Building the Growth Director: how we designed an AI that thinks like a CMO
When we started building Grovar, we made a mistake most AI product teams make: we tried to automate the tools before we understood the thinking.
We spent weeks wiring up APIs — Meta Ads, Google Ads, analytics platforms. We could launch a campaign programmatically. We could pull performance data. But we couldn't answer the question that actually matters: given everything we know about this business, what should we do next?
That's the question the Growth Director Agent is designed to answer.
What a CMO actually does
To build an agent that thinks like a CMO, we had to deconstruct what that actually means.
Strip away the meetings, the decks, the stakeholder management — at the core, a good CMO does three things:
- Translates business objectives into growth hypotheses — "We need 30% revenue growth" becomes "We need to increase AOV by 15% and reduce CAC by 20% in Q3."
- Allocates resources across the highest-leverage experiments — not spreading budget evenly, but making asymmetric bets based on what they know about the business and market.
- Reads signals and adjusts — killing what isn't working, doubling down on what is, faster than competitors.
None of these are creative. All of them are analytical. And that's what made us think AI could do them.
The architecture we landed on
The Growth Director is not one model doing everything. It's a coordinator.
When you give Grovar a goal, the Growth Director takes that goal and decomposes it into a structured execution plan — specific tasks for each specialist agent, with clear success criteria and priorities.
Goal: Increase revenue by 30% in 90 days
↓
Growth Director: builds plan, assigns tasks
↓
Campaign Agent: launches experiments on Meta + Google
Audience Agent: builds and refines targeting segments
Content Agent: generates creatives and copy
Analytics Agent: tracks ROAS, CAC, CTR in real time
Customer Agent: handles inbound leads and DMs
Each agent has a narrow responsibility. The Growth Director is the only one that sees the whole picture.
Why we used Claude
We evaluated several models for the Growth Director role. The criteria were specific: the model needed to produce structured, consistent output — not creative writing, not general conversation, but disciplined JSON that downstream systems could rely on.
Claude (Anthropic) won on two dimensions: instruction-following reliability and reasoning transparency. When the Growth Director produces a plan, you can see the logic. The confidence level, the rationale behind channel allocation, the specific metrics it's targeting — it reasons out loud, which matters enormously when you're trusting an AI with marketing budget.
We also needed the output to be actionable immediately. Not "consider testing Facebook ads" but "launch three ad sets targeting women 25-44 in these five US metros with a combined budget of $1,200/month, optimising for purchase events." The specificity is the value.
The hardest problem: business context
The Growth Director is only as good as what it knows about the business. A plan for a £50k/year Etsy shop should look completely different from a plan for a £5M D2C brand — different channels, different CAC targets, different timeline expectations.
Early versions of the agent produced reasonable-sounding but generic plans. The kind of thing a junior consultant might produce after a two-hour briefing.
The fix was forcing specificity in the input. Business type, products, margins, current revenue, ad budget, target channels. The more precise the context, the more precise the plan. This is obvious in retrospect — you can't reason well about things you don't know — but it took a few iterations to build the input model that reliably produces business-specific output.
What we're building toward
The Growth Director today is reactive — you give it a goal and it builds a plan. The version we're working toward is proactive. It monitors performance continuously, detects when something is underperforming, generates a hypothesis for why, and proposes an adjustment — without being asked.
That's the CMO analogy completing itself. Not "here's a plan when you ask for one" but "here's what's happening and here's what I think we should do about it."
We're not there yet. But the architecture is designed for it.
If you want to see the Growth Director in action, the fastest way is to try it. Set a real goal for your business and see what it produces.