Agnes AI is an early-stage AI application. When it came to Inpander it was adding about 30 new users a day, the classic case of a product that works and growth that has not started. Six months later that number was 70,000+, CPA had fallen by 65%, cumulative impressions had passed 300M, and the app sat in the Top 5 of its category in the target market. None of it came from one lucky viral moment. It came from a data-driven creative iteration machine running continuously.
The challenge
Growth in the AI app category has a characteristic problem. The product is new, users have no established search habit for it, and demand has to be educated into existence through content. Meanwhile the category is crowded, media costs keep rising, and an early team cannot afford broad, unfocused paid media.
For Agnes AI, starting at about 30 daily new users meant there was no historical data to lean on: nobody knew which audience would convert, which way of framing the value would land, or which psychological price point would move people. The traditional route is qualitative research first, then a creative direction. That is too slow. The right move is to engineer the search for answers itself: make every asset a hypothesis test and let the media system do the research in the live market. In other words, the growth team was not building an ad production line. It was building a learning machine.
The strategy
Inpander built three interlocking modules for Agnes AI.
First, the hook factory. The core formula is one script, five hooks. The product narrative stays stable while the opening three seconds are varied at scale, with dozens of new variants going live every week. The benefit is a single variable: when an asset wins, the win can be attributed to the hook itself rather than muddled with script, creator or editing differences. Winning hooks are banked and become the starting point for the next production round.
Second, paid injection. Content that breaks out organically is never left to run on its own. It is immediately amplified with Spark Ads, with dark posting under the creator's handle so the ad runs with the creator's identity without appearing on their profile, and lookalike audience targeting layered on top to push the winning asset to new users who most resemble those who already engaged. Organic reach validates the content, paid media harvests the scale.
Third, the data loop. Budget allocation runs on a 24/7 rhythm. Underperforming assets are cut within the hour, never allowed to keep consuming budget, and high-ROAS assets get more. Every round of media data flows back into the next creator brief, so the whole system gets more accurate the longer it runs. The 65% reduction in CPA is, in essence, the money that used to be wasted on failed assets, recovered piece by piece.
The results
- Daily new users up from about 30 to 70,000+
- CPA down 65%
- 300M+ cumulative impressions in six months
- Top 5 in category in the target market
Source: Inpander campaign review.
Key takeaways
- What an early product lacks most is not budget but learning speed. Dozens of hook variants a week buy the most market feedback for the least money. Test density decides how fast you find a winner.
- Hunt breakout content, do not admire it. The window in which a post is organically taking off is short. Spark Ads plus lookalike targeting extracts its commercial value before the momentum fades.
- Discipline in cutting losses is rarer than skill in scaling. Most teams cannot bring themselves to switch off mediocre assets and keep saying "one more day". Hourly cut-offs are the single biggest source of the CPA drop. Cutting losers and scaling winners are two sides of one coin: without the first, there is no budget for the second.
If your AI product is stuck in early growth and you want to build a creative testing machine of your own, write to hello@inpander.com or talk to the Inpander team.