Before NeoCognition raised its $40 million seed round last week, chief executive Yu Su published research on making AI agents reason more reliably under uncertainty. Whether those papers describe the core technology NeoCognition is now selling — and whether they actually work — is the question the company has not answered.
NeoCognition emerged from stealth April 21 backed by Cambium Capital, Walden Catalyst Ventures, Vista Equity Partners, and angels including Intel CEO Lip-Bu Tan and Databricks co-founder Ion Stoica (TechCrunch). The pitch: current AI agents, including those from Anthropic, OpenAI, and Perplexity, complete their intended tasks only about half the time. NeoCognition is building agents that self-learn and rapidly specialize in any domain, mimicking how humans acquire expertise. The company has roughly 15 employees, most with PhDs.
The claim about the 50 percent failure rate came from Su in a press interview. No independent benchmark was cited.
Su is a real researcher. He is an associate professor at Ohio State, co-directs the OSU NLP group, and was named a Sloan Research Fellow in 2025 (Yu Su Homepage). His academic publications cover reasoning under uncertainty, agentic AI, and generalization in language models — work that is directly relevant to the reliability problem NeoCognition says it is solving. The company has not disclosed which specific papers describe the techniques underlying its product, nor has it published benchmarks showing how its approach compares to the baseline it is criticizing.
This matters for a straightforward reason: a startup with academic credentials and a compelling pitch can claim to have cracked a hard problem. But if the underlying research does not match the product narrative, that gap is the story.
Vista Equity's involvement adds a second layer worth examining. Vista publicly challenged the theory that AI will eat software as recently as March 2026, positioning agents instead as tools that enhance the value of the enterprise software its portfolio companies run (Vista Equity Partners). The firm operates an Agentic AI Factory that includes Gainsight, SimplePractice, and Reslinc (CNBC). NeoCognition is the first seed-stage company Vista has backed that is directly addressing the agent reliability problem it has cited as central to its thesis.
That does not mean the thesis is correct. A $40 million seed investment in a 15-person research lab is not proof that agents can learn like humans. It is a bet that the gap NeoCognition is selling against — the 50 percent failure rate — is real, meaningful, and solvable by a team with academic rather than product credentials.
The evidence available to reporters and readers is thin. NeoCognition has no shipped product, no public benchmarks, and no independent validation of its central technical claims. What it has is a credible research founder, a round sized large enough to signal investor conviction, and a framing that resonates with every enterprise software buyer frustrated by AI tools that fail at the worst moments.
Whether the academic papers behind NeoCognition's pitch match the product it is building — or whether they describe a narrower research contribution being marketed at a much larger scale — is the question this story is designed to answer.
The reporting will focus on Su's published research, comparing the specific techniques proposed in his papers against the product description NeoCognition has shared publicly. If the techniques match, the story is about a research team with real momentum and real risk. If they do not match, the story is about the gap between what a Sloan Fellow publishes and what a startup sells.
Either way, the 50 percent failure rate for current agents is real enough that every major AI lab and enterprise software vendor is now competing to close it. NeoCognition's bet is that the academic approach wins. Whether that holds will take time to know. What can be reported now is whether the pitch is grounded in the paper or built on top of it.