![\"\"](\"https:\/\/www.newsbeep.com\/nz\/wp-content\/uploads\/2026\/03\/image-78-1024x344.png\")
Image by the author.<\/p>\n
A\u00a0National Bureau of Economic Research study<\/a>\u00a0from February 2026 found that 90% of firms reported no measurable productivity impact from AI.\u00a0Less than 30% of CEOs<\/a>\u00a0were satisfied with their GenAI returns. And Gartner placed\u00a0Generative AI squarely in the Trough of Disillusionment<\/a>.<\/p>\n This doesn\u2019t mean AI is useless. It means the bubble is deflating on schedule, the way every technology bubble does. The dot-com bust didn\u2019t kill the internet. It killed the companies that confused hype with product-market fit. The survivors (the ones that sold books and optimized logistics) were the ones obsessed with measurement, experimentation, and unglamorous operational rigor.<\/p>\n The same correction is happening in data science. And the skill set that survives it is the one built on causation, not correlation.<\/p>\n The boat everyone rushed to board is taking on water. The shore they abandoned is looking increasingly solid.<\/p>\n The Anti-Hype Skill Stack<\/p>\n Five skills. Each one is counter-cyclical (becomes more valuable as hype recedes), resistant to LLM automation (requires human judgment that pattern-matching can\u2019t replicate), and directly tied to the business outcomes executives actually pay for.<\/p>\n I didn\u2019t pick these from a textbook. They\u2019re the skills I\u2019ve relied on across four industries (healthcare, retail, higher education, digital advertising) and nearly a decade of applied work. The technical stack barely changed between domains. What changed everything was knowing which of these tools to reach for and when.<\/p>\n 1. Causal Inference: The Skill That Answers \u201cWhy\u201d<\/p>\n What it is<\/p>\n Determining whether X actually causes Y, not just whether they correlate. The toolkit: Randomized Controlled Trials (RCTs), Difference-in-Differences (DiD), interrupted time series, instrumental variables, regression discontinuity, and Directed Acyclic Graphs (DAGs).<\/p>\n Why I believe this is the #1 skill<\/p>\n I once used interrupted time series analysis to isolate the causal impact of a major promotional event on ad revenue forecasts. The predictive model said the event boosted revenue. The causal model told a different story: roughly 40% of that apparent \u201cboost\u201d was cannibalized from surrounding weeks. Customers weren\u2019t spending more; they were shifting when they spent. That single analysis changed how the forecasting team modeled promotional events going forward, improving accuracy by 12% (worth about $2 million annually in a single product vertical).<\/p>\n An LLM can describe instrumental variables. Ask ChatGPT and you\u2019ll get a solid textbook answer. But it can\u2019t\u00a0do\u00a0the reasoning, because causal reasoning requires understanding the data-generating process, intervening on variables, and reasoning about counterfactuals that never appear in any training corpus.<\/p>\n The market signal<\/p>\n A\u00a0Causalens survey<\/a>\u00a0found 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 advertising reported 35% higher ROI than those using correlation-based targeting.<\/p>\n You can predict customer churn with 95% accuracy and still have no idea how to reduce it. Prediction without causation is an expensive way to watch things happen.<\/p>\n 2. Experimental Design: Beyond the Basic A\/B Test<\/p>\n What it is<\/p>\n Designing controlled experiments that isolate the effect of a specific 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\u00a0can\u2019t\u00a0randomize.<\/p>\n Where this gets real<\/p>\n I\u2019ve watched teams deploy machine learning models across multiple retail locations that scored well on holdout sets but failed in production. The reason was always the same: nobody designed the rollout as a proper experiment. No staggered deployment. No matched controls. No pre-registered success metric. The model \u201cworked\u201d on historical data, but without an experimental framework, there was no way to distinguish genuine lift from seasonal noise, selection bias, or regression to the mean.<\/p>\n Running a t-test on two groups is easy. Designing an experiment that accounts for network effects, carryover, and Simpson\u2019s paradox? That takes training most data science programs skip entirely. And it\u2019s the part no AI coding assistant can do for you, because the hard problem isn\u2019t statistical computation. It\u2019s convincing a product team to withhold a feature from a control group long enough to measure the effect.<\/p>\n The market signal<\/p>\n Zalora increased its checkout rate by 12.3%<\/a>\u00a0through a single well-designed experiment on product page copy. PayU gained 5.8% in conversions by testing the removal of one form field. These aren\u2019t ML model improvements. They\u2019re business outcomes from rigorous experimental thinking.<\/p>\n 3. Bayesian Reasoning: Honest Uncertainty<\/p>\n What it is<\/p>\n A framework for updating beliefs as new evidence arrives, quantifying uncertainty, and incorporating prior knowledge into models. In practice: Bayesian A\/B testing, hierarchical models, and probabilistic programming (PyMC, Stan).<\/p>\n Why I learned this out of necessity<\/p>\n When you\u2019re responsible for revenue forecasts that roll up to the CFO, a point estimate is not an answer. \u201cWe expect $X\u201d means nothing without \u201cand here\u2019s the range, and here\u2019s what would make us revise.\u201d I learned Bayesian methods because frequentist confidence intervals weren\u2019t cutting it. A 95% CI that spans a range wider than the entire quarterly target isn\u2019t useful to anyone making a decision. What decision-makers needed was a posterior distribution: \u201cThere\u2019s a 75% probability revenue falls between A and B, and here are the three assumptions that, if violated, shift the distribution.\u201d<\/p>\n Bayesian thinking requires a fundamentally different mental model from the frequentist statistics that dominate most curricula. Probability represents degrees of belief, not long-run frequencies. The learning curve is real. But once you cross it, you stop reporting numbers without uncertainty bands, and you start giving people what they actually need to decide.<\/p>\n The market signal<\/p>\n Bayesian methods<\/a>\u00a0excel 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\u2019re also essential for honest uncertainty quantification, the one thing that point-estimate ML models handle worst.<\/p>\n In a world drowning in AI-generated predictions, the scarcest resource isn\u2019t another forecast. It\u2019s a credible explanation of cause and effect, with an honest confidence interval attached.<\/p>\n 4. Domain Modeling: The Skill You Can\u2019t Bootcamp<\/p>\n What it is<\/p>\n Translating business context into mathematical structure. Understanding the data-generating process (how the data came to exist), identifying the right loss function (what you actually care about optimizing), and knowing which features are causes versus effects.<\/p>\n What four industries taught me<\/p>\n I\u2019ve built models in healthcare (processing millions of patient records daily), retail (forecasting item sales across 15+ locations), higher education (student enrollment pipelines), and digital advertising (econometric models for multi-billion-dollar revenue streams). The Python didn\u2019t change. The SQL didn\u2019t change. What changed\u00a0everything\u00a0was understanding why a hospital\u2019s readmission rate spiked in February (flu season, not a model failure), why a retailer\u2019s 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 rather than an outlier.<\/p>\n AI tools can process data. They can\u2019t understand the context that determines whether a pattern is signal or artifact. That understanding comes from sustained exposure to a specific industry and the ability to think in terms of systems rather than datasets.<\/p>\n The market signal<\/p>\n Domain expertise is why a data scientist in healthcare or finance earns\u00a025-40% more<\/a>\u00a0than a generalist with the same technical skills. The model is rarely the bottleneck. Understanding what the model should optimize is.<\/p>\n 5. Statistical Process Control: Knowing When Something Actually Changed<\/p>\n What it is<\/p>\n Monitoring systems and processes over time to distinguish signal from noise. Control charts, process capability analysis, and root cause investigation. Originally from manufacturing; now applied to ML model monitoring, data pipeline health, and business metric tracking.<\/p>\n A lesson from production ML<\/p>\n I once helped build 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 started drifting and nobody noticed for a month, because there was no process control layer. Once we added control charts tracking detection confidence distributions, inference latency, and feature drift metrics, we could distinguish seasonal shelf rearrangements (noise) from genuine model degradation (signal). The difference: catching a problem in week one versus week five. In inventory management, that gap translates directly to empty shelves and lost revenue.<\/p>\n ML and Statistical Process Control (SPC) are complementary tools<\/a>, not competing ones. Every production ML system needs SPC. Almost none have it, because the skill lives in industrial engineering departments, not data science programs.<\/p>\n The market signal<\/p>\n Manufacturing companies 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 a problem.<\/p>\n Why LLMs Can\u2019t Replace This Stack<\/p>\n The obvious objection: won\u2019t AI eventually learn to do causal reasoning too?<\/p>\n Not anytime soon. The reason is structural.<\/p>\n LLMs are correlation engines. They predict the next token based on statistical patterns in training data. They can\u00a0describe\u00a0causal inference techniques, but they can\u2019t\u00a0do\u00a0causal reasoning, because it requires understanding a data-generating process, intervening on variables, and reasoning about counterfactuals that never appear in any training corpus.<\/p>\n Consider a concrete example. An e-commerce company notices that customers who use their mobile app spend 40% more than desktop users. A predictive model would happily forecast higher revenue if you 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 first explanation is true. No language model can resolve this by pattern-matching over text. It requires designing an experiment, collecting new data, and applying a causal framework.<\/p>\n This is irreducibly human work. And the five skills above are the toolkit for doing it.<\/p>\n The 90-Day Roadmap<\/p>\n Reading about these skills and building them are two different things. Here\u2019s a concrete plan, organized by what you can start this week versus what takes longer to develop. Every recommendation comes from what I\u2019ve personally used or seen produce results.<\/p>\n None of this requires a GPU cluster. None of it requires a subscription to the latest AI platform. A notebook, some data, and the willingness to slow down and think carefully about what you\u2019re measuring and why.<\/p>\n Where This Is Heading<\/p>\n Three shifts are already visible in the market.<\/p>\n The \u201cAI engineer\u201d role will split.\u00a0One track becomes infrastructure (MLOps, deployment, scaling), which is software engineering. The other becomes decision science (causal inference, experimentation, strategic analysis), which is what data science was supposed to be before it got distracted by Kaggle leaderboards.<\/p>\n The premium shifts from prediction to prescription.\u00a0Prediction is commoditizing. AutoML and AI coding assistants can build a decent predictive model in hours. But translating a prediction into a recommendation (\u201craise prices by 3% for this segment, and here\u2019s why we\u2019re 85% confident it increases margin\u201d) requires causal reasoning, domain expertise, and Bayesian uncertainty quantification. That combination is rare.<\/p>\n Trust becomes the differentiator.\u00a0As AI-generated analysis floods every organization, the ability to explain\u00a0why\u00a0a recommendation is credible (here\u2019s the experiment, here\u2019s the confidence interval, here\u2019s what would change our mind) separates analysis that gets acted on from analysis that gets ignored. Statistical rigor becomes the moat.<\/p>\n Prediction is becoming a commodity. The premium is shifting to prescription: \u201cdo X, here\u2019s why, and here\u2019s our confidence level.\u201d<\/p>\n Four hundred billion dollars is chasing a technology whose paying customers can\u2019t explain what they\u2019re getting for their money. The correction will come. It always does.<\/p>\n When it arrives, the people still standing won\u2019t be the ones who learned to prompt a language model. They\u2019ll be the ones who can design an experiment, trace a causal chain, and tell a room full of skeptical executives exactly how confident they should be in a recommendation and exactly what evidence would change their mind.<\/p>\n The bubble is cracking. Underneath it, the ground is solid. Start building on it.<\/p>\n References<\/p>\n IntuitionLabs.\u00a0\u201cAI Bubble vs. Dot-com Bubble: A Data-Driven Comparison.\u201d<\/a>\u00a02025.<\/p>\n Davenport, Thomas H. and Bean, Randy.\u00a0\u201cFive Trends in AI and Data Science for 2026.\u201d<\/a>\u00a0MIT Sloan Management Review, 2026.<\/p>\n Gartner.\u00a0\u201cGenerative AI in Trough of Disillusionment.\u201d<\/a>\u00a0Procurement Magazine, 2025.<\/p>\n Pragmatic Coders.\u00a0\u201cWe Analyzed 4 Years of Gartner\u2019s AI Hype So You Don\u2019t Make a Bad Investment in 2026.\u201d<\/a>\u00a02026.<\/p>\n Acalytica.\u00a0\u201cCausal AI Disruption Across Industries (2025-2026).\u201d<\/a>\u00a02025.<\/p>\n PyMC Labs.\u00a0\u201cFrom Uncertainty to Insight: How Bayesian Data Science Can Transform Your Business.\u201d<\/a>\u00a02024.<\/p>\n Contentsquare.\u00a0\u201c6 Real Examples and Case Studies of A\/B Testing.\u201d<\/a>\u00a02025.<\/p>\n Acerta Analytics.\u00a0\u201cThe Difference Between Machine Learning and SPC, and Why It Matters.\u201d<\/a>\u00a02024.<\/p>\n DASCA.\u00a0\u201cEssential Skills for Data Science Professionals in 2026 and Beyond.\u201d<\/a>\u00a02025.<\/p>\n![\"\"](\"https:\/\/www.newsbeep.com\/nz\/wp-content\/uploads\/2026\/03\/image-79-1024x414.png\")
The hype path has diminishing returns. The bedrock path has a supply shortage and rising demand.
Image by the author.<\/p>\n![\"\"](\"https:\/\/www.newsbeep.com\/nz\/wp-content\/uploads\/2026\/03\/image-80-1024x439.png\")
The stack is hierarchical. Domain knowledge and process monitoring form the base; causal inference sits closest to business outcomes. Image by the author.<\/p>\n![\"\"](\"https:\/\/www.newsbeep.com\/nz\/wp-content\/uploads\/2026\/03\/image-81-1024x677.png\")
Image by the author.<\/p>\n![\"\"](\"https:\/\/www.newsbeep.com\/nz\/wp-content\/uploads\/2026\/03\/image-82-1024x766.png\")
Image by the author.<\/p>\n![\"\"](\"https:\/\/www.newsbeep.com\/nz\/wp-content\/uploads\/2026\/03\/image-83-1024x675.png\")
Image by the author.<\/p>\n