AI-Powered Growth: From Experiments to Systems
Anna Danyi
1 December 20245 min read
AI made it cheap to generate variants. That is not the same as making growth cheap. The teams that win treat models as workers inside a measured system: they produce options, suggest audiences, and compress busywork, while humans still own strategy, brand, and the definition of success. The teams that lose paste outputs into auctions and call the dashboard "learning."
This post is how we move from AI-assisted experiments to AI-powered systems without discarding rigour. It sits next to our production view on AI UGC ads, the AI creatives offering, and the broader stack in building growth systems that scale.
What AI is actually good for in growth
Three jobs show up again and again. Creative variation: scripts, hooks, visual directions, and UGC-style cuts at a volume no studio retainer can match week after week. Personalisation assists: drafting segment-aware copy or offers that a human still approves. Prioritisation: clustering qualitative feedback, summarising experiment logs, and ranking backlog ideas when the constraint is attention, not imagination.
Platform automation is part of the same story. Meta's Advantage+ style products and Google's App campaign automation already optimise delivery with machine learning; their help centres (Meta Advantage+ campaigns, Google App campaigns) are clear that the machine needs conversion signals and creative inputs. AI does not remove the need for a kill line — it makes a bad kill line more expensive because you can waste budget faster.
On the research side, foundation-model vendors publish capability overviews that are useful for scoping what belongs in-house versus in a vendor: see OpenAI's product documentation and Google's Gemini API docs. Use them as engineering references, not as proof that your CPI will fall.
Keep humans in the loop on purpose
Our default loop is simple: models generate, humans gate, systems measure. Strategy stays human — which markets, which promises, which constraints. Brand and compliance stay human — especially in health, finance, and kids categories. Interpretation stays human — a model will happily explain a random spike with a confident story.
That is not Luddism; it is how you avoid silent failure. Creative fatigue still exists when AI floods the auction with near-duplicates. Our creative fatigue piece is about decay curves; AI changes the supply of creatives, not the physics of audience saturation. Diversity of concept still matters more than diversity of caption.
Practically, we keep a human checklist before anything scales: does the hook match the landing promise, is the claim substantiated, is the success metric defined, and is there a kill threshold? The hook analyzer exists to make the first question faster. The D7 ROAS kill line exists so the last question is not a meeting.
From one-off experiments to always-on systems
An experiment is a ticket. A system is a pipeline. AI-native growth means the pipeline learns from every interaction and feeds the next batch of variants without a heroic brief every Monday. Exposure data, winning hooks, failing angles, and retention outcomes should land in a structured log that the next generation step can read.
Build that in stages. First: AI-assisted production with manual upload and manual reading. Second: templated pipelines (script → voiceover → edit → QA) with shared scoring. Third: closed-loop systems where losers are auto-paused against your rules and winners spawn controlled mutations. Jumping to stage three without stage one's measurement is how brands get surprising ads and unsurprising refunds.
Answer engines add another surface. Users ask ChatGPT and peers which app to download; that is a discovery channel with different rules than Meta. Our AEO guide covers how to show up there. Treat it as part of the system — content and product truth that models can cite — not as a separate "AI project" with no owner.
Guardrails that keep AI growth honest
- 01
Define success before you turn the model on
If you cannot state the KPI and the kill line, you are not ready for volume.
- 02
Keep a holdout or calibrated control
when the change is personalisation or large-scale creative automation, so you can estimate lift instead of storytelling.
- 03
Audit outputs for brand and compliance
on a sample every week; automated QA catches some issues, not tone-deaf claims.
- 04
Separate platform-reported ROAS from finance ROAS
Modelled conversions plus AI delivery is a double abstraction — manage to MMP or internal truth, as in what is a good ROAS.
- 05
Watch concept diversity
, not just asset count. Fifty variants of one weak idea is still one idea.
Industry experimentation literature keeps repeating the holdout point for a reason; see Optimizely's experimentation overview for the classic framing. TikTok's creative guidance in the TikTok Creative Center is also a useful reminder that platform-native patterns still beat generic AI sludge.
How we install this at Exp(G)
On engagements we wire AI into the growth system rather than beside it. Unit economics first (Payback Engine, benchmarks), then creative velocity through AI creatives, then weekly rituals that decide fate. Agency and Growth Engine engagements differ in depth; the measurement spine does not.
The destination is not "remove the growth team." The destination is a team that spends its hours on judgement — offers, positioning, product constraints — while the system handles repetition at a pace the auction demands. If you want that installed on your stack, book a discovery call and bring one funnel metric and your current creative throughput per week.
Sources & further reading
Anna Danyi
Founder at Exp(G) — building and scaling mobile apps with AI-powered growth systems. About the team