AI ad production: evaluate the cost of finding a winner
Cheap production does not necessarily make a winning ad cheaper to find. Our AI creative discussions recommend evaluating production cost and testing spend together, then checking whether additional assets improve the results that matter.
By NewForm · Updated
Key takeaways
- Measure the cost of finding a useful winner, including both production and paid testing.
- A lower hit rate can outweigh savings on individual assets; it is a scenario to evaluate, not an inherent property of AI.
- Give automated analysis account context and check its explanations against the underlying evidence.
Production volume is only part of the cost
Our videos argue that easier production shifts attention toward distribution and creative strategy. Producing more ads is valuable only if the account can evaluate them and use what it learns.
The same question applies to human and AI workflows: does the additional volume create better opportunities, or simply increase the budget needed to search for one? The answer depends on the quality, production cost and test cost of the assets in that workflow.
The testing arithmetic of lower creative hit rates
One video uses a distribution sketch to illustrate how a lower average quality could reduce the number of exceptional ads. Its move from roughly one winner in forty to one in a hundred is hypothetical, not a measured comparison proving that AI creative performs worse.
A simpler way to evaluate the scenario is to estimate the production and test cost per asset, then ask how many assets you expect to evaluate before finding a useful winner. Holding per-asset cost equal, testing a hundred assets costs more than testing forty. Whether cheaper production offsets that difference depends on the actual costs and observed hit rates.
The clip also suggests that higher acquisition costs can make testing expensive relative to production. It does not provide enough evidence for a universal CAC threshold at which AI production stops being economical.
Give automated evaluation the context it needs
Automated analysis can mistake a visible correlation for an explanation if it lacks relevant account context. A background color, creator or format may appear in several winning ads without being the reason they won. This is also a risk for human analysis; AI systems are not inherently unable to use prior context.
Our proposed workflow makes the purpose of tests explicit and separates testing results from scaling results. Use model-generated observations to form hypotheses, then check the underlying assets, delivery and downstream outcomes before turning them into creative rules.
Adapted from NewForm’s original videos on creative strategy and paid social.
Explore NewForm Framework →