NeoCognition’s $40M bet on agents that learn like humans

NeoCognition’s $40 million seed stands out because checks this large are still unusual at the seed stage, even in a crowded AI market. TechCrunch and TechBuzz both describe it as one of the biggest seed rounds in AI agents this year, which suggests investors are treating the category as a platform shift rather than a narrow experiment ^[1][2]. The money is a vote of confidence in a specific thesis: the next wave of AI value will come from systems that keep learning after they are deployed.

That framing matters because most venture attention in recent years has gone to scaling foundation models, not redesigning how agents adapt. NeoCognition is effectively arguing that intelligence is not just a function of parameter count, but of ongoing experience and specialization ^[1][6]. If that thesis holds, the company is not building another assistant; it is trying to redefine what an agent is supposed to become.

The timing also reflects a broader market mood. Investors have seen large model labs dominate headlines, but they are increasingly looking for the layer that turns those models into reliable operational tools ^[2][4]. A seed round this large says the market believes the real competition is shifting from raw generation to durable learning.

NeoCognition’s founder, Ohio State professor Yu Su, is building from a research base rather than a conventional startup template ^[4][6]. The lab reportedly has about 15 researchers, many with PhDs, which is a dense concentration of academic talent for an early-stage company ^[4]. That structure fits the company’s ambition: it is trying to invent a learning system, not just package an existing model for a new interface.

Su’s core argument is that current agents do not really learn the way humans do. Human workers enter a new role by building a mental model of the environment, then refining it through experience; NeoCognition wants agents to do the same by constructing domain-specific world models ^[4][6]. In that framing, a system becomes an expert not because it was preloaded with every answer, but because it can absorb rules, relationships, and edge cases over time.

That is why the startup keeps using the phrase “any domain.” It is a bold claim, but the technical logic is clear: if a system can generalize the process of specialization, it can move across industries without being rebuilt from scratch ^[2][6]. The company’s bet is that flexibility is the real missing ingredient in enterprise AI.

NeoCognition’s pitch lands because today’s agents still struggle with reliability. Su says current systems from major providers complete tasks only about half the time, which forces users to treat automation as a leap of faith rather than a dependable workflow ^[4]. TechBuzz adds that most agent stacks are still pre-trained models with retrieval-augmented generation layered on top, which can make them look capable without making them genuinely adaptive ^[2].

That limitation is more than a performance annoyance. If an agent cannot remember what worked, adjust to new conditions, or internalize feedback, every deployment becomes a brittle integration project rather than a learning system ^[2][4]. Enterprises then have to keep humans in the loop for supervision, correction, and exception handling, which reduces the economics of automation.

NeoCognition is trying to solve that gap by making learning itself the product. RichlyAI describes the company’s aim as creating agents that build world models through continuous interaction, mirroring how people refine expertise on the job ^[6]. If the company can improve task success rates materially, it would move agents from impressive demos toward systems that can be trusted in production.

The funding syndicate is as revealing as the round size. Cambium Capital and Walden Catalyst Ventures co-led the deal, with participation from Vista Equity Partners and angels including Intel CEO Lip-Bu Tan and Databricks co-founder Ion Stoica ^[4]. That mix combines deep technical credibility with investors who understand enterprise software buying cycles and infrastructure economics.

Vista’s involvement is especially notable because it points to a commercialization path, not just a research milestone. Companies in its orbit tend to want AI that can modernize software-heavy workflows, not just generate better answers ^[4]. NeoCognition’s value proposition therefore looks tailored to enterprise customers that face expensive, domain-specific labor bottlenecks.

The broader market context helps explain the enthusiasm. TechBuzz cites a projected $300 billion enterprise AI market by 2027, and the startup is aiming at the part of that market where specialized expertise is scarce and expensive ^[2]. In that sense, the round is a bet that the next breakout agent company will win by learning operational nuance, not by generating text faster than everyone else.

For enterprises, the upside is straightforward: if agents can truly learn like employees, they could shorten onboarding, reduce repetitive training, and take on domain work that currently requires scarce specialists ^[1][2]. That has obvious appeal in healthcare, legal services, finance, and other fields where expertise takes years to build. But it also raises the stakes for validation, because a system that adapts over time can also drift.

That means governance becomes part of the product, not an afterthought. A self-learning agent will need stronger evaluation, audit trails, and change-control processes than a static copilot, especially in regulated workflows ^[4][6]. Teams will need to know not just whether the system works today, but what it learned, when it learned it, and whether that new behavior is still acceptable.

NeoCognition is also entering a race against the incumbents. TechCrunch notes that the company is building while OpenAI, Google, and others continue to push larger models, but its wager is that adaptability—not just scale—will define the next frontier ^[1]. If that wager is right, the prize is a new class of software that gets better with use; if it is wrong, the funding still marks how strongly the market believes human-like agents are worth chasing.