Thought Leadership Article by Radha Ramaswami Basu, CEO and Founder, iMerit
For years, AI progress followed the familiar pattern: more compute, more data, bigger models, better results. That pattern is now breaking. While model sizes soar, the incremental gains in real-world performance are shrinking. The bottleneck is no more compute; or even sheer data volume. It’s judgment: the ability to embed domain reasoning, context and values into learning systems. The most credible AI research today points to the same shift: progress now depends less on scaling models, and more on injecting human expertise into them.
Why Bigger No Longer Means Better
Between 2018 and 2024, model parameter counts increased by orders of magnitude, yet benchmark gains have flattened. A recent study titled Four data and model quality challenges tied to generative AI from Deloitte points out that despite massive investments, many generative-AI projects still struggle with data architecture, drift and model-integrity issues.
On the same lines, research from MIT A new way to test how well AI systems classify text found that for text-classification models, designing rigorous evaluation pipelines (including human-expert review) is critical to measuring and improving real accuracy. In effect: piling more raw data yields diminishing returns if the data lacks structure, proven provenance or contextual domain insight. Large corpora optimise for statistical consistency and not judgment-rich reasoning.
Quality, Not Quantity, Moves the Needle
Empirical work confirms the shift in emphasis. Curated datasets, annotated by domain specialists and structured for reasoning tasks, consistently outperform much larger noisy corpora. The Deloitte research further notes that one of the four major obstacles for Gen AI adoption is “data integration and model-accuracy controls,” emphasising that data quality has become the bottleneck, not scale. Simply put, fewer but smarter examples beat masses of unrefined ones.
Human Expertise as a System Component
As models enter regulated, high-stakes environments, expert-in‐the‐loop (EITL) workflows are becoming mandatory. In autonomous mobility, safety validation now demands engineers to review and tag real and synthetic scenarios that capture human behavioural variance. In healthcare, clinicians and linguists annotate ambiguous cases, providing interpretive metadata models cannot infer. This is not “human oversight” in the ethical sense; it’s human reasoning as part of system design. Expert‐in‐the‐loop frameworks inject judgment directly into the data supply chain. So, if compute, algorithms and data were the three pillars of AI infrastructure, a fourth is emerging which is human expertise.
A recent Harvard Business Review article, “ How Gen AI Could Change the Value of Expertise”, in, iterates “In some fields, [Gen AI] is automating the foundational tasks … simultaneously, in other fields, AI democratises access to technical expertise … what remains crucial is company-specific knowledge and less codifiable skills.”- HBR makes it clear that in some fields, GenAI takes over the basic, repetitive, or procedural tasks that humans used to do while in others, AI makes advanced knowledge more accessible to everyone , almost like giving people “expert-level tools” instantly. But even with this automation and democratization, the things that AI cannot replicate become more valuable:
- deep understanding of a company’s unique context,
- human judgment,
- nuanced skills that can’t be written down as rules or instructions.
The Economics of Judgment
Deloitte in its report further notes that one of the key engineering obstacles for Gen AI is handling uncontrolled drift and lifecycle data issues. In systems terms, every percentage point of accuracy you earn via better data and expert annotation saves multiple downstream costs; fewer corrections, less retraining, fewer failures. Thus, human judgment becomes not just a nice to have but a measurable efficiency lever. In 2025, McKinsey’s Superagency in the Workplace: Empowering People to Unlock AI’s Full Potential positions AI not as a replacement for humans, but as a capability that is only effective when paired directly with human expertise, contextual decision-making, and domain judgment. Their framework makes a bold claim: AI only creates value when human expertise is embedded into the system design and workflow.
The Judgment Loop
Generative AI revealed this shift. Reinforcement Learning with Human Feedback (RLHF) made humans part of the loop. Now, the next iteration is expert-in-the-loop feedback: humans not only guide but reason, critique and continuously refine model behavior. It is no longer an oversight mechanism; it is embedded architecture. As MIT’s work on classifier evaluation shows, human-expert tooling (evaluation + remediation) is a critical path to higher trust. The increased focus on evals and RL also places in focus the quality of reward design and certainty of answers.
Industry Perspective: The Indian AI Reality
In India, the race to harness AI is intensifying but so is the tension between opportunity and skill-gap. According to a Deloitte–Nasscom report titled Bridging the AI talent gap to boost India’s tech and economic impact published Aug 2024, India’s AI talent pool is projected to more than double by 2027, rising from 600,000–650,000 to over 1.25 million, while the AI market itself is expected to grow at 25–35% annually. Yet this growth hides a looming concern, reports The Times of India on March 25 India to face AI talent gap, shortfall of more than a million workers by 2027 highlighting that the demand is rising far faster than supply, creating a risk of a severe talent shortfall. Around the same time, a World Economic Forum Future of Jobs in India Report 2025 highlights that 63% of Indian workers may require upskilling by 2030 for roles tied to AI and digital technologies, underscoring that India’s demographic dividend could go unrealized without major investments in continuous learning. And in terms of economic impact, EY India projects that generative AI could transform 38 million jobs by 2030, delivering productivity gains of 2.61% in the organized sector and potentially an additional 2.82% in the unorganized sector but they warn the talent shortage (97% of executives cite lack of in-house AI skills) is one of the biggest hurdles. This paints a clear picture that India stands at a crossroads; the potential upside of embedding human judgment into AI systems is massive, but only if the ecosystem can build, nurture, and scale the expert talent base rapidly.
From Data Factories to Cognitive Infrastructure
The AI supply chain is moving from high-throughput data factories to cognitive infrastructure; systems that encode expertise, verify truth, and adapt dynamically. That transition requires new human skillsets: mathematicians who stress-test LLM logic, domain experts who annotate anomalies, engineers who reason about edge-case scenarios. The infrastructure now is as much human as it is computational.
The Bottom Line
Data is everywhere. What’s scarce is judgment; structured, explainable, accountable judgment.The world’s most respected AI research bodies from McKinsey’s global analysis to Stanford’s AI Index converge on one truth: the next frontier of AI performance won’t be unlocked by GPUs alone. It will be unlocked by human judgment. As models grow more capable, the premium shifts to expert reasoning, domain oversight, and ethical interpretation because high-stakes decisions demand more than statistical predictions; they demand cognition. The organisations that learn to engineer judgment into their data pipelines will define the next generation of reliable, domain-specific and responsible AI systems. Not by scaling bigger, but by reasoning better.







