Working AI agents for marketing and growth decisions
Where we build GenAI that a CMO can use and a CFO can trust. Our flagship One AI Engine for Marketing turns segments, offers and content into one chain of agents, and our demos show agents that call analytical models as tools. The demos run on synthetic data; the same design runs on yours.
One AI Engine for Marketing
Our flagship AI solution: a chain of agents that segments customers, picks the best offer for each one, and writes the campaign, with marketers approving every step.
Marketing runs in silos: segments with BI, offers with the promo team, content with agencies. Every handoff loses time and money.
The result is the same message and discount for everyone, and weeks to launch a campaign. Typically 15–20% of promo spend is wasted.
Dashboards alone don't fix this, because someone still has to act. The gap is not insight. The gap is action.
The agent chain: know, decide, speak
Segmentation Agent
Groups customers by behaviour and value, refreshed every week.
ML + GenAI-written segment profilesPersonalization Agent
Picks the right offer and discount depth for each customer, within budget limits.
Uplift models + offer optimisationCampaign Builder Agent
Writes the email, SMS and push message for each segment, on brand, in minutes.
GenAI-led, grounded in brand knowledgeWhy agentic AI
Agents work as a chain
A segment changes, the offer updates, the message regenerates. No relay race.
Always current
Agents re-run when data changes, not when the calendar says so.
People stay in control
Marketers approve content and budgets; agents do the heavy lifting.
It gets smarter
Every campaign result feeds back into the next cycle.
Inside the engine
A multi-agent architecture on Google Cloud data, Databricks and Gemini, with governance and human approval built in.
Business data inputs
Customer & loyalty (CRM)POS transactions (BigQuery)Clickstream (Cloud Storage)Campaign history (BigQuery)Product catalogue & brand contentMulti-agent engine
Outputs
Weekly segments & personasBest offer per customerReady-to-send campaign variantsPushed to CRM / campaign toolMarketing dashboardCost & quality monitorExpected business impact
| Dimension | Before | Expected with the agent chain |
|---|---|---|
| Time to launch | 3–4 weeks: brief, agency, review, launch | Days: agents write, marketers approve (70–80% faster) |
| Content volume | One message per campaign, the same for everyone | Many versions per segment per channel, written and tested automatically |
| Targeting | The same message to the full list | Live segments and the best offer per customer, refreshed weekly |
| Promo spend | The same discount for all, with 15–20% wasted | 2–4% of promo budget back as margin, with the discount matched to each customer |
| Personalisation reach | Only the big campaigns, because manual effort limits it | Every send personalised; extra versions cost tokens, not headcount |
| Brand consistency | Depends on who reviews | Grounded in brand knowledge, with automatic quality checks on every output |
| Learning | Quarterly campaign reports | Results feed the agents weekly, so every cycle gets better |
Expected impact based on the engine's design and typical marketing benchmarks. Actual results depend on each client's data, channels and promotion strategy.
Working agent demos
Three demos of the same pattern: an LLM agent that calls real analytical models as tools.
Every number in an answer comes from a tool or a governed table. The language model plans and explains; it never invents figures.
Marketing Intelligence Agent
Ask: "Which accounts should we prioritise next quarter, and what is the upside?"
An LLM orchestrator works through a B2B growth question using four tools. It looks up segments, pulls customer lifetime value, runs a CLV what-if on a survival and growth model, and searches internal documents. It then reasons the results into one recommendation, with sources.
Media Mix Agent
Ask: "What happens to revenue if we move 15% of search budget into video?"
Built for a D2C e-commerce brand on top of a marketing mix model. The agent reads channel performance, MMM contributions and a scenario simulator, then checks campaign briefs, creative learnings, media plans, the promo calendar and pricing guardrails before it answers.
Pipeline Attribution Agent
Ask: "Which touchpoints actually create pipeline, and what if we cut events?"
A B2B agent built on Markov-chain attribution at MQL and opportunity level. Its what-if tool uses the removal effect: take a channel out of the journeys and measure how much pipeline disappears. That gives a defensible answer to 'what is this channel worth?'
How the agents are built
The same reference architecture sits under every demo, and under the agents we build for clients.
Governed data
Customer, campaign and pipeline tables in Unity Catalog, with access control and lineage.
Models as tools
CLV, MMM and attribution models served as tools the agent can call, tracked in MLflow.
Document search
Briefs, playbooks and results indexed with vector search, so answers cite their sources.
Orchestration
A frontier LLM plans the steps, calls tools in order and reasons over the results.
Evaluation
Test questions with expected answers, so quality is checked before every release.
Want an agent on your own models?
We can scope a pilot on one decision, such as budget allocation, account prioritisation or pricing, using your data.