A wrong saved reply scales the error
Canned replies gradually become outdated. The product changes, the limit is altered, and the screen is given a new name. The saved reply keeps being sent, right in tone and wrong in content, hundreds of times before anyone notices.
Nobody owns this. Agents use the responses without checking them, and the person who wrote them has moved on, so staleness builds up in the most frequently sent material.
Staleness comes in two types that need different handling. A factually incorrect response must be handled urgently. A poorly worded but accurate item belongs on the backlog. If you treat both as equal priority, neither gets addressed.
The more often something is used, the more urgent it becomes. A mistake made twice a year and one used forty times a week are not the same issue.
How the agent is built in Agent Studio
Usage data for each pre-written response turns this from an audit into a priority list. An HTTP Request node reads the response library and retrieves the send counts.
The system prompt distinguishes a factual error from a wording problem and requires that each finding be assigned one of the two, along with a severity level.
Prompt skills include facts that change often in your product: limits, plan boundaries, and feature names, so the agent knows where to look rather than treating everything with equal suspicion.
MCP connections provide access to the current documentation, and each response is checked against it.
Agent Chat hands the library over when a Start Scheduled trigger fires. Factual errors go to Slack immediately, with the response name and the specific claim that is now incorrect. Wording issues go to a Notion backlog ordered by usage.
Automatic editing does not take place. A saved reply sent out incorrectly is already bad. A reply silently rewritten by an agent that has not been reviewed is worse, because the error then becomes traceable.
What this agent is built from
- System prompt: requires every finding to be named as a factual error or a wording problem.
- Prompt skills: the facts that change often in your product, so the agent knows where to look.
- MCP connections: current documentation to check each response against.
- HTTP Request (Util): reads the response library and includes send counts, so findings are ranked by exposure.
- Agent Chat (Util): a scheduled flow hands the library over for review.
- Slack (Communication): immediately carries factual errors. Notion receives the wording backlog.
- When an ActionFlow is enough: If flagging responses that have not been touched for a year is enough, build the workflow. The agent judges whether a response is actually wrong.
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.