An AI Media Buyer Says It Cut CAC 30%. The Test Details Are Missing.

Notch says its AI media buyer reduced customer acquisition cost by 30% for a direct-to-consumer brand called Scentless. It released no spend level, CAC baseline, sales total, control group or attribution method.

The product claim and the performance claim need separate standards of evidence.

The company says its agent, Max, ran a two-month campaign with a Shopify connection, a Meta ad account and a folder of raw B-roll. It says Max analyzed trends, made video ads, replaced fatigued creative and reallocated budget across Meta and TikTok. The closing stunt involved launching a laptop running Max into space while the campaign stayed active.

Vinay Jain, Notch’s founder and CEO, calls it “an autonomous marketer with a continuous feedback loop.” His point is clear: the company wants to automate the work between an idea and a live campaign, not simply generate copy or images.

The product page gives a more grounded description. It says Max reads account data and competitor activity, posts a weekly plan in Slack, produces creative for review, moves approved ads live and reports on results. It also says nothing goes live without approval or preset rules. That is materially different from the release’s “minimal human oversight” framing.

The distinction is important. An agent can prepare a lot of work without owning the outcome. Someone still has to set the brand guardrails, choose which claims are legal to run, approve budget changes and decide whether a reported conversion is worth the acquisition cost.

Notch says it has shipped more than 50,000 ads for 4,300 brands and agencies. Its press release instead says 85,000 ads for hundreds of companies. The difference could reflect different reporting dates or definitions, but the company hasn’t explained it. That alone is enough to avoid treating either total as a clean market signal.

Nitzan Leibovitch, head of creative at Schemathics, said creative capacity had been the constraint and that the team can now move from an idea to a batch of video variations in one session. That’s the part a marketer can test now. If an agent cuts the time between a signal and a reviewable creative batch, it can increase the number of useful experiments.

The 30% CAC result needs a much higher bar. Ask for spend, period, channel split, holdout, creative approval rate, conversion window and the definition of CAC. Then compare it with a matched manual campaign.

Autonomous media buying will be measured on its exceptions, not its demo. A strong system should make its decisions visible, let the team reverse them, and show whether the next batch beat a real baseline.

What would make the CAC claim useful

The Notch release says the Scentless trial used a Shopify connection, Meta access and raw B-roll. It also says the agent refreshed decaying creative and moved budget. Those are multiple changes to one campaign. A lower customer acquisition cost could come from creative, audience selection, spend timing, offer changes, seasonality or an attribution setting.

The Notch product page describes a more ordinary workflow: read account data, prepare a plan, create a batch, move approved work live, then report the winners and losers. That is still valuable. A paid-social team can use an agent to reduce the time between a performance signal and an approved creative response.

Notch’s case-study library gives prospective buyers examples to review, but it doesn’t replace a controlled result for the Scentless claim. The company should publish the spend range, comparison period, creative count, approval rate and measurement window. It should also say whether the 30% number is blended acquisition cost, platform-reported cost or a first-order customer metric.

The working marketer needs a defined human approval boundary for every agent action. The same issue is visible in Amazon’s new seller workflows, where the seller can choose a recommendation or an action, and in AI-led lifecycle platforms, where a model chooses whom to contact.

Run a narrow comparison first. Give the agent one product line, a fixed spend ceiling and a preapproved set of offers. Keep a manual control campaign. Review every creative and budget action, then compare CAC, contribution margin, frequency and customer quality. That tells a team whether it bought additional learning or simply automated a larger volume of guesses.

The agent’s speed is plausible. The 30% result remains a vendor claim until the experiment is reproducible.