Context: In July 2026 the Coalition for Innovative Media Measurement (CIMM) published “Models as Masters? Marketing Mix Modeling and AI-Driven Media Decision Making”. Its central warning deserves to be taken seriously: as MMM outputs get wired into planning systems, automated optimization, and agentic buying, the assumptions inside a model stop being argued over by analysts and start allocating budgets at scale. We agree with most of the diagnosis although we reach a different conclusion about the cure.
Marketing Mix Modeling (MMM) is safe to use in an AI-driven world but only under two conditions: the model runs on a platform whose data, math, and assumptions can be inspected by the people relying on it, and a human still owns the decision. The risk worth worrying about is not that MMM is a model rather than a fact. Every measurement method is a model. The real risk is automation acting on any estimate as if it were the truth. Improve the transparency and accountability of the system and MMM becomes more trustworthy under automation, not less and more importantly, AI, when used correctly, is what makes that transparency possible.
The report's core move is correct, even if we would frame it differently. MMM is shifting from what many treat as a rear-view analytical tool into decision infrastructure. The caveat we would add is that MMM was always meant to inform decisions: treating it as a purely backward-looking exercise was never a function of MMM itself, but a shortcoming of some of the people using it. Now more than ever, MMM outputs have the potential to shape budget allocation, feed optimization systems, and most likely will feed more and more autonomous buying agents over time.
This is a strong point, and we are glad to see an organization such as CIMM making it. We also agree that the governance principles CIMM lands on are the right ones: transparency, auditability, data completeness, separation of measurement from media-selling, and human oversight. The report draws a sharp line between two ways automated MMM can be built: one where the model's logic stays opaque and decisions execute without anyone checking the reasoning, and one where every input, assumption, and output stays inspectable and human-owned. The one caveat we would add is that transparency as a concept really ought to have been a priority all along.
Our opinions differ on two things: what is actually new here, and how long the fix takes.
“MMM does not produce objective facts,” the report says. That is true, and it is true of every method marketers have ever used. Even at the height of multi-touch attribution (MTA), before the “gardens became walled,” MTA never measured incrementality, so its “objectivity” was never as objective as it looked. A proper RCT (randomized controlled trial) is the gold standard for causality yet lift studies and matched-market tests in the marketing world are rarely able to meet the high technical bar of a true RCT, and even when they do, they rarely generalize cleanly to next year's plan.
The genuinely new thing in an AI-driven world that is different: automated buying will act on an estimate continuously, at scale, with no human pausing to ask whether the number is solid. Automated systems do not always buy the true optimum and that holds regardless of whether the signal comes from MMM, MTA, or a platform's own conversion API. The useful response is not to single out MMM as uniquely fragile. It is to make sure everyone relying on automation understands they are optimizing against estimates, and to give them the tools to see how good those estimates are.
This is where we most want to extend the report. CIMM seems to treat AI mainly as a risk multiplier. It can be. But the same technology, pointed the other way, solves a problem the industry has struggled with for over two decades: making MMM understandable to the people who act on it.
We launched our first MMM platform in 2006, and from day one it exposed all of its data handling and modeling methodologies to any client interested enough to look. For many MMM firms, certain methodological choices sat behind a veil of “proprietary math” but even so, if clients had internal experts, they were typically able to gain at least a high-level understanding of the methods. The gap was never capability. It was that non-technical decision-makers rarely knew what to ask, and rarely had the time to learn more, so the folks responsible for the decisions that MMM was driving never ended up with true transparency.
Agentic AI closes that gap- or at least it has the potential to do so. On a trusted platform, a less-technical user can now ask an AI assistant why a model produced a given result, and whether the finding has any obvious weaknesses, and get a clear answer grounded in the model's actual documented assumptions. Crucially, the agent is not doing the math or choosing the method, the platform did that, transparently and reproducibly. The agent explains. That distinction is what keeps it honest: an agent sitting on top of a well-built, deterministic model has nothing to hallucinate, because it is reading the model, not inventing one.
The result is a genuine best of both worlds. The trusted platform supplies rigor, documented priors, and reproducible math. The agent supplies plain-language explanation, handles the tedious work, and most importantly democratizes expertise that used to live with a handful of specialists. Over more than 20 years, the industry has produced countless “how to use MMM responsibly” guides. The responsible-use principles have not changed. What has changed is that AI can now enforce and explain them at the point of decision, for far more people than any pre-AI system ever reached.
The report is right that AI plus open source can be a risk, and can enable irresponsible MMM. It underweights the opposite:
CIMM's opaque-versus-accountable table is, read from the side of the MMM buyer, a checklist for choosing an MMM and optimization vendor in the age of automation. If you are evaluating one, ask whether it can:
A platform that meets this list is what the report calls an accountable system. The point we would add is that this is available now. It does not require the industry to first agree on shared standards, evidence libraries, and RFP templates, worthy as those efforts are.
MMM is safe to use with AI and automated buying when it runs on a transparent, inspectable platform with documented assumptions and a human owning the decision.
The real risk of automation is not that MMM is an estimate - every method is - but that automated systems act on estimates as if they were facts.
Used on a trusted platform, agentic AI increases MMM transparency rather than eroding it: the platform does the math, the agent explains it, and expertise gets democratized.
Judge any MMM vendor against an “accountable system” checklist: inspectable math, documented priors, uncertainty ranges, cross-media data, experiment calibration, human oversight, and separation of measurement from selling.