The AI stack running in production today.
Eighteen months of production AI for our active client book. Not a "we're exploring AI" deck. Not a "this is on our roadmap" line. A working stack that ships client work every week. Here's what runs, what doesn't, and the human-review pipeline that keeps it accountable.
What runs, six components
Gemini 2.0, strategy synthesis, brief expansion, long-form content drafting from senior-strategist outlines, brand-voice continuity across multi-month campaigns, NL2SQL for BigQuery-backed analytics queries.
Imagen 3, brand-trained ad creative variant production. Twenty-plus variants per campaign at the same quality bar as a hand-finished hero. The variant volume Meta's algorithm needs.
Veo 3, short-form video at retainer cadence. 15-30 second product, lifestyle and capability video. Brand-consistent, motion-engineered, in-house-finished.
Vertex AI, analytics pipelines feeding client-specific dashboards. Campaign performance, attribution modelling, audience clustering, anomaly detection on ad-account spend velocity.
BigQuery, multi-brand portfolio dashboards. Six brands, five Google Ads accounts, four Meta accounts, one consolidated dashboard. Refreshed daily.
NL2SQL, natural-language query interface to BigQuery. Plain-English questions answered against your data in seconds. No dashboard navigation, no data-team ticket.
What we won't do
From our AI Governance Policy v1.0, the explicit non-list:
- We will not use client-confidential or patient personal information to train any general-purpose AI model.
- We will not ship AI-generated regulated content (healthcare, financial, legal claims) without the relevant compliance review.
- We will not publish AI-generated likenesses of identifiable real people without their consent.
- We will not use AI to fabricate testimonials, reviews, statistics or case-study claims.
- We will not deploy autonomous AI agents to take final external-facing actions on behalf of a client without a human approval step.
The human-review pipeline
Every AI-generated output that affects, refers to, or could be seen by a person passes a human review gate before it ships. The senior strategist edits, substantively, not cosmetically. AHPRA review for healthcare creative. Procurement-claim review for industrial capability statements. Australian Consumer Law review for consumer-brand claims. Founder-pass for consumer brands where the founder voice is a brand asset.
For our regulated clients, every output carries provenance metadata: which model generated it, which prompt produced it, who reviewed and approved it, when it was published.
What this changes for clients
Three things, observably:
- Variant volume. Meta campaigns get 12-24 creative variants instead of 2-4. The algorithm rewards variation, the AI stack provides it at retainer-economic prices.
- Reporting at portfolio scale. Multi-brand portfolios get consolidated BigQuery + Looker dashboards, not 6× separate platform-native dashboard screenshots in a deck.
- Capped retainers stay capped. Production efficiency is what makes capped retainers profitable at retainer scale. Without the AI stack, the Charleston's capped retainer wouldn't be economic.
Why we're not noisy about it
Most agencies shipped "AI capabilities" announcements 12-18 months ago and called it done. We've spent those 18 months actually deploying the stack into client work, quietly, with senior strategists driving the production, with governance documented, with the boundaries explicit.
Today the stack ships work for our active client book. That's our headline. Not "we use AI", but "AI is how a 25-person agency does the work of a much larger team, with human accountability on every output, while keeping our retainer economics reasonable."
Read more:
Want to see the stack in action?
Forty-five-minute call. We'll walk through a live BigQuery dashboard, show you brand-trained Imagen output, and demo NL2SQL against an anonymised dataset.
Book a discovery call