Auto-generated review decks

Marketing ops leads assemble performance review decks on a recurring cadence: weekly standups, monthly business reviews, quarterly planning. The work is familiar and repetitive. Pull metrics from ad platforms, analytics, CRM, and spreadsheet models. Reconcile date ranges. Drop numbers onto a slide template. Write short commentary. Fix formatting. Hand the deck to someone who will present it.

An agent can take the assembly step. It reads from connected data sources, applies an agreed slide structure, and drafts a deck that a human then edits and presents. Tools such as Plus or Rows fit this pattern when they sit on top of live sheets or warehouse extracts and can push charts and tables into slides without a full rebuild each cycle.

This page is for the ops lead who owns the review ritual, not the analyst who invents the metrics. The goal is speed with a clear review gate: the agent drafts; people decide what stays on the screen.

Related reading: Anomaly root-cause explanations, Auto-narrative performance reports, and Conversational reporting in Slack.

Auto-generation pays off when the review format is stable and the inputs are already connected. Typical triggers:

• The same template is reused every period, with only the date window and campaign filters changing.
• Numbers live in more than one system, and copy-paste between them is the main delay.
• Multiple stakeholders need the same core slides before local teams customize for their audience.
• Prep time compresses into the evening before the meeting, so assembly quality depends on who is free that night.

It is a poor fit when the story changes every cycle, when leadership wants a new narrative frame each time, or when source systems are not connected. In those cases, drafting slides by hand is still cheaper than debugging an incomplete agent run.

Treat the agent as a pipeline with explicit inputs and a fixed output shape.

Required inputs. The agent needs (1) a defined review period with start and end dates, and (2) at least one connected data source that covers the metrics in the template. If either is missing, the agent returns empty output and does not invent placeholder slides. That empty result is intentional: a half-built deck with guessed ranges is worse than no deck.

Template contract. Ops defines the slide order once: title and period, KPI scorecard, channel or campaign breakdown, funnel or conversion view, spend and efficiency, notable changes, open questions. The agent fills cells, charts, and callout fields that map to named metrics. It does not invent new slide types mid-run.

Data pull. Using connectors (spreadsheets refreshed by Plus or Rows, warehouse views, platform exports already synced to a sheet), the agent filters to the review period, aggregates to the grain the template expects (campaign, channel, region), and writes values into the slide bindings. Currency, percentage, and decimal formats follow the template rules so presenters do not reformat every cell.

Draft narrative slots. Short captions under charts can be drafted from period-over-period deltas already in the data model (for example, “Paid search CPA up vs prior period”). Captions stay factual and tied to computed fields. Opinion, blame, and strategy recommendations stay out of the draft; humans add those while editing.

Handoff. The output is a draft deck in the team’s slide tool or a linked workbook that generates slides. Ownership then moves to a person: check the period label, spot-check two or three known campaigns, tighten wording, remove slides that do not apply this cycle, and present.

Human-in-the-loop is not optional packaging. It is the control that keeps leadership trust.

Reviewers should confirm:

• The period on the title slide matches the meeting agenda.
• Totals reconcile with a trusted dashboard or export for the same window.
• Campaign or channel filters match what the room expects (brand vs performance, paid vs organic, region).
• Draft captions do not overstate causality. A delta is not a root cause.
• Confidential or incomplete rows are hidden before sharing outside the ops team.

Edit patterns that work well: keep the KPI and breakdown slides; rewrite the “what we think” and “asks” slides by hand; archive slides that show zero activity for the period rather than leaving blank charts. Presenters own the live narrative. The agent’s job ends when the draft is ready for that edit pass.

If connected sources or the review period are missing, do not bypass the empty-output rule by pasting last period’s deck and changing the title. That creates silent error. Fix the connection or set the period, then re-run.

Assign clear roles so the automation does not become orphaned.

• Ops lead: owns the template, metric dictionary, and connector health. Decides when the template changes.
• Agent run: scheduled after data refresh completes for the period (for example, morning of review day, or the night before once daily syncs finish).
• Editor: usually the ops lead or a channel owner, who spends a short window editing and approving.
• Presenter: may be a different person; they receive an edited deck, not a raw agent export.

Version the template when leadership changes the scorecard. Do not ask the agent to “figure out” a new structure from chat. Structure changes are product decisions for the review ritual; the agent only fills the current contract.

Track failure modes that matter in practice: stale credentials, sheets that stopped refreshing, period parameters left blank, and metrics renamed upstream. Surface those as empty or partial runs with a clear reason, not as slides filled with zeros that look like real performance.

Auto-generated review decks improve speed when assembly is the bottleneck. They do not replace judgment about which campaigns to highlight, how to explain a miss, or what to ask the room for. Pair this use case with deeper analysis elsewhere: anomaly explanations when a metric breaks the pattern, narrative reports when the audience needs prose instead of slides, and Slack-based Q&A when stakeholders want answers between formal reviews.

Start small. Connect one trusted scorecard source, lock a five-to-eight-slide template, require period and source before any run, and measure how long the edit pass takes compared with building from scratch. Expand connectors only after presenters trust the draft on a steady cadence.