Large language models, along with the ecosystem of techniques developed around them—from new forms of structured reasoning to increasingly sophisticated multi-agent architectures—are demonstrating unprecedented capabilities that allow us to automate complex tasks that, until recently, required human thought.
Over the past decade, expectations around data science often led to frustration. Investments in talent, infrastructure, and governance were not always enough to translate analytical capabilities into effective decision-making. Recurrent challenges included retaining scarce talent, deploying models into production, translating technical results into business needs, driving adoption within the organization, and coordinating cross-functional teams. In many organizations, transformation has been slow or still underway, delivering partial results far from the aspirational goal of being data-driven.
A new paradigm emerges
In recent years, however, breakthroughs have laid the groundwork for a new model: the agentic automation of analytical tasks. Key milestones include:
- Published studies demonstrate GPT-5’s capacity to contribute to scientific discovery in fields such as mathematics, physics, biology, and astronomy. It is no longer limited to executing instructions: it explores alternatives, interprets evidence, and guides lines of inquiry in ways that resemble the work of expert human teams.
- Solutions like Harmonic’s Aristotle show that LLMs can conduct complex analyses and validate their logic through formal mathematical proofs using languages like Lean 4, which can detect common reasoning errors in these models. At the 2025 International Mathematical Olympiad (IMO), Aristotle performed at gold medal level, delivering verified solutions with reliability on par with advanced models from OpenAI and DeepMind.
- The next wave of models is progressing toward orchestrated multi-agent systems composed of specialized agents, offering enhanced reasoning, transparency, and explainability. OpenAI, for instance, is developing domain-specific agents for future iterations of GPT-6 in areas such as advanced programming, mathematics, scientific research, business consulting, and healthcare.
Multi-agent systems for data science
Specialized multi-agent systems are now emerging for data science. A prime example is described in the paper Kosmos: An AI Scientist for Autonomous Discovery, which has attracted attention for its ability to formulate hypotheses, design experiments, analyze results, iteratively refine models without continuous human oversight, and derive meaningful conclusions.
In a corporate-adapted multi-agent system:
- One agent defines the analytical strategy and formulates hypotheses, drawing from a knowledge base of data science approaches relevant to the business domain.
- Another agent selects the most suitable quantitative methods, develops the necessary transformation logic, performs data analysis, trains predictive models, and builds complete end-to-end pipelines.
- A third agent executes the code, manages errors, and optimizes resource usage. Tools such as ChatGPT’s Advanced Data Analysis already anticipate this direction by seamlessly integrating analytical reasoning with executable code.
- Another agent reviews results, compares methodologies, and identifies patterns. In the near future, formal verification tools such as Lean could be used to validate the logical and deductive structure of reasoning processes.
- A fifth agent synthesizes findings, proposes refinements, and iterates to enhance solution quality.
- Finally, a business-oriented agent converts strategic challenges into targeted analytical questions and evaluates the relevance and impact of the resulting insights. Without this interpretive layer, automation risks drifting away from real business value.
For an agentic data science system to function effectively, robust data governance is essential—allowing agents to reliably identify, locate, and access the knowledge and data required for analysis. Without strong guarantees of quality, traceability, and contextual relevance, even the most advanced agents would struggle to deliver meaningful results. Without strong guarantees of quality, traceability, accessibility, and context, even the most advanced agents would face significant limitations in their ability to generate useful results.
This is not a theoretical proposition. At NTT DATA, we have piloted agentic systems in real-world client environments, aligning them with business objectives and achieving tangible results.
Redefining the role of the data scientist
In this new landscape, the data scientist is redefining their role and gaining strategic relevance. As agents automate much of the technical workload, the differentiating value lies in the human ability to interpret results and discern which findings matter—and what they mean for the business. Expert judgment will play a decisive role in challenging assumptions, anticipating risks, and translating conclusions into solid, strategically aligned decisions.
Organizations that position their specialists as leaders in analytical interpretation and oversight will be better equipped to capitalize on the agentic era. Agentic AI is not about replacing human judgment—but amplifying it.
This new leap is already unfolding. NTT DATA is here to guide you through it. Continue exploring our perspective on the future of analytics and artificial intelligence here.