Can you use AI to build an MMM? Yes. But picture how it can go wrong.

Say you hand an AI coding agent your spend, your sales, and an open-source MMM. Say it hands back a model that passes its health checks, and podcasts show a crazy high iROAS (incremental return on ad spend).

So do you move budget into podcasts?

Not yet. AI is very good at building and checking a marketing mix model. It can’t make an iROAS true. A strong out-of-sample fit doesn’t confirm a channel’s return. A lift test is how you check it.

Here’s how I split the work. The AI builds the model, runs the checks, and flags where its confidence is high or low. Lift tests anchor the channels they’ve tested. Untested channels get marked low confidence. And a human signs off on every budget change.

If you’re still deciding whether you need a model at all, start with do you need an MMM yet?

Where does AI help build an MMM?

AI helps with the build and the checks. It can set up an open-source MMM from your data and take in your promo calendar and seasonality. It can suggest decay rates. It adds test results in the form each library takes them. Then it runs data checks that flag where its confidence is high or low. It depends on what you give it. What it can't do is make a channel's iROAS true. Lift tests check that, and a person signs off on every budget change.

Here’s the workflow I use:

  1. Build from open-source code. I’ve used Meridian (Google), Robyn (Meta), and PyMC-Marketing. I built all three with Claude Code working from their GitHub codebases, with data pulled from Google BigQuery and Google Sheets.
  2. Add context. Promo calendars and known seasonality go in, so the model doesn’t credit a sale to the wrong cause.
  3. Suggest decay rates. The decay rate describes how long an ad’s effect lingers. In model terms, it’s how much of this week’s effect carries over into next week (often called adstock). I estimate it from click journeys, post-purchase surveys (“When did you first hear about us?”), and geo tests. A geo test can show how long sales take to move after upper-funnel media starts.
  4. Add test results as priors, where the model accepts them. (A prior is a starting belief the model weighs against the data.)
  5. Run the checks. I rely on AI to take all of this in, run data checks including out-of-sample tests, and say when and why its confidence in each iROAS is high or low.

That last step matters most. You want a model that tells you where it’s unsure.

Why does the decay rate matter so much?

Because upper-funnel media can take months to show up. A vendor case study shared in August 2026 tracked a 26-week upper-funnel YouTube geo experiment (Dec 2025 to May 2026). Cumulative lift hovered around zero for roughly the first five weeks. About 87% of the measured effect came after the first two months. It reached about half its final level around three months in.

Upper-funnel impact also needs commitment. In The Effectiveness Code (Field & Hurman, 2020), some campaigns scored higher on budget and on number of media channels. They also ran longer. Those campaigns had higher average effectiveness scores. An illustrative TransUnion chart in the IPA’s Econometrics and the C-suite (2023) makes the same point. About 90% of performance media’s impact shows up by week 4. Brand media takes about 10 weeks to get there.

So if the model’s decay rate is too short, it undercounts upper-funnel media. And a four-week test may end before the effect even shows up.

AI does the heavy lifting on the build. It doesn’t get the final word on what’s true.

For the inputs themselves, see what should I put in my MMM?

What can AI and MCP tools do for MMM and incrementality testing?

They let an AI assistant work with your data and a fitted model. So it can pull numbers, run checks, and explain what the model says in plain language. They don't make an iROAS true, and they don't replace a lift test. A person still signs off on budget.

MCP is the Model Context Protocol, an open-source standard for connecting AI applications to outside systems like data sources, tools and workflows (modelcontextprotocol.io, checked October 18, 2026).

As of October 2026, Google's open-source Meridian has its own add-ons for AI coding assistants, called Meridian agent skills, and they pause for your review at key steps.

Where do these tools help?

Where don't they?

They can't turn an untested channel's iROAS into a fact. They can't pick the comparison for a test, decide what the business does with a result, or sign off on a budget move. And a scenario or "optimizer" result is what the model's fitted response curves imply under its assumptions (by default, your past weeks, spend pattern and costs). It's a conditional estimate, not a forecast.

Before you trust any AI or MCP tool with your model, ask:

For the context an AI needs to answer well, see what an AI agent needs before it can answer a marketing question and the marketing context layer template.

If you want a second pair of eyes on what an AI tool told you about your model, bring it to me.

Does a good fit mean the iROAS is true?

Say a team hears this in a review: “The model passed every check. Holdout accuracy looks great.”

So they might think the channel numbers are right. But a model can predict total sales well and still split the credit wrong between channels.

An out-of-sample test (hiding some weeks from the model, often the most recent ones, then checking how well it predicts them) tells you the model fits your business overall. It doesn’t tell you whether podcasts or paid social earned that fit. Passing health checks, out-of-sample prediction included, doesn’t make an iROAS accurate.

A lift test is how you check it.

How do lift tests anchor the model?

Run a lift test on a channel, then give the result to the model. As of October 2026, how depends on the tool:

  1. Test the channel with a geo holdout or another lift study.
  2. Give the model that result. Meridian takes it as an ROI prior for the channel. Robyn uses it as a calibration target: one of the scores it tries to shrink is the gap between the model and the test (MAPE.LIFT). PyMC-Marketing can add lift-test results to the model’s likelihood (the part that scores how well the model matches the data).
  3. Let the model fit everything else around it.

When an MMM accepts test results, the test anchors that channel so the model can’t drift too far. The rest of the portfolio still comes from the data.

It’s worth saying that not every test belongs in a model. A test run in an odd season, or one that was underpowered, may only speak for that moment. See what an incrementality test can calibrate in an MMM.

A tested channel is anchored. An untested one is still a guess with good manners.

What about channels you haven’t tested?

Mark them.

Some channels haven’t been tested yet. Some may be hard to lift test because of how the channel works, like national TV, affiliates, and influencers. For those, the model’s iROAS is a read, not a finding.

Here’s a typical scenario a brand may face (a hypothetical, not a client result). Picture the model handing back a sky-high podcast score. Here’s how I’d put it: “Podcasts have a crazy high iROAS score, but we’ve never pushed the channel to any significant volume, we’ve never seriously varied the spend, and we don’t have a geo test. So we will simply mark this score as low confidence.”

No volume. No variation. No test. That’s three reasons to doubt the number, no matter how good the fit looks.

Put the confidence label right next to the iROAS in every report, so nobody reads a low-confidence score as a green light. If a channel can be tested, check whether a geo test is feasible.

When is an MMM read enough to move budget?

I think of model reads like weather forecasts. You don’t cancel a wedding on a 30% chance of rain. You might bring an umbrella.

Take an MMM read seriously for a budget change only if at least one of these holds:

  1. The change is small as a percentage.
  2. The change is small in dollars. Doubling podcasts from $1k to $2k a month, when podcasts are under 1% of the budget, is a cheap way to learn.
  3. The MMM has high confidence in that channel.
  4. You can run a lift study on the channel, or recently did one, to inform the change.

Ideally more than one holds at once.

A small bet on a low-confidence channel is fine. A big shift on a low-confidence channel is how a model ends up running your business.

So who makes that call?

Who signs off on budget changes?

Who approves the move: the AI or a person?

A person. Every time.

A human signs off on every budget decision. The AI can build the model, check it, and explain its confidence. But it doesn’t know what’s changed since the data was pulled: a promo added last week, a channel that just changed its pricing, or a CFO who needs a smaller bet this quarter. You do.

If you want help setting up that review, see marketing mix modeling for DTC brands.

FAQs

Can AI build a marketing mix model?

Yes. Start with an open-source MMM codebase, an AI coding agent, and your own data. I’ve built Meridian, Robyn, and PyMC-Marketing models this way with Claude Code. You still need to supply context (promos, seasonality, decay) and review the result.

Is a high out-of-sample score enough to trust a channel’s iROAS?

No. It shows the model fits your total sales. It doesn’t confirm how credit is split. A lift test is how you check it.

What should we do with a channel that’s hard to test?

Keep it in the model, label its iROAS low confidence, and only make small changes on it. Comparing more than one model can also help. See why use multiple MMM models.

Should AI make budget changes on its own?

No. A human approves every budget change. No model owns the truth.

Need help? Speak with me here.

Next, what to put in an MMM to assemble the outcome, media, and context records the model needs, why multiple MMM perspectives help to pressure-test a consequential model recommendation, or what a test can calibrate in an MMM to decide how much authority a test result should have in a broader model reading.