Marketing teams now ask ChatGPT and Claude about almost everything, including the budget. That raises a fair question: do you still need a marketing mix model, or can the assistant do the analysis itself? The short version is that the assistant is a very good way to use a model, and a poor replacement for one. This guide explains the difference, what MCP changes, and what to check before you connect an MMM platform to your own AI assistant.
A general-purpose assistant cannot run reliable marketing mix modelling on its own. Give it the same spreadsheet twice and you will get two different analyses. What it does well is act as the interface to a real model. Through MCP (the Model Context Protocol, an open standard for connecting assistants to other systems), Claude, ChatGPT or Copilot can call an MMM platform directly, run scenarios against your fitted model and return that model's numbers in plain language.
Marketing mix modelling estimates how each channel's spend drives sales, after accounting for the baseline, seasonality, delayed effects and diminishing returns. Doing that properly takes a fixed model structure, prior knowledge about how marketing works, enough history, and a check that the model forecasts sales it has not seen. It also has to give the same answer on Tuesday as on Monday.
An assistant asked to "do MMM on this data" improvises a new method each time. It may produce a plausible regression, but there is no stable model behind it, no record of the assumptions, no uncertainty you can trust and nothing to verify next month. That is fine for exploring an idea. It is not something to set a budget on.
We use AI heavily ourselves, to develop and improve the models and the framework around them. The decisions still run on a structured model with guardrails, and the assistant is how people talk to it.
MCP lets an assistant use tools in another system on your behalf. When an MMM platform offers an MCP server, the assistant can ask the platform to run a scenario, an optimisation or a forecast, and get back the same numbers the platform computes, with an explanation of what they mean. Nothing is invented in the chat.
In practice, that means a CMO or CFO can ask questions like these directly, without logging in to anything:
Behind each answer, the platform runs the actual scenario on the fitted model, including the simulations that produce the uncertainty range. The assistant explains the result and can follow up. Because the setup of the model is exposed as well, a finance lead can get a direct answer while an analyst digs into the assumptions.
The same connection works before a model exists. Once the data is collected, you can ask about your own history: what was spent where, how TV prices have moved, which ads the platforms report as strongest. The structured dataset is useful on day one.
Partly. The optimisation itself is done by the model: given a budget and constraints, it finds the allocation with the highest expected return. An agent can run that optimisation, compare several scenarios and write up the options. What it should not do on its own is move money in your ad accounts. The Odins MCP recommends how much to spend, where, and what to test. Acting on it stays with you, and our team reviews the monthly recommendations before they are delivered.
The list is growing quickly. As of autumn 2026:
The protocol is the same everywhere. What differs is what sits behind it.
When a vendor says its MMM uses AI, that can mean three different things, and it helps to separate them. The estimation itself is statistics: Bayesian sampling that fits carry-over, saturation and baseline with uncertainty. AI helps the people who build and maintain models work faster and check more. And assistants connected through MCP change who can use the model, from a few analysts to anyone in the leadership team. Only the first decides whether the numbers are right, which is why the questions about priors and verification in our vendor selection guide still matter most.
Not reliably on their own. They can explore data and explain concepts, but they do not hold a stable, verified model, so the same question can give different answers. Used through MCP on top of an MMM platform, they become a practical way to query a real model.
Odins, Lifesight and ScanmarQED offer MCP access to their products, and Google Meridian 2.0 includes an MCP-based module for teams running Meridian in-house.
Any MMM platform with an MCP server can be queried from Claude, ChatGPT or Copilot. With Odins, you can ask about budget size, allocation, scenarios and your raw marketing data in the same conversation.
Not with Odins. The model recommends; your team decides and acts.
To see what a conversation with your own model looks like, read about the Odins MCP, start with our guide to marketing mix modelling, or book a demo.