Building Data Quality Checks Into an AI Agent

An agent that writes to your CRM, sends emails, or triggers workflows is only as safe as the data it acts on. Building quality checks into the agent's loop turns "acted on garbage" into "flagged a problem and stopped."

Check before you act

Before an agent performs a consequential action, have it validate the records with Guard and, for larger batches, profile them with the data-quality endpoint. Both return structured results the agent can branch on.

Agent plan:
  1. guard(records)              -> any failures? route to human
  2. data-quality(records)       -> completeness/consistency below bar? stop
  3. repair(records)             -> fix accepted records
  4. act (write / send / route)

Validate at the boundary

curl -X POST https://zapinner.com/api/v1/guard \
  -H "Authorization: Bearer $ZAPINNER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{ "records": [ ... ], "expect": { "email": "email", "phone": "phone" } }'
# branch on passed / failures[]

Profile a batch

For bulk work, the data-quality endpoint reports completeness, consistency, and anomalies with the specific records behind each finding — enough signal for the agent to decide whether to proceed, repair, or escalate.

Fail loudly

The point of the checks is a clear stop condition. When validation fails or quality is below your bar, the agent should surface the evidence and hand off — not push through and hope.