The reasoning edge: Why the future belongs to problem framers, not just problem solvers

In an AI-augmented world where machines can execute complex quantitative methods faster and more reliably than any human, the question that will define careers is no longer “can you run the model?” It is, “do you know why you are running it?” Reflections from Rodreck David, Professor of Data and Information Systems.

Practical experiment

In March 2025, I ran a small experiment. I gave two data science students the same dataset on hospital readmissions and the same prompt: “Predict which patients are likely to be readmitted within 30 days.” Both students were technically capable. Both had access to the same tools. Student A got to work immediately. Within three hours, she had cleaned the data, engineered features, run a logistic regression, tuned hyperparameters, and produced a model that cored 0.87 on the area under the ROC curve (AUC) – a common measure of predictive accuracy where 1.0 is perfect and 0.5 is no better than chance. The model was technically impeccable. She presented it with confidence. Student B paused before touching the data. He asked one question first: “What is the hospital going to do with this prediction?” When he learned that the clinical team wanted to target interventions for preventable readmissions, he reframed the problem entirely. Rather than predicting who would be re-admitted, he modelled which patients were both high-risk and likely to benefit from a specific intervention, a subtly but consequentially different question. His model scored lower on standard metrics. However, it carried a far more useful potential for practice.

The difference between these two students had nothing to do with technical skills. It had everything to do with where their reasoning began. Student A solved the problem she was given. Student B questioned whether it was the right problem to solve. In an AI-augmented world, that distinction is the whole game.

The quantitative comfort zone

When we train students in data science, programming, statistics, or quantitative finance, we teach them a particular kind of reasoning: methodical, quantitative reasoning. This mode of thinking is rule-bound and procedural. It asks: given this type of data and this type of question, which method applies? It is the reasoning behind choosing a chi-square test over a t-test, or regularizing a regression when multicollinearity is suspected. It is rigorous, learnable, and increasingly automatable.

AI systems have become increasingly capable at this layer. Large language models (LLMs) can write many complex types of statistical models, debug pipelines, select appropriate statistical tests, interpret output tables, and flag violations of model assumptions. Tools like GitHub Copilot, Code Interpreter, Claude Code, and specialized data science agents now compress hours of technical work into minutes.

Some scholars have argued that this means AI is itself moving upstream, encroaching on higher-order thinking and not just execution. Others are not so convinced and squarely disagree. This is a serious argument and deserves an honest response.

Is AI really moving upstream?

The evidence that AI is encroaching on abstract reasoning is real but asymmetric. AI can surface analogies across domains, generate creative hypotheses when prompted, and restructure arguments when asked. These are not trivial capabilities. But there is a fundamental difference between performing abstract reasoning when the structure of the problem has already been handed to it, and independently recognizing that the problem needs to be restructured in the first place.

Think of a world-class navigator who can chart any course with precision and speed. Now ask: who decides the destination? AI has become a remarkable navigator. The question of whether to sail, where, and why remains distinctively human.

What AI is genuinely exceptional at, and will continue to improve at, is the full stack of quantitative and computational reasoning: mathematics, statistics, programming, data wrangling, model selection, and optimization. These are domains where the rules are formalized, feedback is measurable, and improvement scales with compute. The upstream movement of AI has been primarily within this quantitative domain, going deeper and faster. The reasoning modes that remain stubbornly, distinctively human operate on a different axis entirely.

Four reasoning modes that matter

Methodical, quantitative reasoning is the foundation. It involves the structured application of formal methods: inferential statistics, algorithmic thinking, computational modelling. This is the reasoning that fills most of a data science curriculum, and it is the reasoning most exposed to AI substitution. Students who believe their career advantage lies here alone are building on sand.

Abstract reasoning is categorically different. It involves operating on concepts rather than instances: seeing what two structurally different situations have in common, or recognizing that a concept from one field illuminates a puzzle in another. A supply chain analyst who realizes that epidemic propagation models from epidemiology describe information cascades in their logistics network is engaged in abstract reasoning. This mode is not about finding the right method. It is about seeing the problem at a level of generality that makes the right method obvious.

Intuitive and counter-intuitive reasoning describes two complementary capacities. Intuitive reasoning is the expert pattern recognition that tells an experienced analyst something feels wrong before they can articulate why, the instinct that a suspiciously clean result in messy data deserves scrutiny, or that a strong correlation is likely spurious before any diagnostic test confirms it. Counter-intuitive reasoning is its productive inversion: the deliberate capacity to question what seems obvious. When everyone in the room agrees that more data will solve the problem, counter-intuitive reasoning asks whether the issue lies with the data itself, or with the objective function, the feature space, or definition of success. Both are skills. Both atrophy without deliberate cultivation.

Problem reformulation is perhaps the most consequential and the least taught. It is the capacity to return to the problem definition itself and ask whether the question being answered is the right one. Student A in my experiment failed not technically but analytically: she produced a correct answer to the wrong question. In professional practice, the most expensive mistakes are rarely wrong answers to right questions. They are right answers to wrong questions. AI will optimize any objective function you give it. It cannot yet determine whether that objective function reflects what actually matters.

Transfer learning as a human competency

In machine learning, transfer learning refers to applying a model trained in one domain to solve a problem in another. The concept has a direct human analogue that is arguably more important. A researcher trained in time-series econometrics who recognizes that survival analysis in clinical trials and customer churn prediction in technology are structurally identical problems is practicing human transfer learning. They are not running a new algorithm. They are relocating their conceptual toolkit across a domain boundary.

This capacity to transfer reasoning frameworks, not just methods, but the deeper structural intuition behind those methods is one of the most durable forms of competitive advantage a quantitatively trained person can develop. It is slow to build, hard to automate, and compounds over time. It becomes more valuable, not less, as AI handles more of the procedural execution.

The upstream shift

The human reasoning edge does not disappear as AI advances. It moves. What changes is where in the problem-solving chain human judgment is most scarce and most valuable. As AI absorbs more of the execution layer, the premium shifts to the pre and post-execution layers: deciding what problem to solve, choosing the level of abstraction at which to engage it, questioning assumptions before and after they get encoded into a pipeline, and recognizing when a technically correct result is answering the wrong question.

This is what it means to say that the future belongs to people who can frame problems, not just solve them. Framing is upstream of solving. It requires abstract reasoning to see the problem in its most useful form, intuitive reasoning to sense when the frame is wrong, counter-intuitive reasoning to interrogate what everyone has taken for granted, and transfer reasoning to borrow conceptual structures from unexpected places.

None of this makes quantitative skill irrelevant. The opposite is true. A person who cannot reason quantitatively cannot evaluate whether a problem has been well-framed, because they cannot model the consequences of different framings. The two modes of reasoning are not in competition. They are in iteratively complementary. Quantitative depth enriches abstract reasoning. Abstract reasoning makes quantitative depth purposeful.

What this means for students today

The students who will be ahead of the curve when they graduate are not necessarily those who know the most tools. They are the ones who have internalized why the tools exist: what conceptual problems they were designed to solve, where they break down, and what assumptions they silently encode. That understanding is the substrate on which all four reasoning modes described here are built.

Student B in my experiment was not more technically gifted than Student A. He had simply learned to think before he computed. In the age of AI, that pause that moment of deliberate problem reformulation before the first line of code, is not a delay. It is the work.