Most signals are noise, and acting on noise costs the relationship
Follow-up sales outreach happens when usage increases. In half of those cases, the spike came from a single data import or from an administrator clicking a feature by accident. The customer still received a sales email even though they were not actually seeing growth.
The cost of sending that email is greater than the cost of missing the expansion. It tells an existing customer that the attention on their account is not only automated but also wrong, and that impression is hard to correct.
The evaluation takes durability into account. A spike that holds for weeks is not the same as one that reverts. Distinguishing the two means looking at the customer's own history rather than a fixed threshold.
Timing is the next issue. Even a true expansion signal on an account with an open escalation is still the wrong time, and usage metrics do not show that.
How the agent is built in Agent Studio
The system prompt defines signal durability: how long a change must last to count. It also instructs the agent to compare against the account's own baseline rather than the fleet average.
Suppression rules should also apply: an ongoing escalation, a recent outage that affected that customer, and renewal negotiations that are underway. No usage metric captures those.
Prompt skills carry your expansion paths, so a qualified signal arrives with a specific recommendation rather than only "consider an upsell conversation."
A Start Scheduled flow runs the review. An HTTP Request node retrieves usage and seat counts. Salesforce or HubSpot supplies support history and account status. The Agent Chat node then requests an evaluation.
Only qualified signals reach Slack, because the account manager can see both the evidence and the suppression check, and so knows what was considered.
Google Sheets logs each signal by type and whether it leads to a conversion. Retire any signal types that never convert. That decision must come from the log, not from impression.
What this agent is built from
- System prompt: how long a signal must hold and the instruction to compare against the account's own baseline.
- System prompt: suppression rules for escalations, outages, and live renewal negotiations.
- Prompt skills: your expansion paths, so that a signal arrives with a specific recommendation.
- HTTP Request (Util): pulls usage and seat counts.
- Salesforce (Integration): supplies support history and account state. HubSpot fits the same slot.
- Google Sheets (Integration): logs signals against conversion so unproductive types can be retired.
- When an ActionFlow is enough: If a seat-limit threshold alert is enough, that is a workflow that catches the obvious cases. The agent is for the signals that a threshold misses.
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.