Most quality alerts describe your business correctly
Quality rules fire on anything unusual, and a large share of what is unusual is simply true. A spike in orders on a promotion day, nulls in a field that only enterprise customers fill, a currency appearing for the first time because you entered a market.
Each of those is a correct observation and a false alert. After a month, the quality report is filtered and saved to a folder.
Distinguishing them requires knowing the business, not the schema. Nothing in the table says that last Tuesday was a campaign. Nothing in the distribution says that a new country is an expansion rather than corruption.
Severity is the second judgment, and it is about consequence. A broken field in a dormant table and a broken join key feeding billing are both errors. Treating them as the same priority is how the billing error waits.
How the agent is built in Agent Studio
What "normal" looks like for your business is the configuration, and it belongs in prompt skills, where an analyst maintains it: seasonal patterns, known campaign effects, legitimately sparse fields, and deliberately dormant tables.
The system prompt defines severity by consequence rather than by size. It requires each finding to name what breaks downstream.
The agent reads the raw issue set produced by a Validate-based flow. Mechanical detection remains in the flow. Only judgment is delegated. That split keeps costs down and detection auditable.
Agent Chat delivers the issue set on a schedule, along with recent business context from a Google Sheets node.
Slack receives only what the agent judges to be genuine problems, ranked by consequence. A Notion node holds everything the agent dismissed, with the reason.
Reviewing the dismissed list monthly is the discipline that keeps this safe. An agent quietly suppressing a real problem is the failure mode, and only the dismissal log reveals it.
What this agent is built from
- Prompt skills: seasonal patterns, known campaign effects, and legitimately sparse fields.
- System prompt: severity defined by downstream consequence, named on every finding.
- Validate (Data): mechanical detection stays in the flow. Only the judgment is delegated.
- Google Sheets (Integration): supplies recent business context alongside the issue set.
- Agent Chat (Util): a scheduled flow hands over the issue set and context.
- Notion (Integration): holds every dismissal with its reason for monthly review.
- When an ActionFlow is enough: If the work is type, range, and null checks, keep those checks in the workflow. The agent only decides what the results mean.
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.