forecasting roughly $50 billion in promoting revenue using econometrics, time-series models, and causal inference. When a senior VP asked how confident we ought to be in a number, I couldn’t hand them a degree estimate and shrug. I needed to quantify the uncertainty, trace the causal chain, and explain which assumptions would break the forecast in the event that they turned out to be fallacious.
None of that work involved a Large Language Model (LLM). None of it could have.
In case you’re an information scientist who’s been feeling left behind by the AI wave, this text is the reframe. The talents the industry is abandoning are the precise ones becoming scarcer, more demanded, and higher compensated. While everyone else chases the following foundation model, the market is quietly repricing the basics.
This piece lays out five specific skills (I call them the Anti-Hype Stack), explains why each resists automation, and provides you a 90-day roadmap to construct them. But first, a fast take a look at why the hype is cracking.
The $300 Billion Gap
In 2025, hyperscaler corporations committed nearly $400 billion in capital expenditure on AI infrastructure. Actual enterprise AI revenue? Roughly $100 billion. That’s a 4:1 ratio of spending to earning.
A National Bureau of Economic Research study from February 2026 found that 90% of firms reported no measurable productivity impact from AI. Lower than 30% of CEOs were satisfied with their GenAI returns. And Gartner placed Generative AI squarely within the Trough of Disillusionment.
This doesn’t mean AI is useless. It means the bubble is deflating on schedule, the best way every technology bubble does. The dot-com bust didn’t kill the web. It killed the businesses that confused hype with product-market fit. The survivors (those that sold books and optimized logistics) were those obsessive about measurement, experimentation, and unglamorous operational rigor.
The identical correction is going on in data science. And the skill set that survives it’s the one built on causation, not correlation.
The boat everyone rushed to board is taking over water. The shore they abandoned is looking increasingly solid.
The Anti-Hype Skill Stack
Five skills. Every one is counter-cyclical (becomes more worthwhile as hype recedes), proof against LLM automation (requires human judgment that pattern-matching can’t replicate), and directly tied to the business outcomes executives actually pay for.
I didn’t pick these from a textbook. They’re the talents I’ve relied on across 4 industries (healthcare, retail, higher education, digital promoting) and nearly a decade of applied work. The technical stack barely modified between domains. What modified every thing was knowing which of those tools to achieve for and when.
Image by the writer.
1. Causal Inference: The Skill That Answers “Why”
What it’s
Determining whether X actually causes Y, not only whether or not they correlate. The toolkit: Randomized Controlled Trials (RCTs), Difference-in-Differences (DiD), interrupted time series, instrumental variables, regression discontinuity, and Directed Acyclic Graphs (DAGs).
Why I consider that is the #1 skill
I once used interrupted time series evaluation to isolate the causal impact of a serious promotional event on ad revenue forecasts. The predictive model said the event boosted revenue. The causal model told a distinct story: roughly 40% of that apparent “boost” was cannibalized from surrounding weeks. Customers weren’t spending more; they were shifting once they spent. That single evaluation modified how the forecasting team modeled promotional events going forward, improving accuracy by 12% (price about $2 million annually in a single product vertical).
An LLM can describe instrumental variables. Ask ChatGPT and also you’ll get a solid textbook answer. But it could’t the reasoning, because causal reasoning requires understanding the data-generating process, intervening on variables, and reasoning about counterfactuals that never appear in any training corpus.
The market signal
A Causalens survey found Causal AI was the #1 technique AI leaders planned to adopt, with nearly 70% of AI-driven organizations implementing causal reasoning by 2026. Organizations applying causal methods to promoting reported 35% higher ROI than those using correlation-based targeting.
You may predict customer churn with 95% accuracy and still don’t know easy methods to reduce it. Prediction without causation is an expensive strategy to watch things occur.
2. Experimental Design: Beyond the Basic A/B Test
What it’s
Designing controlled experiments that isolate the effect of a selected intervention. This goes well beyond splitting traffic 50/50. It includes multi-armed bandits, factorial designs, sequential testing, and (critically) quasi-experimental methods for situations where you randomize.
Where this gets real
I’ve watched teams deploy machine learning models across multiple retail locations that scored well on holdout sets but failed in production. The rationale was all the time the identical: no person designed the rollout as a correct experiment. No staggered deployment. No matched controls. No pre-registered success metric. The model “worked” on historical data, but without an experimental framework, there was no strategy to distinguish real lift from seasonal noise, selection bias, or regression to the mean.
Running a t-test on two groups is straightforward. Designing an experiment that accounts for network effects, carryover, and Simpson’s paradox? That takes training most data science programs skip entirely. And it’s the part no AI coding assistant can do for you, since the hard problem isn’t statistical computation. It’s convincing a product team to withhold a feature from a control group long enough to measure the effect.
The market signal
Zalora increased its checkout rate by 12.3% through a single well-designed experiment on product page copy. PayU gained 5.8% in conversions by testing the removal of 1 form field. These aren’t ML model improvements. They’re business outcomes from rigorous experimental considering.
3. Bayesian Reasoning: Honest Uncertainty
What it’s
A framework for updating beliefs as recent evidence arrives, quantifying uncertainty, and incorporating prior knowledge into models. In practice: Bayesian A/B testing, hierarchical models, and probabilistic programming (PyMC, Stan).
Why I learned this out of necessity
Once you’re answerable for revenue forecasts that roll as much as the CFO, a degree estimate is just not a solution. “We expect $X” means nothing without “and here’s the range, and here’s what would make us revise.” I learned Bayesian methods because frequentist confidence intervals weren’t cutting it. A 95% CI that spans a variety wider than the whole quarterly goal isn’t useful to anyone making a choice. What decision-makers needed was a posterior distribution: “There’s a 75% probability revenue falls between A and B, and listed below are the three assumptions that, if violated, shift the distribution.”
Bayesian considering requires a fundamentally different mental model from the frequentist statistics that dominate most curricula. Probability represents degrees of belief, not long-run frequencies. The educational curve is real. But when you cross it, you stop reporting numbers without uncertainty bands, and also you start giving people what they really need to make a decision.
The market signal
Bayesian methods excel in small-data environments where classical approaches break down: clinical trials with limited participants, early-stage product experiments, and risk modeling with sparse history. They’re also essential for honest uncertainty quantification, the one thing that point-estimate ML models handle worst.
In a world drowning in AI-generated predictions, the scarcest resource isn’t one other forecast. It’s a reputable explanation of cause and effect, with an honest confidence interval attached.
4. Domain Modeling: The Skill You Can’t Bootcamp
What it’s
Translating business context into mathematical structure. Understanding the data-generating process (how the info got here to exist), identifying the appropriate loss function (what you really care about optimizing), and knowing which features are causes versus effects.
What 4 industries taught me
I’ve built models in healthcare (processing tens of millions of patient records each day), retail (forecasting item sales across 15+ locations), higher education (student enrollment pipelines), and digital promoting (econometric models for multi-billion-dollar revenue streams). The Python didn’t change. The SQL didn’t change. What modified was understanding why a hospital’s readmission rate spiked in February (flu season, not a model failure), why a retailer’s demand forecast collapsed in week 47 (Black Friday cannibalization, not a distribution shift), and why an ad revenue forecast needed to treat a tentpole event as a structural break somewhat than an outlier.
AI tools can process data. They will’t understand the context that determines whether a pattern is signal or artifact. That understanding comes from sustained exposure to a selected industry and the flexibility to think when it comes to systems somewhat than datasets.
The market signal
Domain expertise is why an information scientist in healthcare or finance earns 25-40% more than a generalist with the identical technical skills. The model isn’t the bottleneck. Understanding what the model should optimize is.
5. Statistical Process Control: Knowing When Something Actually Modified
What it’s
Monitoring systems and processes over time to differentiate signal from noise. Control charts, process capability evaluation, and root cause investigation. Originally from manufacturing; now applied to ML model monitoring, data pipeline health, and business metric tracking.
A lesson from production ML
I once helped construct an object detection pipeline for automated retail inventory monitoring. The model hit 95% mAP on the test set. It went to production. Three weeks later, accuracy began drifting and no person noticed for a month, because there was no process control layer. Once we added control charts tracking detection confidence distributions, inference latency, and have drift metrics, we could distinguish seasonal shelf rearrangements (noise) from real model degradation (signal). The difference: catching an issue in week one versus week five. In inventory management, that gap translates on to empty shelves and lost revenue.
ML and Statistical Process Control (SPC) are complementary tools, not competing ones. Every production ML system needs SPC. Almost none have it, since the skill lives in industrial engineering departments, not data science programs.
The market signal
Manufacturing corporations using SPC alongside ML achieve measurably lower defect rates by catching process anomalies before they cascade. In tech, SPC-based monitoring catches model degradation weeks before accuracy metrics flag an issue.
Why LLMs Can’t Replace This Stack
The apparent objection: won’t AI eventually learn to do causal reasoning too?
Not anytime soon. The rationale is structural.
LLMs are correlation engines. They predict the following token based on statistical patterns in training data. They will causal inference techniques, but they will’t causal reasoning, since it requires understanding a data-generating process, intervening on variables, and reasoning about counterfactuals that never appear in any training corpus.
Consider a concrete example. An e-commerce company notices that customers who use their mobile app spend 40% greater than desktop users. A predictive model would happily forecast higher revenue if you happen to push more people to download the app. A causal thinker would stop and ask: does the app cause higher spending, or do high-spending customers just prefer apps? The intervention (pushing downloads) only works if the primary explanation is true. No language model can resolve this by pattern-matching over text. It requires designing an experiment, collecting recent data, and applying a causal framework.
That is irreducibly human work. And the five skills above are the toolkit for doing it.
The 90-Day Roadmap
Reading about these skills and constructing them are two various things. Here’s a concrete plan, organized by what you possibly can start this week versus what takes longer to develop. Every advice comes from what I’ve personally used or seen produce results.
None of this requires a GPU cluster. None of it requires a subscription to the most recent AI platform. A notebook, some data, and the willingness to decelerate and think twice about what you’re measuring and why.
Where This Is Heading
Three shifts are already visible out there.
The “AI engineer” role will split. One track becomes infrastructure (MLOps, deployment, scaling), which is software engineering. The opposite becomes decision science (causal inference, experimentation, strategic evaluation), which is what data science was purported to be before it got distracted by Kaggle leaderboards.
The premium shifts from prediction to prescription. Prediction is commoditizing. AutoML and AI coding assistants can construct a good predictive model in hours. But translating a prediction right into a advice (“raise prices by 3% for this segment, and here’s why we’re 85% confident it increases margin”) requires causal reasoning, domain expertise, and Bayesian uncertainty quantification. That combination is rare.
Trust becomes the differentiator. As AI-generated evaluation floods every organization, the flexibility to elucidate a advice is credible (here’s the experiment, here’s the arrogance interval, here’s what would change our mind) separates evaluation that gets acted on from evaluation that gets ignored. Statistical rigor becomes the moat.
Prediction is becoming a commodity. The premium is shifting to prescription: “do X, here’s why, and here’s our confidence level.”
4 hundred billion dollars is chasing a technology whose paying customers can’t explain what they’re getting for his or her money. The correction will come. It all the time does.
When it arrives, the people still standing won’t be those who learned to prompt a language model. They’ll be those who can design an experiment, trace a causal chain, and tell a room filled with skeptical executives exactly how confident they ought to be in a advice and exactly what evidence would change their mind.
The bubble is cracking. Underneath it, the bottom is solid. Start constructing on it.
References
- IntuitionLabs. “AI Bubble vs. Dot-com Bubble: A Data-Driven Comparison.” 2025.
- Davenport, Thomas H. and Bean, Randy. “Five Trends in AI and Data Science for 2026.” MIT Sloan Management Review, 2026.
- Gartner. “Generative AI in Trough of Disillusionment.” Procurement Magazine, 2025.
- Pragmatic Coders. “We Analyzed 4 Years of Gartner’s AI Hype So You Don’t Make a Bad Investment in 2026.” 2026.
- Acalytica. “Causal AI Disruption Across Industries (2025-2026).” 2025.
- PyMC Labs. “From Uncertainty to Insight: How Bayesian Data Science Can Transform Your Business.” 2024.
- Contentsquare. “6 Real Examples and Case Studies of A/B Testing.” 2025.
- Acerta Analytics. “The Difference Between Machine Learning and SPC, and Why It Matters.” 2024.
- DASCA. “Essential Skills for Data Science Professionals in 2026 and Beyond.” 2025.
- Wikipedia. “AI Bubble.” (Accessed February 2026).
