Most week-over-week movement means nothing
Weekly reporting leads to overreaction. A 12 percent drop when volume is low is just noise, but it still prompts a meeting and a change in strategy. The next week the metric returns to normal, and the decision made that week is not revisited.
To decide whether a change goes beyond normal variation, you need volume and history, not only the percentage. A 12 percent figure is meaningless for 200 sessions and serious for 20,000.
'Inconclusive' must be one of the possible answers. A system that always gives a verdict will be wrong about as often as the data is poor. Poor data occurs most weeks for most campaigns.
A budget-change recommendation sets a much higher standard. Observation, interpretation, and recommendation are three distinct stages of confidence. If you collapse them, noise moves money.
How the agent is built in Agent Studio
Variance thresholds are settings for this agent. The system prompt specifies the volume below which nothing is considered significant, and it requires sample size with every claim.
Prompt skills include your campaign taxonomy and known seasonal patterns, so a December drop in a B2B channel is read as a December effect rather than a failure.
A Start Scheduled flow assembles the inputs. An HTTP Request node retrieves current and historical performance data. A Google Sheets node supplies the campaign calendar. An Agent Chat node passes the package along.
The report divides the findings into three distinct parts: significant, inconclusive, and noise. Slack receives the report. The inconclusive section stays visible on purpose, because that section is what prevents action.
Before any budget recommendation is issued, a minimum observation window is always required. That rule lives in the system prompt and stays in effect even during a busy week.
The only way to tell whether the agent's judgment of variance fits your volumes is to compare the agent's calls with what took place every quarter.
What this agent is built from
- System prompt: states volume thresholds and requires sample size with every claim.
- Prompt skills: carry your campaign taxonomy and known seasonality.
- HTTP Request (Util): pulls current and historical performance.
- Google Sheets (Integration): supplies the campaign calendar for cross-reference.
- Agent Chat (Util): a scheduled flow hands over the assembled package.
- Slack (Communication): receives significant, inconclusive, and noisy as three visible sections.
- When an ActionFlow is enough: A performance summary is a workflow, and most teams should build that first. Add the agent when overreaction is the actual problem.
Frequently asked questions
Start building AI workflows
Create a free account, open a template or a blank canvas, and run your first ActionFlow.
Newsletter
Get product updates
New nodes, agents, and product notes. We send mail only when we have something worth opening.