What happened:
Curtis Howland, co-founder of DTC ad agency Misfit Marketing, posted the 5-layer AI workflow he uses to manage $3M/month in Meta spend across multiple brands.
The stack: AI-tagged creative across 40+ dimensions, a gap-analysis heatmap, a routing layer that decides which client's data to pull, SQL queries instead of LLM math, and a verification loop on every output.
His thesis, from a recent appearance on the AI Marketing Pioneers podcast: “AI makes me faster at finding insights. It doesn’t replace knowing which ones actually matter.”
More Insight:
How he does it: Howland’s workflow, in order:
Every static and video ad gets run through an AI model and tagged on 40+ dimensions: format, selling point, who’s in it, setting, audio transcript. Performance gets pulled against those tags. He can then draw conclusions like “UGC about pain point in kitchen outperforms founder ads in office about features by 40%.”
The tags build a creative heatmap, with selling points (e.g., for toothpaste: “whiter teeth” or “better breath”) on one axis and formats (UGC, founder interview, on-the-street) on the other. Gaps reveal combinations he hasn’t tested yet. Winners get scaled across the other axis. “Without this you either repeat what worked until it fatigues or throw random stuff at the wall.”
A routing layer sits in front of every AI query. Its job isn’t to answer; it decides which client’s data to pull and which sources to read (call transcripts, Slack threads, Meta data, internal DB). “We manage multiple brands and cannot have the wrong client’s data leaking into analysis.”
AI writes SQL against a rebuilt data warehouse instead of analyzing raw exports. SQL forces segmentation. “I never let it analyze ad copy, CTAs, and landing pages as one aggregate. Break it by persona.”
Every output gets checked. “Even one extra loop of ‘is this accurate?’ catches most hallucinations.”
Zoom out: The heatmap exists because Meta ads fatigue fast. Howland says the average ad loses 25% of its performance the month after launch, so “you have to replace basically a quarter of your ads every month if you just want to stay flat.”
What he’s saying: AI’s job, Howland says, is to make humans more effective, not replace the final call. “It can give us good data. It can give us good ideas. But it’s still up to us to ensure how we implement best to reduce cost, maximize output, and overall drive profit.” He’s wary of acting on every suggestion the model generates. “I could spit out a hundred ads before I finish the sentence, but it’s up to me to ensure we’re balancing testing costs and output.”
The big picture: Howland frames the risk of AI in advertising with a line from The Incredibles: “Once everyone’s a superhero, nobody is.” His version: “Once everyone is perfectly generic, nobody is interesting.” If every advertiser uses the same models the same way, output converges to the middle. The differentiator is the brand-specific context being fed in, and the human judgment about which AI insights to act on.
Yes, but: Howland says the infrastructure under the stack is the hard part. Misfit rebuilt its data backend before AI could write SQL against it, and on the podcast he said keeping the infrastructure current is “really challenging.”
Looking ahead: Howland sees three layers AI has to climb. The numbers (already there). The middle context of whether the data is even accurate (getting there). The high-level strategy of seasonalities, competitor moves, and brand differentiation (the last to fall).