The Same Three Hours, Every Week
Monday morning. Open GA4, pull last week's sessions and conversions. Download the Naver Ads report. Screenshot the Meta dashboard. Paste it all into a spreadsheet. Write comments explaining why the numbers moved. Format it into slides. Two to three hours — and next week you do it again.
The problem isn't that the work is hard. It's that it's repetitive yet needs a human every time. The data comes from fixed sources, the analysis frame is the same each week, and the report format is fixed. That's exactly the shape of a task worth automating.
This post is a recipe for automating that weekly report with a single Marblo board. Assuming you've already tried the first multi-agent tutorial once, we'll build one real workflow end to end.
Why a "Board," Not Just a Chatbot
"Couldn't I just paste the data into ChatGPT and ask for a report?" — Yes. Once. The trouble is it has to run every week, automatically, and reliably.
A weekly report is really a bundle of four different kinds of work:
- Data collection — pulling numbers from several sources accurately (no room for error)
- Analysis — interpreting what the numbers mean (long-context reasoning)
- Writing — turning it into a human-readable report (fluent generation)
- Verification — checking that no numbers were invented (independent judgment)
Ask one chatbot to do all four and each step runs on that model's weak spot. Split the work across models chosen for their strengths and both quality and cost improve — this is why heterogeneous agents beat a single model. A Marblo board lets you compose that division visually, run it on a schedule, and trace every step.
Recipe Architecture
The board wires four stations in sequence:
[Trigger: every Mon 09:00]
│
▼
collector (data collection) ──┐
│ │
▼ │
analyst (analysis) │ (raw data also flows to the checker)
│ │
▼ │
writer (report drafting) │
│ │
▼ │
fact_checker (verification) ◀──┘
│
▼
[Human approval gate] → send to Slack / email
The key is the second connection, collector → fact_checker. It makes the verifier see both the raw data and the drafted report, so if a number appears in the report that isn't in the source, it gets caught. It's the same structure as the research pipeline in the tutorial, applied to marketing data.
Station by Station
1. collector — Data Collection
The most important station. If the numbers are wrong, everything downstream is wrong. So this station needs accuracy, not creativity.
In Marblo, data comes in through MCP connectors or each platform's API. Realistic wiring:
- GA4 — the most stable path is to export GA4 into BigQuery and query it with SQL (see the GA4 + BigQuery guide). The collector agent runs a fixed query to pull last week's metrics.
- Naver / Meta ads — an MCP tool that calls each platform's reporting API to collect spend, impressions, clicks, and conversions.
- Sources with no MCP — attach a tool that lets the agent call the API directly. For what MCP is and why it became the tool standard, see the MCP guide.
Role: collector · A light, accurate model (e.g. Claude Haiku or Gemini Flash) is plenty — the job is running a fixed query, not reasoning. System prompt, in essence:
You are a data collection agent. Use the connected tools to gather last
week's (Mon–Sun) metrics and output structured JSON only.
Collect: sessions/conversions/conversion rate by channel, spend/ROAS by
ad channel, week-over-week change (%).
Never estimate or fill in numbers. Leave any value the tools did not
return as null.
That last line matters — if it's empty, don't fill it, leave it null. This blocks invented numbers from ever reaching the report.
2. analyst — Analysis
This station pulls the story out of the numbers. Not "conversion rate rose 12%" but "rising Naver brand-search traffic lifted the conversion rate." It needs long-context reasoning, so Claude fits.
Role: analyst · Input: the collector's JSON. System prompt, in essence:
You are a marketing analysis agent. Take the structured weekly metrics
and distill 3–5 key insights on 'what moved and why'.
Rules:
- Ground every insight in the numbers provided
- If a cause can't be confirmed from the data, mark it "Hypothesis:"
- Include 1–2 recommended actions
3. writer — Report Drafting
Turns insights into a report a person reads. Use a GPT-class model, strong at fluent generation. Put your own house report style into the system prompt and you get a consistent format every week.
You are a marketing report writing agent. Take the analysis insights and
write a weekly report using the fixed template.
Structure: (1) one-line summary (2) key-metrics table (3) 3–5 insights
(4) next-week actions. Attach week-over-week change to every number.
Never add anything not present in the analysis.
4. fact_checker — Verification
Cross-checks the drafted report against the raw data. Crucially it uses a different vendor (Gemini) — a model checking its own writing shares its own blind spots. Inputs: both collector (raw) and writer (report).
You are a verification agent. Take the raw metrics JSON and the drafted
report, and check that every number in the report matches the source
exactly.
Output:
- Match: same number as source
- Mismatch: differs from source (needs fixing, state the source value)
- Unsupported: a number/claim in the report that isn't in the source
Flag anything where a number, percentage change, or date is off.
Where the Human Comes In — the Approval Gate
This is what lets you hand the automation your trust. Instead of sending the verified report straight out, put a human approval gate just before dispatch.
In Marblo you can hold the pre-send node as pending approval. The report lands in Slack, a teammate skims it for 30 seconds, and it only goes out when they hit Approve. And if fact_checker flags even one "Mismatch" or "Unsupported," it automatically raises a warning to that teammate.
This one layer removes the "can I really send an AI-written report to an executive?" anxiety. Run it with manual approval for the first few weeks to build trust; once verification passes reliably, switch to auto-send. For how to design this kind of trust layer for in-house agents, see the 5 principles of AI agent governance.
Weekly Auto-Run + Real Numbers
Put a schedule trigger (every Mon 09:00) at the top of the board, and before your teammate even clocks in Monday, a fully verified report is sitting in Slack awaiting approval. Every run leaves a trace, and cost is tracked per agent.
To be honest, automation doesn't replace human judgment. The analyst's insights are a draft, and calls like "how should we read this A/B result?" are still a person's job. What this recipe removes is labor, not judgment — the two-to-three hours of pulling, pasting, and formatting.
On our own team, weekly report prep dropped from roughly 2.5 hours per round to about 10 minutes of review and approval. The time saved moves from making the report to deciding on it.
FAQ
Q. Couldn't I just write an API script instead of using Marblo? You could. But then you build the multi-model role split, the run tracing, the approval gate, and the per-agent cost accounting yourself. A Marblo board replaces all of that with node connections.
Q. My data sources go beyond GA4, Naver, and Meta. The station structure stays the same. Add one more tool (MCP/API) to the collector and you add a source. The architecture scales regardless of source count.
Q. What does it cost?
Multi-agent boards are available from Marblo Pro (from ₩19,000/month), plus per-run API charges for each model. Placing cheaper models on collector and fact_checker cuts per-run cost substantially.
Q. Does the verifier actually catch errors? The most common catch is the writer rounding a number slightly or miscomputing a percentage change. Because a different-vendor model checks against the source, it filters these mismatches better than a model reviewing its own writing.
Build It Yourself
If you'd like to build this board against your own metrics, book a free 30-minute session. The Marblo team will design your first board around your data sources and report format. Or just start free on Marblo.
Related Posts
- Building Your First Multi-Agent System with Marblo
- Why Heterogeneous AI Agents Beat a Single Model
- Connecting GA4 and BigQuery to Build a Marketing Dashboard
- 5 Principles of In-House AI Agent Governance
Last updated: 2026-07-09. Marblo's UI improves often, but the board recipe pattern is stable.