffic, backlinks), content (publishing velocity, engagement, assisted conversion). QBR / quarterly business review (2-3 hours target). Audience: CEO, CFO, board. Focus: strategic narrative, outcomes vs goals, headwinds/tailwinds, next quarter's thesis. Length: presentation with 15-30 slides + 1-page exec summary. Prompt: 'Generate the narrative and bullet points for a Q[X] business review. Sections: TL;DR (5 bullets), results vs goals (funnel-level), key drivers (top 3), key headwinds (top 3), experimental results, forward thesis for next quarter. Use structured JSON output with slide-level bullets. Do not invent financial numbers; flag any inputs missing.' Tradeoff: heavy templating accelerates production but can produce 'all our reports look the same' syndrome. Refresh templates every 6 months; add context about unique quarters (product launches, market shifts) that breaks the pattern deliberately. Pitfall: assuming the CFO report equals the CMO report with different numbers. Audience determines framing, CFO wants unit economics and efficiency; CMO wants pipeline and brand metrics; CEO wants strategic narrative with rare numeric anchors.

The Art of the Executive Summary

The executive summary is the most read, most re-used, and most scrutinized paragraph in any marketing report. Executives make decisions based on this paragraph alone. The rule every sentence must pass: the 'so what' test. 'Traffic grew 12% week over week' fails because it doesn't connect to a business outcome. 'Traffic grew 12% week over week, which contributed to 34 additional MQLs and $12K in pipeline' passes. AI defaults to fail the 'so what' test because training data is full of report summaries that list metrics without consequence. Your prompt must force the translation. Prompt: 'Write an executive summary for this report that connects every data point to a business outcome. For each metric cited, specify its impact on revenue, pipeline, or strategic goals. Do not include metrics that don't tie to an outcome. Maximum 5 bullets or 150 words.' Audience tuning matters. CMO summary: emphasizes pipeline, campaign effectiveness, brand metrics, team performance. 'Paid search CAC improved 18% as we shifted budget from display to long-tail keywords.' CFO summary: emphasizes cost efficiency, ROI, cashflow impact. 'Paid budget came in 8% under plan; CAC of $142 is 22% below the $180 FY target.' CEO summary: emphasizes strategic narrative, not numeric detail. 'We're ahead of plan on acquisition efficiency, behind plan on enterprise pipeline, trade-off reflects our Q3 bet to optimize SMB demand generation.' Prompt the AI with the audience explicitly: 'Write three versions of this summary for (a) CMO, (b) CFO, (c) CEO. Each should reflect the audience's decision context.' Common mistakes the AI will make unless prompted otherwise: burying the lede (putting the most important number in the third bullet instead of the first), over-detailing in areas the executive doesn't decide on (writing three bullets about email deliverability for a CEO), ignoring the tradeoff (saying 'CAC improved' without noting the volume drop that enabled it). Prompt pattern to counter: 'Lead with the single most important result for this audience. Each subsequent bullet should deepen the narrative, not restart it. If a result involves a tradeoff, describe both sides in the same bullet.' Pitfall: writing the executive summary last. Do it first (based on the topline data) then the detail sections fill the body; this forces clarity on what actually matters. Tradeoff: AI produces fluent summaries but needs explicit prompting on emphasis. A flat 'summarize this data' prompt produces an average summary; a 'what would the CFO act on here?' prompt produces a targeted one. The 15-20 minute human review of the summary is the highest-leverage edit in the entire report.

Dashboard Commentary and Anomaly Annotation

Dashboards without commentary are numbers without meaning. Most BI tools (Looker, Tableau, Mode, Hex, Omni, Metabase, Power BI) support text annotations; few teams use them consistently because writing commentary is tedious. AI removes the friction. The dashboard commentary pattern: for each dashboard section, feed AI the relevant data slice plus a short context note, and request 2-3 sentences of plain-English commentary. Prompt: 'This dashboard section shows [metric] over [time window]. Key data points: [paste]. Context: [one-line note about relevant external events]. Write 2-3 sentences of commentary explaining the trend, notable points, and what to watch next. No jargon. No speculation beyond the data.' Apply this to: funnel dashboard (conversion rate section commentary), channel dashboard (per-channel performance), cohort dashboard (retention shifts), attribution dashboard (multi-touch path changes). Update commentary weekly or with each refresh. Anomaly annotation is the higher-leverage pattern. When a metric spikes or dips, annotate it in the moment with context: why it happened, what action was taken, what we expect next. Six months later, that annotation is invaluable when someone asks 'why did traffic spike on March 14?' AI prompt: 'On [date], [metric] showed [anomaly value vs baseline]. Context: [what happened]. Write a 1-2 sentence annotation suitable for future analysts. Include the cause, the response, and expected duration of the effect.' Integration patterns. Most modern BI tools support annotations via API (Looker's Markdown tiles, Tableau Annotations, Mode Report Builder, Hex Notebooks). A lightweight workflow: weekly scheduled script pulls the dashboard data, passes to AI with context template, posts the generated commentary back to the dashboard. Build this in Zapier + Make (no-code), Airflow + Python (code), or Databricks workflows (data-team-led). Pitfall: AI speculates when data is thin. A prompt that says 'explain why X happened' when the data doesn't support causation produces confident hallucinations. Safer prompt: 'Describe the observed change. If multiple plausible causes exist, list them without claiming certainty.' Pitfall: one-time commentary with no refresh. A dashboard labeled 'insights last updated 6 months ago' signals a stale data asset. Set a cadence. Pitfall: jargon-laden commentary. 'Funnel mid-stage velocity improved sequentially on a rolling 28-day basis' communicates nothing. 'More prospects moved from demo to proposal this month than last; we shipped a shorter demo script on day 12' communicates something. Tradeoff: deep commentary on every chart is expensive in attention; prioritize the 3-5 dashboards executives actually read.

Automated Workflows and Report Governance

A semi-automated reporting workflow captures most of the efficiency gain without the risk of fully automated reports shipping without human eyes. The canonical 35-45 minute weekly pattern. Step one, data collection (10 min): pull metrics from dashboards into a structured template (CSV or formatted block). Tools to automate the pull: Mode Python notebooks that output to Slack/Google Sheets, Hex apps with scheduled refresh, Sigma's scheduled exports, or a custom job using Fivetran/dbt → warehouse → API. The template should include current period, prior period, 4-week average, and 52-week trend. Step two, AI narrative generation (5 min): feed the template plus context notes into the prompt. Context notes are the differentiator: 'we launched the new pricing page on Tuesday,' 'the paid campaign paused on Friday due to budget exhaustion,' 'engineering shipped a site-speed optimization mid-week.' Without these, AI narrates noise. Step three, human review and context (15 min): the editor's job is to add strategic interpretation, verify numbers, kill jargon, flag anything that looks wrong, and ensure the exec summary passes the 'so what' test. Most errors originate here if skipped. Step four, format and distribute (5 min): convert markdown to Slides/PDF/email, add charts (AI-generated with Mermaid or a charting tool like Datawrapper, Flourish, or embedded Tableau), publish. Governance elements that separate a good program from a risky one. Verification: spot-check 3-4 numbers per report against source. Numbers that disagree with the source must be fixed before publishing. Attribution: tag every metric in the narrative with the dashboard or source it came from so reviewers can trace any number. Versioning: keep a report archive with dates and authors (including 'AI-assisted' if AI drafted). Style guide: a shared voice-sample library for the team so AI outputs remain brand-consistent across reporters. AI-disclosure policy: decide whether to internally/externally label reports as AI-assisted; policies vary by organization and regulator. Attribution pitfall: AI blending numbers. When fed two data sources (e.g., Google Analytics + a CRM export), AI can confuse which metric came from which source, producing confidently mis-attributed statements. Build source-column headers in your data template and require AI to preserve them. Pitfall: 'ship unreviewed reports to executives' is the single highest-risk operational change. Never. Tradeoff: fully automated reports ship fastest but carry the highest downside if a number or narrative is wrong and reaches a board deck. Semi-automated with the 15-minute human review step captures the 80% efficiency gain at a 5% risk cost.