As AI continues to reshape marketing measurement and commercial decision-making, we're seeing an increasing number of approaches to building and deploying models. Some of these involve foundational models trained on large datasets (as in open source world this week), while others use libraries of existing models, shared parameters or candidate models to accelerate the modeling process. There are legitimate technological questions around how these approaches work, but we think there is a much more fundamental question that enterprise customers should be asking: What happens to my data, and to the information my organization generates, when it becomes part of an AI modeling system?
At Mutinex, our position on this is very simple. Our customers' data is their data. Unequivocally. We don't believe that working with a technology provider should mean that your proprietary commercial information, your historical marketing performance or the models built to understand your business should become an asset that can subsequently be leveraged for another customer, especially a direct competitor.
As increasingly sophisticated AI architectures enter the marketing measurement market, we think it is important to be transparent about exactly how Mutinex approaches model development, generalized priors and data pooling.
How we use generalized priors
Like many modern Bayesian modeling systems, Mutinex uses generalized priors as a starting point for certain model specifications. These are deliberately broad distributions, designed to allow models to learn from the individual customer's data rather than impose narrow assumptions about what the answer should be.
We continuously invest in research and development to improve the performance of our modeling systems. Part of this involves testing new model specifications across our customer base to understand whether a particular methodological change improves overall model stability, predictive performance or reliability.
This is an important distinction. We are testing the quality of the modeling methodology, not building a shared understanding of individual customers or industries that can be transferred between clients. Our R&D evaluates whether a general specification performs better across a broad range of modeling environments. It is not an exercise in extracting customer-specific findings or analyzing individual categories in isolation to establish assumptions for other businesses.
That means we can continue to improve our technology while preserving the independence of the models we deliver to individual enterprises.
Why we don't use category-specific priors
Mutinex does not use category-specific priors derived from pooling customer models.
There is a methodological reason for this, alongside the commercial and confidentiality considerations. Markets such as Australia are relatively small. Within many categories, there simply aren't enough genuinely comparable businesses to establish reliable category-level assumptions. Even within the same industry, businesses can differ materially in customer composition, brand maturity, pricing, distribution, media execution and competitive position.
The risk is that a model begins importing assumptions from what other businesses have experienced, rather than independently estimating what is actually happening within the enterprise being analyzed.
We don't think that's a particularly good foundation for enterprise decision-making. Similar businesses can have very different growth dynamics, and the purpose of an independent measurement system is to understand those differences, not obscure them.
More importantly, we do not believe a customer's decision to invest in better measurement should result in their proprietary commercial learnings becoming an input into the measurement systems of their competitors.
Our position on data pooling
We want to be particularly clear about what Mutinex does not do.
- We do not pool customer datasets to build production models for other customers.
- We do not maintain a pooled repository of customer models that can be leveraged to produce models for another enterprise.
- We do not use competitor-specific or category-specific priors derived from aggregating customer modeling results.
- We do not deploy production models that depend on pooled customer data from other enterprises.
These are deliberate decisions about how we believe enterprise modeling technology should operate.
There is an important difference between improving the underlying software and methodology used to build models, and reusing the proprietary information generated by those models. The former is an essential part of building a better technology company. The latter raises significant questions about data ownership, model independence, commercial confidentiality and the obligations a technology provider has to its customers.
We believe those boundaries should be explicit.
Why we believe this matters for enterprise customers
Consider two competing businesses that both invest significant resources in understanding the effectiveness of their marketing, pricing and commercial strategies. Their modeling results can reveal sensitive information about their economics, customer behavior, relative channel effectiveness and potential growth opportunities.
That information has commercial value. It is part of what the enterprise has paid to understand.
We do not believe it is appropriate for a technology provider to take the proprietary knowledge generated through one enterprise's investment and use it to improve the specific modeling outcomes available to a competitor without that enterprise's explicit authorization.
Even where the underlying data is not directly disclosed, the reuse of model parameters, outputs or learned commercial relationships can raise legitimate questions about whether commercially sensitive information is being transferred indirectly.
This is an important distinction as the industry moves toward more sophisticated foundational and pretrained modeling approaches. Faster inference, more automated modeling and stronger AI capabilities are all worthwhile objectives. But they should not come at the expense of enterprise trust or introduce dependencies between competing customers that those customers do not understand.
Innovation and data independence should reinforce each other
We are enormously optimistic about what AI can do for commercial decision-making. We believe modeling will become faster, more accessible, increasingly automated and capable of answering questions that have historically required substantial manual analytical work.
We are investing heavily in that future ourselves.
But our view is that the next generation of enterprise AI should be built on clear principles around data ownership, transparency and independent modeling. Better modeling architectures should create more value from an enterprise's own information, rather than depend on transferring proprietary commercial knowledge between enterprises.
For Mutinex, that means continuing to invest in generalizable modeling methods, rigorously evaluating their performance and ensuring that our production models remain independent of pooled customer data.
Our customers should benefit from improvements in our technology without their commercial information becoming a competitive advantage for somebody else.
That is the standard we believe enterprise technology providers should be held to, and the standard we intend to maintain as modeling technology evolves.