Your AI-Built Marketing Mix Model Is Fast. But How Do You Know It's Right?

Full analysis

The news here is small, and the speaker has a book to sell. Madan Bharadwaj, founder and CEO of measurement firm M-Squared, went on AdExchanger's Inside the Stack to argue that AI makes marketing mix modeling faster, which makes human experts more valuable. The episode is sponsored by his own company. So treat every claim as a sales pitch until proven otherwise.

But there's a real signal buried in the pitch. A beauty brand CMO built his own marketing mix model with free open-source tools and an AI assistant, got a nine-figure budget recommendation, and acted on his gut instead. Marketing mix modeling, for the record, is a statistical method that estimates how much each channel (TV, search, Amazon, and so on) actually drives sales. It used to take six months and a team of PhDs. Now a CMO can run one over a weekend. That part is true and it matters.

What's being decided: Nothing, by the reader, today. This is a read on where measurement spend and trust are heading. Easy to undo in the sense that no operator has to act this week. But the underlying shift, cheap DIY modeling, is already loose in the market and hard to put back.

What sets the deadline: Bharadwaj's own claim that "next year" causal measurement goes mainstream. Convenient timing for a vendor, but the DIY tooling is here now regardless.

The Skeptic

This is a vendor telling you the robot got it wrong and the human (who happens to run a paid managed service) got it right. Convenient. The whole Pacifica Beauty story is a sales narrative: cheap tool says cut $30 million from Amazon, CMO panics, hires M-Squared, M-Squared says keep spending, revenue goes up 5 to 7 percent. Every number is self-reported and unaudited. Nobody shows the counterfactual where the brand actually cut the Amazon spend and measured what happened. The "5 to 7 percent top-line growth" could be the overall market, a new product launch, or seasonality. For an operator deciding whether to buy managed measurement over a dashboard, this proves nothing except that granular models give different answers than coarse ones, which everyone already knew.

The Operator

Here's what actually happened Tuesday morning, and it's the part that should worry any measurement vendor. A CMO with ~200 products ran open-source modeling himself, no data scientist, no six-figure contract, and got to a nine-figure recommendation. That genie is out. The model was probably wrong, but the behavior is now normal. Every DTC and retail brand finance team will start pressure-testing channel budgets this way, and the first casualty is retail media spend, because Amazon is the easiest channel to flag as "incremental or not." Publishers and SSPs with heavy DTC advertiser bases should expect more "prove it's incremental" conversations and fewer always-on buys. The second-order effect at 90 days: budgets get yanked mid-flight on the strength of a model nobody validated.

The Customer / End User

Put yourself in the CMO's chair. You get a model that says cut $30 million. Do you trust it? Bharadwaj's actual useful idea is the test he uses: does the model's answer match your go-to-market story? If a model tells you to kill your best-performing channel and you can't explain why in business terms, the model is probably broken. That's genuinely good advice, and it's free. The CFO conversation is the other real thing here. Translating marketing into contribution margin (revenue minus the cost of goods and shipping) gives marketing and finance a shared number. CFOs have wanted that for twenty years. If AI-cheap modeling finally delivers it, that's a durable change in who controls marketing budgets, and the answer is finance.

The Market Analyst

For public measurement names, the strategic question is whether cheap DIY modeling commoditizes the dashboard or raises the value of the service wrapped around it. Bharadwaj is betting on the service. He's half right. The dashboard-only vendors (the ones selling a login and a chart) are the exposed ones if any CMO can spin up Robyn, which is Meta's free open-source modeling tool, over a weekend. The ones with embedded analysts and managed delivery have a story. But "we're more necessary because the robots are faster" is exactly what every incumbent says right before a cheaper tool eats their floor. The real pressure lands on the mid-tier: good enough to charge a premium, not deep enough to justify it against free tools plus a competent in-house analyst.

The CFO

The expensive part of measurement was never the model build. It was trusting the output enough to move money. AI collapsed the cheap part (the build) and left the expensive part (the judgment) untouched. So Bharadwaj's core claim is actually sound: if building is free, the premium moves entirely to validation. The problem is his economics. Managed services don't scale like software. Building "hundreds of SKU-level models" with humans in the loop is a consulting business, and consulting businesses carry people, not margins. For an operator deciding build versus buy: the cheap path (open-source plus an AI assistant plus one skeptical analyst) now gets you 80 percent of the way, and the last 20 percent is the most expensive mile anyone sells.

Where the council splits

Two real disagreements. First, the Operator and the CFO both think DIY modeling is now good enough to change behavior, while the Skeptic says the Pacifica story proves the opposite, that DIY produced a dangerously wrong answer. They're both right, and that's the tension: the tool is cheap enough to use and not good enough to trust, which is the worst combination for anyone selling trust.

Second, the Market Analyst and the CFO disagree on who wins. The Analyst says the service wrapper survives. The CFO says a managed-service model can't scale and the real winner is in-house teams with free tools. The hinge is whether validation can be productized or whether it stays a human consulting job forever.

What it actually hinges on

Whether "model validation" is a software problem or a people problem. If it's software, the dashboard vendors build it in and the managed-services pitch collapses. If it's people, Bharadwaj wins but in a low-margin consulting business, not a scalable one. Everything else in the episode is methodology and philosophy with a sponsor's logo on it.

Where the council leans: the DIY shift is real and permanent, the specific "AI was wrong, hire us" narrative is a sales story, and the pressure lands hardest on retail media budgets and on mid-tier measurement vendors who sell dashboards without analysts.

Prediction: At least one publicly traded measurement vendor (DoubleVerify, Integral Ad Science, Comscore, iSpot, or Nielsen) will announce an AI-assisted marketing mix modeling or incrementality product by its Q2 2027 earnings call in August 2027.

Confidence: Medium. The tooling is free and the competitive pressure is obvious, but timing across a specific vendor set is the soft spot.

Why: Open-source modeling tools like Meta's Robyn, run through an AI assistant, now let a single CMO build a channel model over a weekend, which hollows out the floor under any vendor selling a dashboard and a login. The measurement incumbents have watched this category move from six-month builds to days, and the defensive move is to wrap the cheap modeling in their own branded product before clients route around them entirely. The opposite outcome, that none of these five launches anything in this space, would require the public measurement names to ignore a free tool actively taking conversations out of their hands, which is not how incumbents under margin pressure behave.

Revisit by 2027-08-31: We're right if any of DoubleVerify, Integral Ad Science, Comscore, iSpot, or Nielsen ships or announces an AI-built MMM or incrementality offering by its Q2 2027 earnings call. We're wrong if none of them does.