ai
Brilliant Machines, Human Breakthroughs.
Why the most powerful AI in history still can’t replace the people asking the right questions?
Every week, a new AI capability lands on your LinkedIn feed. Autonomous research agents. Agentic workflows that spin up, selfcorrect, and deliver results overnight. Models that ace PhDlevel science benchmarks. The pace is dizzying and deliberately so.
So here is a question worth sitting with: if AI is this good, why do R&D pipelines still stall? Why are threequarters of companies still struggling to generate meaningful value from their AI investments, according to the World Economic Forum? Why are 55% of AI pilots in biotech failing not because the models are bad, but because the conditions for good AI outcomes were not there to begin with?
“AI is sprinting. But the problems worth solving require something it has not learned yet: discernment.”
At NineSigma, we sit at the intersection of that tension every day. We work with companies on their hardest innovation challenges the ones where the answer is not in a database, and where the difference between a promising lead and a genuine breakthrough comes down to human curiosity, crossdomain instinct, and the ability to ask a question no one else thought to ask.
This is a map of where AI ends and where the real work begins.
The state of AI in R&D: a genuinely impressive ledger
Let us be honest about what AI has achieved. According to Stanford’s 2026 AI Index, models now meet or exceed human expert performance on benchmarks designed to test PhDlevel science, mathematics, and language understanding. Software engineering benchmark scores jumped from around 60% to nearly 100% in just one year.
In pharmaceutical R&D, AI is reshaping the early pipeline. Protein structure prediction tools are now used by 73% of leading biotech organisations. Predictive docking models have cut timelines from years to months in certain targetidentification workflows.
Biotech AI adoption 2026 Biotech AI Report
Source: 2026 Biotech AI Report
Agentic AI systems that do not just answer questions but pursue goals, orchestrate tasks, and loop back for human validation has moved from research labs into production. IBM, Google, Anthropic, and others are racing to standardise multiagent protocols. The automation of complex enterprise workflows is no longer science fiction; it is a procurement question.
In short: AI is a formidable accelerant. For welldefined problems, with clean data, in established domains, it delivers speed and scale that no human team can match.
Where it breaks down
The failure modes are structural and consistent across industries.
Confident
Wrong answers, at scale
Stanford researchers call it “jagged intelligence” AI that surpasses experts on one task, then fails at something a child could do. In open innovation, a scouting exercise that surfaces wrong signals, confidently and at scale, is worse than no scouting at all.
55%
Pilots that went nowhere
Poor data quality is the number one reason AI pilots fail, cited by 55% of organisations. R&D data is messy, siloed, and shaped by organisational memory that never made it into any system.
Safer ≠
Innovation that converged
Harvard Business School research found that AI recommendation systems risk homogenising creative output. When evaluators rely on AI first, they converge on higheraverage solutions but miss the outliers. In innovation, it is the tails that matter.
75%
Value that never materialised
The WEF found that around threequarters of companies have yet to generate meaningful value from AI not because the models are bad, but because organisational processes were not built to host them.
“The most important innovations often look like noise to a pattern recognition system.”
What this means for innovation
NineSigma’s work is fundamentally about brokering knowledge across boundaries between industries, disciplines, geographies, and organisational contexts that would never otherwise connect. This is precisely the kind of work that AI, for all its power, cannot yet do alone.
Consider what a highquality innovation facilitation process actually requires:
Problem framing
Capturing what the client actually needs, not just what they have articulated. The right question is usually not the one that was asked.
Crossdomain translation
Recognising that a materials science solution might answer a food industry challenge, or that a defence technology might unlock a sustainability application.
Ecosystem sensing
Knowing which experts, research groups, and startups are doing quiet work that has not yet made it into the literature AI is trained on.
Synthesis and judgment
Not just summarising what exists, but making a call on what matters and why. This requires context no dataset fully captures.
Relationship intelligence
Understanding who in a network is the right person to speak to, and how to earn their genuine engagement.
None of this is beyond augmentation by AI. All of it still requires a human at the centre.
The right frame: augmentation, not replacement
The most important shift happening in serious AI conversations right now is the move from replacement thinking to augmentation thinking. IBM’s 2026 tech outlook describes this well: the goal is not to replace human roles, but to decompose them automating the highvolume, repetitive tasks so that the humans in the room can focus entirely on judgmentintensive work.
AI productivity gains where they are (and are not) significant
Source: Stanford AI Index 2026. Gains of this magnitude are not observed in tasks requiring complex contextual judgment.
CapTech’s 2026 trends report puts it plainly: AI systems excel at speed, scale, and automation, but they still lack the nuanced judgment, ethical reasoning, and contextual understanding that humans provide. The failure mode is not AI replacing humans it is humans being removed from the loop too early, replaced not by AI but by overconfidence in AI.
The effective model: humanAI iterative loop
What smart companies are doing differently
The organisations pulling ahead in 2026 are not the ones with the best models. They are the ones who pair model capabilities with human depth. Three patterns stand out.
They invest in problem quality before solution speed. The fastest AI pipeline will not compensate for a poorly framed challenge. Leading organisations spend more time not less on understanding the root question. Domain expertise has become infrastructure: in a world where AI can produce a plausiblesounding analysis of almost any topic, the people who know a field deeply enough to spot when an output is technically correct but practically wrong become more valuable, not less.
They build for humanAI loops, not handoffs. The most effective deployments are not “AI does it, human approves.” They are iterative: human sets direction, AI searches and synthesises, human interrogates the output, AI refines. This loop structure is where the real gains live. And they stay curious about the edges the unexpected disciplines, the emerging researchers, the quiet laboratories working on something that has not been named yet.
Conclusion
AI is not the end of expertise. It is a formidable accelerant for problems that are welldefined, datarich, and bounded. But the evidence from failed pilots, stalled portfolios, and unrealised value is clear: when human judgment is removed too early, AI does not just underperform it can actively mislead.
The companies that will lead in innovation are not the ones who automate the most. They are the ones who stay relentlessly focused on the human insight at the centre of every breakthrough: the sharp question, the unexpected connection, the judgment call that no dataset could have predicted.
“Speed is AI’s gift. Discernment is yours. The organisations that understand this will define what innovation looks like in the next decade.”
That is what open innovation has always been about and why the most important thing a brilliant machine can do is help a brilliant person think better.
Work with us
At NineSigma, we combine the best of human expertise and AI capability to deliver innovation outcomes that neither can achieve alone. If you are facing an innovation challenge that has resisted easy solutions, we would love to talk.
Get in touch →Other topics :
Get in touch
NineSigma Europe BV
Koning Leopold I straat 3B-3000 LeuvenBelgium
+32 16 24 42 80
europe@ninesigma.com
NEWS
