Marketing & growth leaders
CMO Agent Suite
In productionA marketing team of agents that finds, scores and engages pipeline, and never publishes without sign-off.
The challenge
Marketing teams want AI to research prospects, prioritize pipeline and produce content. They can't let a model invent lead scores or post to the company's LinkedIn on its own.
Our approach
A supervisor agent breaks a brief into steps and routes each to an isolated specialist. Lead scores come from a deterministic BigQuery model, never from the LLM. Anything outward-facing is stopped by a platform-level approval gate until a human approves it from a phone or the CLI. We run this suite on AI2 Consulting's own pipeline every day.
Architecture
A supervisor and its specialists.
Orchestrator
Breaks inbound briefs into steps and routes each to the right specialist.
Lead Generation
Finds new candidate companies with live, citation-backed web search. Never scores them.
Lead Scoring
Grades existing leads with a single parameterized SQL tool. It has no code executor, so it can't compute a score itself.
Client Acquisition
Turns a scored lead into a contact plan and outreach draft. Humans send.
Content
Drafts long-form posts on a stronger model tier. Publishing is gated.
Social
Adapts published material for LinkedIn. Posting requires approval.
Vertex AI Agent EngineGoogle ADKGeminiBigQueryCloud Run + IAPTerraformOpenTelemetryGitHub Actions
Guardrails, enforced structurally
- Scores are reproducible: every result carries the SHA-256 of the SQL that produced it
- Approval gate is a platform plugin, so new agents inherit it automatically
- Approvals are bound to a hash of the exact arguments and expire after one hour
- Orchestrator identity holds zero data permissions; tools impersonate scoped accounts
- Token cap plus Cloud Monitoring budget alert
- CI enforces the security invariants on every push
What it changes
- Lead grades you can explain to a sales leader, and reproduce later
- Content and social output move at agent speed, with human sign-off before anything goes public
- An architecture that passes a security review because the controls are structural
agent trace (illustrative)
$ chat "score our top enterprise leads and draft a post"-> routing to lead_scoring-> lead_scoring calls score_leads(segment='enterprise')<- 12 leads, grades={A:3, B:5, C:4}, sql=8a2724aeab3e-> routing to content-> content calls publish_blog_post(...)!! BLOCKED pending human approval [req 7f3c…]
Next suite
CFO Agent
Want the CMO Agent Suite for your team?
We'll adapt the specialists, data connections and approval rules to your systems in a scoped pilot.