Bot Or Not? Investors hunt for real AI startups in a sea of hype

Every startup is now an AI startup. Every pitch deck says “AI-native”. Every demo looks like magic. Every graph points up and to the right. Every founder can say they are stitching together frontier models.

Those were the thorny discussion issues behind “How investors are picking startups in the AI era”, an evening in London brought together by YourStory PR and Pathfounders, gathering founders, investors and enterprise leaders to ask how capital, customers and builders can separate genuine AI-native companies from the noise.

The discussion, moderated by Pathfounders Founder and Editor Mike Butcher, brought together Ozan Dagdeviren, founder of AI skills assessment startup Aisa.to; Agata Nowicka, Managing Partner at Visionaries and founder of Female Foundry; and Ekaterina Almasque, Founding Partner of BlankPage Capital.

The central question was that given AI has made it easier than ever to build something, what remains defensible? How are investors approaching this, and how can founders stand out from the crowd?

For Nowicka, the term “AI-native” has already evolved. A couple of years ago, it could simply mean a company using AI in a meaningful way. But that bar is now too low.

The better definition, she argued, is whether AI fundamentally changes the unit economics of the product, the workflow and the organisation. In other words: does the company create value that would not have been possible without AI?

That distinction matters because many startups are still doing old-world software with AI sprinkled on top. An AI note-taker, for example, may make an existing workflow more efficient. But a truly AI-native company starts from the customer’s problem and asks whether the entire process can be reimagined around AI.

Dagdeviren framed it as two layers: the technology layer and the people layer. The AI-native company does not merely bolt AI onto an existing organisation. It structures itself differently. It considers token spend, human judgment, agentic workflows, orchestration layers, and where humans actually need to remain in the loop.

One of the evening’s strongest themes for discussion was that the most interesting founders are not the ones shouting about AI. They are the ones starting with a problem that could not previously be solved, until AI came along. 

Nowicka said the founders she is most excited by are not leading with: “Here is how AI optimises our internal stack.” They are saying: “We are solving a problem that has not been solvable before, and AI is the reason we can now attack it.”

That changes what investors should be looking for. The prize is not a clever wrapper around an AI model, but a new solution space opened up by the new platforms.

Almasque drew a distinction between two very different species of AI company. One is the frontier AI company: often born from years of research, developing novel algorithms, architectures or approaches that may be less energy-hungry, more adaptive or fundamentally different from today’s dominant models. The other is the company that uses other people’s models but applies them early, aggressively and intelligently to a real market.

Both can matter. But investors need to know which animal they are looking at.

Her key due diligence question was simply what have you actually developed yourself, and what have you built on top of someone else’s product?

The panel split the AI market into camps.

At the frontier (for instance, AI for protein folding, drug discovery, materials science, quantum, semiconductors, cybersecurity), deeply technical founders still matter. Perhaps more than ever.

Almasque pointed out that these are not companies where someone can “run some ChatGPT or Claude” and get the answer. These are teams developing new models, new architectures and new ways of solving hard scientific problems.

But Nowicka built on that, noting the equally important founder archetype of the domain expert.

AI may be creating a golden age for people who have spent 15 or 20 years inside a narrow, unglamorous industry. Previously, their market might have looked too niche, too slow or too operationally messy for venture capital. Now, with AI tools, they can build faster, go to market with less capital and use their hard-won context as an advantage.

The founder who knows the industry, owns the customer relationships, and deeply understands the workflow may suddenly be far more investable than a generic young technical team that thinks it can solve the problem purely with a tech platform.

Of course, the best companies may require both. Almasque described a new kind of startup where domain experts and machine-learning experts have to work under one roof. Historically, these groups did not work together well. Each side thought it was superior. In the AI-native era, the two are coming together.

Perhaps the hardest question of the night came during the Q&A: what happens to defensibility when everyone can build almost anything quickly?

The panel argued that defensibility is shifting away from code alone.

Almasque repeated her point that the most durable AI companies will combine domain expertise with proprietary data. If a company can build a unique dataset, understand how to expand it, and use AI to turn it into a flywheel, then it may be able to build a real platform.

Nowicka added another layer: recursiveness. The best AI products should improve with every interaction, every workflow completed and every output generated. The product should not merely serve the customer. It should learn from serving the customer.

Dagdeviren asked: What can you see, know, structure or execute that others cannot?

AI may make it easier to build a prototype, but that has not made fundraising easier, and in many cases, it has raised expectations.

Nowicka argued that at pre-seed, investors now expect much more evidence of success than before. If an MVP can be built over a weekend, then a founder showing up with only a deck is less compelling. Investors increasingly want customer conversations, early revenue, market evidence, fast iteration and proof that the founder can maximise scarce resources.

The old 18-month fundraising cycle is also breaking down. Instead of raising against neat calendar milestones, founders increasingly need to think in terms of value inflexion points. How quickly can the company reach the next moment that proves it deserves more capital?

Even if a founder has a runway, she said, they may have only around 12 months to prove they are taking off. In the AI era, the time available to validate a company has compressed.

For all the talk of models, agents and automation, the panel repeatedly came back to old-fashioned founder qualities.

Almasque said she looks for three things: crazy ambition, credibility and passion.

She recalled investing early in an early quantum computing company, when it had only a handful of people and patents. The founders said they were going to build one of the leading quantum computers in the world and sell machines within a year. It sounded absurd. They did it.

She also described meeting the founders of Graphcore, where a breakthrough chip architecture was sketched out on a napkin. Experts could not say for certain whether it would work because the idea was so novel. At some point, the investor had to decide whether the founder was credible enough to believe.

Passion also mattered. Almasque recalled pitching Siemens’ leadership as a young technologist on Java standardisation and decoupled architectures. The CEO admitted he had not understood the technical jargon, but he understood the conviction.

Dagdeviren added that passion is often coupled with frustration. The builder sees something broken, cannot persuade others that it should be better, and eventually has to go and prove it.

Nowicka added two more traits: curiosity and the ability to learn rapidly.

In other words, AI has not changed the need for founder obsession.

One audience question went pretty meta: how much AI is being used to pick AI startups?

Almasque said she had spent years building a platform to support AI-driven investment decisions. The idea was not to let AI make the decision, but to surface signals. Yet, when tested against past success stories, the model failed to pick hugely successful companies such as UiPath and Graphcore, confirming that while AI can analyse the past, it is not (yet) able to conjure the human instinct for the future.  

Nowicka added that there is also an incentive problem. VC investors are paid to pick winners. If AI could do that better, the value of the intermediary would itself come under pressure.

Her survey work suggested that only around a quarter of investors are currently using AI meaningfully in diligence or startup discovery. So the switch to full AI systems is far from complete.

The final audience question was deliberately provocative: what comes after the AI era?

Dagdeviren argued that AI will expose weaknesses across political systems, democracies, nation-states and the way money moves through the world, and will put pressure on institutions.

The rest of the panel noted that AI is not just another software cycle, but a ‘compression machine’. It compresses the time to prototype and time to market. It also compresses the time investors give to founders to prove themselves.

In doing so, it makes things like judgment, context, courage, taste, domain expertise, and trust stand out. And that’s still pretty human.