Automating Ad Operations with AI Agents on GCP

Ad trafficking — the process of setting up and managing ad campaigns in an ad server — is one of those tasks that's too complex for simple automation but too repetitive for skilled humans. It involves parsing insertion orders from emails, validating campaigns against an OMS, generating trafficking sheets, and updating the ad server. I built an AI agent system on GCP that handles this end-to-end with a human approval step.

The Architecture

The system uses a two-agent architecture: a Planner Agent that processes and validates, and an Executor Agent that applies changes. Between them sits a human approval step via email. This separation ensures no changes hit the ad server without explicit human sign-off.

# Agent workflow (simplified)
# 1. Email arrives with IO attachment
# 2. Planner Agent processes:
planner_tools = [
    validate_email_tool,      # Extract & validate IO attachment
    validate_campaign_tool,   # Check against Boostr OMS
    generate_sheet_tool,      # Create trafficking sheet
]

# 3. Approval email sent to sender
# 4. On "approved" reply, Executor Agent runs:
executor_tools = [
    update_gam_tool,          # Push to Google Ad Manager
]

GCP Service Selection

Every component maps to a managed GCP service, keeping operational overhead minimal.

The Planner Agent

The Planner Agent runs three validation tools sequentially. First, it extracts the IO attachment from the email and validates the file format (XLS/XLSX). Second, it calls the Boostr OMS API to validate the campaign exists and the line items match. Third, it generates a standardized trafficking sheet with all the fields Google Ad Manager needs.

# Email subject determines validation depth
# IO_TAGS      → Quick mode: skip OMS validation
# IO_TAGS_FILE → Full mode: validate against Boostr + GAM

subject_rules = {
    "IO_TAGS": {"oms_validation": False, "gam_validation": False},
    "IO_TAGS_FILE": {"oms_validation": True, "gam_validation": True},
}

Human-in-the-Loop Approval

After the Planner Agent completes, an approval email goes back to the original sender with a summary of what will be created in GAM. The user replies with 'approved' to trigger the Executor Agent. This pattern adds minimal friction while preventing automated mistakes from reaching the ad server.

Secrets Management

The system integrates with three external APIs (Gmail, Boostr, Google Ad Manager), each with different auth mechanisms. All credentials live in Google Secret Manager with IAM-scoped access. The Cloud Function retrieves secrets at runtime — no credentials in code or environment variables.

Lessons Learned

Three things I'd do differently. First, implement idempotency from day one — email processing can trigger retries, and you need to handle duplicate IOs gracefully. Second, add structured logging for every agent tool call — debugging agent decisions requires complete audit trails. Third, start with the simplest possible agent architecture (sequential tools) before reaching for complex orchestration patterns.