I have been working in artificial intelligence since

before most people in this industry started using the term.

At NASA’s neurocomputing institute in the 1990s, I built neural

networks. Real ones — not the marketing version of AI, the

engineering version, where you are accountable to whether the

system actually works.

So when I hear that a competitor has moved to a single

generative AI model for data science, I am not impressed. I am

concerned — for their clients.

Let me be precise, because the term AI is being used to describe

two very different things right now. There is true AI — statistical

models, machine learning, expert systems. And there is

generative AI — large language models that synthesize content

based on patterns in training data.

Generative AI is genuinely powerful for what it does. It is not

what statistical modeling requires.

When I am building a model to identify who is likely to respond

to a specific offer, the model needs to find real signals in real

data. Generative AI will generate something that looks right.

But it does not know that the offer is a seasonal promotion, that

there is a 20% coupon on the cover, or what the macroeconomic

environment looks like this mail season.

Humans know those things. Data scientists with domain

expertise know those things. My job — and what we built this

platform to do — is to give those humans better tools, not

replace them with a system that generates plausible-sounding

outputs.

The companies going all-in on generative AI for modeling are

going to see their results fall. Some already are.

Path2Response was built without the co-op playbook. I came

from processing massive, complex data sets for airlines. I had

never built a data cooperative, and I did not want to replicate how

they had always worked.

We embrace data in its natural, unstructured form. No fixedwidth

fields, no translation loss, no signal trimmed at the edges

to fit a predefined structure. Data arrives intact, the models

see it intact, and intact data processed at high recency is where

performance comes from.

We run on elastic cloud infrastructure — hundreds of servers,

thousands of nodes when a client needs it, scaled back down

when they do not. Same-day and next-day model turnaround

is standard. That speed enables iteration: run a model, review it,

adjust it, run it again. The best results come from that loop, not

from a single output handed back after a long wait.

And we let the model work. When clients and brokers overspecify — loading up pre-selects and forcing the model toward

a predetermined answer — performance degrades. When we take those constraints back, apply the right variables, and let

the model find its own signals, it consistently outperforms what human pre-selection would have produced. The model finds nuances that experienced people miss.

That requires human expertise at the front end to provide

context — and then the discipline to step back.

Those results do not come from a generative AI model. They

come from a platform built with the right philosophy, by people

who understand what data actually is.

Michael Pelikan is Co-Founder and Chief

Technology Officer of Path2Response.

Path2Response’s formal AI usage policy

prohibits generative AI in production modeling

or decisioning. Client data is never used to

train models and is never shared with external

AI systems.

Reach out directly at inquiries@path2response.com to share your ideas and insights.