AI agents use marketing mix modelling (MMM) in two ways: they help analysts build and check the model, and they read its outputs, chiefly marginal return on investment and response curves, to propose or make budget decisions. The first use is now documented in open-source tools such as Google's Meridian and PyMC-Marketing. The second changes what the model is for, because an estimate built to inform a quarterly plan becomes an input to software that can act every day.

TL;DR. Meridian and PyMC-Marketing publish agent skills that guide data preparation, model fitting and budget optimisation, and Meridian's skills pause for human confirmation at set checkpoints. An agent reading a finished model should work from marginal ROI and its uncertainty, keep spend inside the range the model has observed, and treat the model as one estimate to be checked against experiments. The questions a model cannot settle for itself, such as which assumptions hold and which of several well-fitting models to believe, stay with people and belong in the agent's mandate.

What does a marketing mix model tell an agent?

A marketing mix model estimates how much each channel contributed to sales or another outcome, using aggregated data over time. Meridian's documentation on GitHub describes MMM as "a statistical analysis technique that measures the impact of marketing campaigns and activities to guide budget planning decisions and improve overall media effectiveness", and notes that it does not use cookie or user-level information. The same page lists the questions it answers: how channels drove revenue, what the return on investment was and how to allocate future budget.

For an agent, the most useful output is marginal ROI. Meridian's glossary defines it as the derivative of the response curve, "approximately the ROI on the next monetary unit (such as dollar) spent beyond current spend level". Average ROI describes the past, while marginal ROI estimates where the next pound is likely to work hardest, which is the question an agent reallocating budget is actually asking.

What changes when an agent, rather than a planner, reads the model?

Three things change. The first is cadence: a model refreshed monthly or quarterly is now read by software that can move budget daily, so the gap between the evidence and the action widens. The second is feedback, because the agent's own spending becomes the data the next version of the model is fitted on.

The third is statistical. Meridian's guide to MMM as causal inference explains that valid results depend on the control variables including "all confounding variables, which are variables that causally affect both media execution and KPI". An agent that raises spend whenever sales rise creates exactly that kind of variable, so its decision rules need to be known to whoever builds the next model.

Feedback also makes the spend record matter more. If the model is told a channel received spend it did not receive, or in weeks it did not run, the estimate is wrong however careful the statistics. As a worked example, Alkimi records agent-negotiated terms on a DealSheet that is written at the moment of agreement, timestamped and held by both buyer and seller, so either side can retrieve what was bought and when, and the spend column of a model has a source both parties recognise.

How do AI agents help build a marketing mix model?

Open-source MMM packages now ship instructions written for AI coding assistants. Meridian's agentic skills cover five stages: a documentation consultant, model building, result visualisation, budget optimisation and a scenario planner. The documentation says the agent will pause for review before finalising column mappings, setting Bayesian priors, running long sampling jobs and executing generated scripts.

PyMC-Marketing takes a similar approach. Its mmm-modeling skill states that language models understand Bayesian inference in general but need domain-specific patterns, and lists errors an unaided model might make, such as holding out test data for the final fit or using flat priors on saturation parameters. Robyn, the open-source package from Meta Marketing Science, uses ridge regression and evolutionary hyperparameter optimisation, and its public code can also be run by a general-purpose agent.

Julian Runge, an assistant professor of marketing at Northwestern University, told AdExchanger in August 2026 that an agent can be asked to pull an open-source package, load the data and estimate the model, taking someone with little modelling background "almost from zero to one". He added a condition: "You still need human intelligence that can understand what's happening."

How are advertisers using agents with MMM outputs?

The clearest published example sits on the planning side. Digiday reported in May 2026 that Hershey spent a year developing a system of several AI agents that provide media modelling for campaigns "over the course of weeks, not months". Vinny Rinaldi, vice president of consumer connections at Hershey, told Digiday: "With a modernized MMM, I think you can build a planning agent around hyper-focused data and outputs."

Rinaldi described the system being used to forecast the revenue impact of moving budget between channels. He also argued that the brand should own such an agent, "because it's our data fueling it". In that arrangement the model informs the plan, people approve it and buying follows.

Which model should the agent believe?

An agent can fit many candidate models quickly, which makes choosing between them harder rather than easier. Meridian's causal inference guide warns that "multiple models can have good fit and predictive power yet provide different ROI and optimization results", and that a model with 99% out-of-sample R-squared can still be a poor model for causal inference. Fit statistics alone cannot pick the model an agent should act on.

Madan Bharadwaj, founder and chief executive of M-Squared, made a related point on AdExchanger's Inside the Stack in October 2026. AI has cut model building from months to days, he said, which makes expert judgement more important, and his method is to run MMM and incrementality tests side by side and see where they agree. He also warned against tidy results: "If you get a very clean read, you probably got something wrong."

Calibration is how the two are joined. Robyn's analyst's guide strongly recommends calibrating the model with experimental results treated as ground truth, such as conversion lift or geo-based incrementality studies, run on a regular basis. Meridian 2.0, released in September 2026, added native geo experimentation through Meridian GeoX and automated calibration that brings lift results into the model, and the design of those tests when an agent is buying is covered in how to measure incrementality when an AI agent buys the media.

How much autonomy should an agent have over MMM-driven budgets?

Autonomy should match the strength of the evidence. A useful framework is earned autonomy, which sets out five levels at which an agent can work with a model:

Observe: the agent reads model outputs and reports them, for example a weekly summary of marginal ROI by channel.
Recommend: the agent proposes a reallocation and shows the response curves and uncertainty behind it.
Draft: the agent prepares the revised plan or orders for a person to review.
Human-approved action: the agent executes only once a named person signs off.
Bounded automatic action: the agent moves budget on its own within limits agreed in advance.

Bounded action needs bounds the model can support. Meridian's budget optimiser by default lets each channel rise to no more than 130% of its starting spend in a fixed-budget scenario, which keeps recommendations near levels the model has observed. As a hypothetical, a model fitted on weekly connected TV spend between £20,000 and £60,000 says little about what £150,000 would do, and the agent's limits should reflect that.

Refresh schedules matter too. The IAB's Modernizing MMM Best Practices for Marketers guide, published in December 2025, advises balancing "speed and stability through frequent data refreshes and disciplined retraining cadences" and integrating MMM with attribution, experimentation and financial systems. IAB Australia's MMM landscape report makes the same point: MMM should not be used in isolation.

What should advertisers settle before an agent reads their model?

Four decisions come first: who owns the model and its assumptions, which autonomy level the agent starts at, which experiments will check the model and how spend will be recorded. Each belongs in the agent's mandate, as set out in how to write a media buying mandate. The agent can then move faster than a quarterly planning cycle without moving further than the evidence.

Rehearsing that arrangement before real money moves is what Alkimi's Gym is for: buyers and sellers build their own agents, set the rules they work to and spar against counterparty agents first. The model still needs people to decide what it means.

Frequently asked questions

How do AI agents use marketing mix modelling?

They help analysts build and diagnose models, and they read outputs such as marginal ROI and response curves to recommend or make budget moves. Open-source tools including Meridian and PyMC-Marketing publish agent skills for the first use.

Can an AI agent build a marketing mix model on its own?

An agent can prepare data, fit a model and run diagnostics, but choosing priors, control variables and the final model needs human judgement. Meridian's agent skills are designed to pause for confirmation at those points.

Which MMM output should a buying agent act on?

Marginal ROI is the most relevant, because it estimates the return on the next unit of spend rather than the average of past spend. The agent should also see the uncertainty around it and stay within the spend range the model has observed.

How often should a marketing mix model be refreshed when an agent uses it?

The IAB's MMM guidance recommends frequent data refreshes balanced against disciplined retraining. Whichever schedule is chosen should be written into the agent's mandate so the age of its evidence is always known.

Do agents replace incrementality testing?

No, because experiments are what calibrate the model and check that it is estimating causes rather than correlations. Robyn's and Meridian's documentation both describe bringing lift results into the model.