The Art of Subtraction: Training AI Agents in Marketing Mix Modeling with 'Via Negativa'

Most discussions around AI Agents focus on what the agent can do. While strictly speaking, the AI Agents don't have agency yet. They get the agency or pretend to have agency by virtue of what it is being prompted to do or not do.

Which brings me to the topic of How we at Aryma Labs train our MMM AI Agents.

Improvement through subtraction

We follow a philosophy called 'Via Negativa'. It is basically improvement through subtraction rather than addition or just 'What not to do'.

For many complex domains, it is easier to identify what is wrong than to precisely define what is right.

And MMM is a perfect example :)

Examples from our prompt templates

I will cite few examples from our prompt templates to illustrate the principle of 'Via Negativa'.

Example 1

An MMM model shows:

TV Contribution = 35%

Paid Search Contribution = 20%

An poorly trained AI agent may immediately conclude:

"TV is the most effective channel. Increase TV budgets."

A good MMM practitioner knows that this conclusion lacks nuance.

The agent should first ask:

• What are the current spend and saturation levels?
• What are the marginal ROAS ?
• Is TV already operating near diminishing returns?

Hence Via Negativa prompt is:

"Do not recommend budget increases solely because a channel has high contribution".

or

"Do not recommend reallocations without checking saturation, spend constraints and implementation feasibility."

Example 2

AI Agent : OOH only contributed only 1%, let's cut it.

The agent should know:

• OOH often acts as an amplifier and could manifest its effect through other channels.
• OOH may create long-term brand equity.

Via Negativa Prompt: "Do not recommend cutting channels before examining interaction effects and halo effects."

Example 3

AI Agent: Base is the top driver of your sales or seasonality is the most efficient channel

The agent should know that:

Base is not technically an active driver of your sales. It is the organic sales or brand equity. The sales that you get even when you have zero marketing or media spends.

Base hence can't be treated like a media channel. It is basically a manifestation of your previous marketing efforts (see related post in comment).

Similarly an agent should not construe seasonality as an active driver of sales. It is just an enabler or catalyst at best.

Via Negativa Prompt: Don't treat Base and seasonality like other marketing channels.

The future of MMM AI Agents

The most knowledgeable agent will not necessarily be the one that has read the most MMM content.

It will be one that has accumulated the largest collection of 'Not to Do':

The most knowledgeable agent will not necessarily be the one that has read the most MMM content.

It will be one that has accumulated the largest collection of ‘Not to Do’:

• Things not to conclude / infer / recommend
• Things not to optimize

How experienced MMM consultants operate

In reality, this is how experienced MMM consultants also operate.

They develop a mental catalogue of mistakes they no longer make. We certainly have such a codified catalogue.

The next frontier of MMM AI is not teaching agents more MMM. It is teaching them more MMM restraint.

Check out our website here - https://www.aryma.ai/

Thanks for reading.

For help with MMM, Causal Marketing Experiments and Experimentation, get in touch with us.

We also build some pretty cool AI products to aid Marketing Measurements. Check out our products page to know more -

https://www.aryma.ai/