Hyperscalers are burning cash on generative AI infrastructure while the return for an average business stays close to nothing. Somewhere along the way the default advice became “just add an LLM,” and it costs you performance and budget at the same time. The ground everyone walked away from, causal inference in data science, looks steadier by the month.
I have spent more than a decade on broken WooCommerce checkouts and fragmented data pipelines, and I have watched plenty of “innovative” prototypes fall over for want of basic logic. An LLM will write you a product description. It will not tell you whether a promotion caused a revenue lift or whether customers simply moved their spending a week earlier. Get that wrong and your strategy is an expensive guess.
The $300 billion gap in AI strategy
AI infrastructure spending reached nearly $400 billion in 2025 against barely $100 billion of revenue. A 4:1 ratio like that is a warning, and Gartner has already put generative AI in the Trough of Disillusionment. So if you feel left behind by the wave, don’t. The market is repricing the fundamentals, and the skills that come out of the correction intact are the ones built on causation rather than pattern matching.
I made a related point in human strategy in an AI workflow: automation without reasoning is a faster way to make mistakes.
1. Causal inference in data science: the “why” vs. the “what”
Working out whether X actually causes Y is the scarcest skill in tech right now. LLMs are correlation engines. They predict the next token from statistical patterns and cannot reason about counterfactuals, because counterfactuals never show up in a training set. Take the app trap in WooCommerce: app users spend 40% more, and a naive model concludes that everyone should be pushed into the app.
The causal question is a different one. Does the app cause higher spending, or do the people who already spend a lot prefer the app? If it is the second, herding users into the app adds no revenue and annoys your desktop customers. Answering it takes causal inference in data science tools such as directed acyclic graphs (DAGs) and instrumental variables.
<?php
/**
* Naive vs Causal Logic for Discount Impact
* Prefixed: bbioon_
*/
// NAIVE: Assume all revenue during sale is "lift"
function bbioon_naive_lift_report( $sale_revenue, $baseline ) {
return $sale_revenue - $baseline;
}
// CAUSAL: Account for cannibalization (simplified example)
function bbioon_causal_lift_report( $sale_revenue, $baseline, $cannibalization_rate ) {
$real_lift = $sale_revenue - ($baseline * (1 + $cannibalization_rate));
return $real_lift;
}
2. Experimental design beyond A/B tests
A t-test on two groups is easy. Designing an experiment that survives network effects, Simpson’s paradox and selection bias is where most data science programs fall down. I have seen models score beautifully on a holdout set and then crash in production because nobody modeled the data-generating process. That kind of experimental work resists automation for a dull reason: it needs human judgment and organizational buy-in, and no API supplies either.
3. Bayesian reasoning and honest uncertainty
Decision makers need distributions, not point estimates. “We’ll make $1M” tells a CFO nothing. Bayesian methods update beliefs as evidence arrives, so you can say instead: “there is a 75% probability revenue falls between $900k and $1.1M, and here are the three assumptions that would break this forecast.” That approach to causal inference also holds up in small-data situations where frequentist statistics give out.
4. Domain modeling: the human moat
Domain expertise does not come out of a bootcamp. Knowing that a hospital’s readmission rate spikes in February because of flu season and not a broken process, or that a retailer’s demand collapses in week 47 because Black Friday already ate it, is work a person has to do. AI tools process the data without understanding the context around it. That gap is why a specialized data engineer earns considerably more than a prompt engineer.
5. Statistical process control (SPC)
Accuracy drifts in production ML, and without SPC you will not notice for weeks. Catching the drift in week one instead of week five is the difference between empty shelves and a profitable quarter. Control charts do the unglamorous work of separating signal, meaning real model degradation, from noise like a seasonal shift.
If this kind of analysis is eating your dev hours, I can take it on. I have been wrestling with WordPress since the 4.x days.
The shift from prediction to prescription
Prediction is turning into a commodity now that Auto-ML builds a decent predictive model in hours. The premium sits with prescription: do X, here is why, and here is our confidence level. That is where I would spend the learning hours, rather than on the next foundation model release.
