By Jacques Bughin
How combining intelligence, experimentation, and learning could transform the economics of scientific discovery
The management question AI may do more than speed up individual research tasks. By connecting search, prediction, experimentation, and learning, organizations could make each discovery cycle more productive than the last. The opportunity is substantial; whether it yields persistent gains remains an empirical question.
The innovation paradox: more research, harder discoveries
Innovation drives economic progress. New materials, medicines, industrial processes, and technologies create opportunities for productivity growth, new products, and entirely new industries. Yet a difficult question has troubled economists and business leaders for decades: why does sustaining technological progress appear to require ever more research effort?
Uncertainty can fall while the underlying technology remains unchanged.
Bloom, Jones, Van Reenen, and Webb (2020) documented declining measured research productivity across multiple technological domains. Their semiconductor example was striking: maintaining Moore’s Law required more than 18 times as many researchers as in the early 1970s. The finding is not that scientific progress has stopped. Rather, in these domains, sustaining progress appeared to require increasing research effort.
The broad interpretation remains debated. Results depend on how researchers define ideas, progress, and research inputs, and the pattern need not apply uniformly across fields (OECD, 2023). Still, R&D executives recognize the operating challenge: knowledge is more complex, experimentation is costly, and research outcomes are uncertain. Together, these conditions make it difficult to sustain improvements in scientific output per unit of investment.
Artificial intelligence could alter this equation—not only by helping researchers work faster or by generating more hypotheses, but by changing how many possibilities organizations explore, how they choose among them, how cheaply they test them, and what they learn from each result. Each capability can matter on its own. Their combination may change the economics of discovery.
Consider a pharmaceutical team that can generate one hundred times more candidate molecules. If testing capacity is fixed, most candidates cannot be evaluated. A prediction model can help prioritize them, but its advantage may remain limited if experiments take months. Now connect generation and prediction to automated experimentation and a learning system that updates candidate selection using verified results. Broader search creates options; better selection raises experimental yield; experiments produce evidence; and evidence improves the next search.
The system then does more than accelerate research. It can change the relationship between research effort and validated discovery. For leaders, the question shifts from “Which AI tools should our scientists adopt?” to “How should we redesign discovery so that one improvement amplifies another?”
1. From research productivity to discovery productivity
R&D is often described as a sequence: define a problem, generate hypotheses, test possible solutions, evaluate results, and build knowledge. This is useful for managing projects, but it can obscure the economics. Every research organization faces five connected constraints:
- Search: researchers cannot explore every possible idea or technological combination.
- Information: they cannot know in advance which ideas will work.
- Performance: promising ideas may still fail to meet scientific or technical requirements.
- Experimentation: testing consumes time, money, equipment, and skilled labor.
- Learning: evidence from one experiment may not be retained or reused in later work.
Historically, organizations addressed these constraints through specialization, accumulated expertise, laboratories, partnerships, and research routines. AI can intervene across the same system: it can search knowledge bases, propose directions, estimate outcomes, design tests, interpret results, and help researchers adapt subsequent experiments. The distinctive possibility is to connect these functions rather than optimize each in isolation.
A useful management measure is quality-adjusted discovery productivity:
Quality-adjusted validated discoveries ÷ total effective R&D resources
Quality adjustment matters: ten trivial improvements are not necessarily more valuable than one major discovery. The denominator should include AI, computation, equipment, scientific expertise, and integration costs—not only laboratory headcount. The test is whether an organization produces more valuable, verified scientific knowledge per unit of effective research input.
2. Five ways AI can change the discovery equation
Greater discovery throughput
AI can expand the number of hypotheses, designs, and technological combinations researchers can generate or examine. Wider search raises the opportunity set, but it does not guarantee that more candidates will be useful or testable.
Reduced decision uncertainty
AI can estimate which candidates are worth testing, identify likely failure modes, and update beliefs as evidence accumulates. Its contribution is not the elimination of scientific uncertainty; it is better decision-making under uncertainty.
Better technical performance
AI can help optimize designs, improve experimental precision, and raise the probability that a candidate meets a defined requirement. This improves the yield or quality of discovery rather than merely its volume.
Faster and cheaper experimentation
AI-assisted experimental design, simulation, and laboratory automation can reduce the cost or duration of testing. This makes more candidates economically testable and returns feedback sooner.
Cumulative learning
AI systems can combine verified results with prior experiments to improve later search and selection. An experiment can therefore create value beyond its immediate outcome by making future research more productive.
These mechanisms should not be conflated. Throughput can rise without success probability improving. Uncertainty can fall while the underlying technology remains unchanged. Technical performance can improve even when experiments remain expensive. Nor must every mechanism depend on an autonomous agent: predictive models can support scientific judgment without executing experiments. Agentic systems matter when they reliably connect search, choice, action, evaluation, and adaptation.
3. Why combining capabilities could produce larger gains
Suppose a research group evaluates 100 candidates a year, with a 10% chance that each candidate meets its scientific criteria. Expected output is 10 validated discoveries. Now consider an illustrative system that evaluates 50% more candidates, raises the success rate among those evaluated from 10% to 15%, and increases the average scientific value of a successful result by 20%.
Exhibit 1: How complementary improvements compound in an illustrative scenario.
Starting output: 100 × 10% × 1 = 10 units. After AI-enabled changes: 150 × 15% × 1.2 = 27 units, a 170% increase. The values are assumptions for illustration, not estimates of AI’s causal effect. Costs are held constant.
The result is 27 quality-adjusted units, 170% above the baseline. This arithmetic shows how changes in multiple production factors can compound. It does not, by itself, demonstrate technological complementarity: independent improvements in multiplicative factors also compound.
True complementarity means that one capability’s contribution becomes larger when another is present. A prediction model may add little without experimental feedback, yet become much more useful when an automated laboratory supplies informative results. High-throughput generation may create little value alone, yet become productive when paired with reliable prioritization. The return on one AI investment may therefore depend on the capabilities around it.
This changes the investment problem. A tool with modest standalone returns may be essential to a system that generates larger gains. Conversely, a firm may invest heavily in idea generation only to discover that testing capacity is the binding constraint.
4. From isolated tools to an integrated discovery loop
Many organizations are introducing AI through separate applications: one team summarizes literature, another predicts candidate properties, another automates experiments, and a fourth analyzes results. Each tool can improve local productivity, but opportunities for interaction remain limited when data, workflows, and decisions are disconnected.
A more ambitious design links those activities in a repeated cycle. The loop does not eliminate human scientific judgment. Researchers set objectives, define acceptable evidence, investigate unexpected results, and supervise consequential decisions. Its purpose is to ensure that verified outcomes inform the next round of work.
Exhibit 2: A connected AI-enabled discovery system.
The cycle depends on valid measurement and reusable records. Human oversight sets goals, scrutinizes evidence, and redirects research. Dashed arrows represent continuing scientific supervision.
In a fragmented organization, an unsuccessful experiment may simply be recorded as a failed project. In an integrated system, the same result can update a model, eliminate part of the search space, and guide the next experiment. An experiment can create value even when it does not produce a successful candidate.
This is the logic of active learning and sequential experimentation. The most useful next experiment is not always the one most likely to succeed; it may be the one most likely to resolve uncertainty. Ranković, Griffiths, and Schwaller’s (2026) GOLLuM study illustrates the potential: across 23 experimental optimization tasks, their uncertainty-calibrated method matched traditional Bayesian optimization with more than 40% fewer experiments. The result is evidence about sample-efficient optimization in benchmark tasks, not proof of a general productivity effect across R&D.
5. What the evidence does—and does not—show
Evidence is emerging in specific scientific settings. The strongest verified example in the source material is a study of uncertainty-calibrated optimization that reduces the experiments needed to reach comparable performance on a set of tasks (Ranković et al., 2026). Reviews of self-driving laboratories describe growing integration of AI, robotics, and automated experimentation, while emphasizing challenges in scalability, generalizability, data quality, and complete experimental records (Canty & Abolhasani, 2026).
A widely circulated preprint by Toner-Rodgers (2024) reported large gains in materials discoveries, patents, and downstream product innovation after AI was introduced to researchers. Those findings should not be treated as evidence: in May 2025, MIT publicly stated that it had no confidence in the provenance, reliability, or validity of the data or in the veracity of the research, and said the findings should not be relied on in academic or public discussion (MIT Department of Economics, 2025). This episode underscores why high-profile claims require independent verification.
The available evidence therefore supports a measured conclusion. AI can improve particular research tasks and experimental choices. The self-driving-laboratory literature shows that integration is technically advancing. But no cited study establishes that combining all five capabilities produces a universal nonlinear gain in scientific productivity. That is the central proposition for future research, not a settled empirical result.
6. A compact model of discovery and learning
A simple framework clarifies the economics. Let quality-adjusted discovery output in period t be:
Dₜ = qₜ × pₜ × min(nₜ, Kₜ)
Here, nₜ is the number of candidate ideas available for evaluation, Kₜ is effective experimental and verification capacity, pₜ is the probability of a validated success among candidates evaluated, and qₜ is the average scientific value per success.
The expression yields three immediate implications. First, more search raises output only while the organization can process the additional candidates; once verification capacity is saturated, extra generation has little immediate value. Second, better selection or technical performance raises the yield of candidates actually tested. Third, lowering the cost or duration of experiments expands effective capacity, so more AI-generated opportunities can be validated.
Uncertainty reduction affects the composition of the portfolio: better information improves which candidates and experiments are chosen. It need not raise the success probability of every candidate. Learning makes the process dynamic. Let Hₜ represent reusable scientific knowledge accumulated through verified experimentation:
Hₜ₊₁ = (1 − δ)Hₜ + η Iₜ Eₜ
Eₜ is completed experimentation, Iₜ is its average information content, η is the organization’s ability to convert feedback into reusable knowledge, and δ captures depreciation or obsolescence. More experiments improve future productivity only when they produce informative results and the organization can absorb them.
Let a denote AI-supported search and b denote AI-supported experimentation and learning. Genuine complementarity requires the cross-effect to be positive:
∂²D / ∂a ∂b > 0
In plain language, the discovery benefit of better search rises as experimentation and learning improve. This is not guaranteed. When verification capacity is binding, search has low marginal value. When feedback is unreliable, more experiments may not improve future predictions. When a system favors familiar solutions, wider search may still miss genuinely novel possibilities. Organizations must engineer the conditions for complementarity by connecting capabilities, relieving bottlenecks, and preserving trustworthy feedback.
7. Which organizations can combine the capabilities?
The ability to integrate AI into discovery will vary across organizations. AI-native scientific innovators can design processes around machine-supported search and experimentation, but may lack large laboratories, deep domain expertise, or the scale to generate reliable feedback. Established R&D-intensive firms possess domain knowledge, proprietary data, laboratories, and experienced scientists; their challenge is to connect these assets across organizational boundaries. Research platforms and ecosystems can combine models, universities, laboratories, equipment providers, and specialist organizations, but face interoperability and knowledge-transfer problems.
A practical maturity ladder distinguishes four configurations:
- Assisted research: AI improves isolated scientific tasks, producing local efficiency gains.
- AI-augmented discovery: search, prediction, and design improve opportunity quality and research output.
- Integrated discovery: search, selection, testing, and evaluation are connected, creating the possibility of complementarities.
- Adaptive discovery: verified feedback continuously informs future search and experimentation, potentially supporting persistent productivity gains.
The most advanced configuration is not always economic. Limited data, high physical costs, or hard-to-automate experiments may favor a narrower design. The right level of integration depends on scientific uncertainty, verification costs, and the extent to which results can be reused.
8. Could AI reverse declining research productivity?
The answer depends on whether AI produces a one-time increase in output or a persistent improvement in valuable discoveries per unit of R&D effort. If research productivity would otherwise decline by 4% annually, a one-time 30% gain raises the level but does not change the downward trend. Reversing a persistent decline requires recurring gains large enough to offset the forces that make discovery harder.
Learning creates one route to such gains: better experiment selection can build a more useful knowledge base; that knowledge can improve later predictions; and improved predictions can raise the yield of future experiments. Yet learning can face diminishing returns. Remaining questions may be harder, information may become less novel, and physical research may grow more expensive. Productivity depends on the balance between cumulative learning and rising scientific complexity.
Exhibit 3: Three hypothetical research-productivity paths.
Illustrative index, year 0 = 100: baseline falls 4% per year; isolated AI gives a one-time 30% level increase followed by the same decline; integrated learning grows 6% annually. These assumptions are scenarios, not estimates or forecasts.
The integrated-learning path is the most consequential possibility and the least established empirically. Testing it requires longitudinal evidence on AI-enabled discovery systems, not only short-run comparisons before and after a single tool is introduced.
9. The managerial imperative: redesign the discovery system
The conventional investment case begins with applications: which tasks can AI automate, and how much time or money will each save? A discovery-system perspective begins elsewhere: what constrains the organization from producing more high-quality, verified discoveries per unit of research investment?
- Measure quality-adjusted discoveries, not just model accuracy, generated hypotheses, or experiments completed. Track validation, scientific value, downstream use, and the full cost of computation, equipment, labor, and integration.
- Treat AI investment as a portfolio of connected capabilities. Some tools may have modest standalone returns yet be essential to the performance of the complete system.
- Invest in reliable, informative experimental feedback. Without credible verification, a learning loop can reinforce errors rather than improve discovery.
- Protect scientific exploration. A system that repeatedly optimizes familiar solutions can raise short-term productivity while narrowing the range of ideas considered.
A productive R&D system should discover more of what works and improve its ability to find and validate what is genuinely new.
10. The boundary of the promise: more AI is not always better discovery
It is premature to conclude that AI will inevitably reverse declining research productivity. First, candidate solutions require evidence. Plausibility is not discovery. Second, physical experiments depend on biological processes, measurement time, and specialized equipment that may not be accelerated indefinitely. Third, scientific diversity matters: a system that favors historically successful strategies can perform well in the short term while underexploring unconventional approaches. Finally, AI creates persistent gains only when teams have the expertise, incentives, infrastructure, and governance to evaluate proposals and retain useful knowledge.
These conditions determine whether AI capabilities reinforce one another or undermine one another. The practical challenge is to identify when complementarity is possible, what blocks it, and what it costs to relieve the constraint.
Conclusion: connect intelligence, experimentation, and learning
The finding that research productivity has declined in several important domains poses a challenge for innovation-led growth: maintaining technological progress can require increasing scientific effort. AI creates an opportunity to change that pattern, but the opportunity is not simply faster work or a larger volume of ideas.
AI creates an opportunity to change that pattern, but the opportunity is not simply faster work or a larger volume of ideas.
AI can help organizations search more widely, choose experiments more intelligently, improve technical performance, lower validation costs, and accumulate knowledge through feedback. Each capability can improve R&D productivity. Their greatest potential lies in integration: prediction improves experimental choice; experimentation produces evidence; and evidence informs the next search.
This does not guarantee a universal reversal in declining research productivity. Scientific constraints remain, and positive feedback does not eliminate diminishing returns. But it changes the question innovation leaders should ask. Rather than focusing only on whether AI makes research faster, they should ask whether it changes the relationship between research effort and valuable discoveries.
AI’s most consequential contribution to innovation may be its ability to make each act of discovery improve the productivity of the next.
Jacques Bughin
