What VCs Actually Want in AI Startups Right Now: A Panel Breakdown

Video: https://www.youtube.com/watch?v=_zUnVik1xvE

This panel brings together venture capitalists discussing where AI investing stands today – what is working, what is not, what kinds of founders get funded, and how they think about moats, pricing, and the future. Here is every major point, explained.

1. Everything Is AI Now

Every new startup these investors fund is either an AI company or AI-adjacent. Even data infrastructure plays get funded because they benefit from the explosion of AI applications. The era of “is this an AI company?” as a distinguishing question is over. The question is now “how deeply does this company use AI?”

The foundation model layer – OpenAI, Anthropic, Google, Meta – is funded and running. Those bets are placed. New investment focuses on companies that deeply understand a specific vertical and apply AI in technically differentiated ways. The bar has moved past wrappers. If your product is a thin interface over someone else’s API, you are not fundable.

The speed of change is the defining challenge. The trajectory from Jasper (AI copywriting wrapper) to foundation models eating that use case, to vertical apps, to model vendors themselves integrating up the stack, happened in under two years. That speed makes it hard to underwrite durable differentiation. Investors compensate by overweighting founder quality and resilience – betting on the person more than the product, because the product will have to change.

2. What They Look For in AI Founders

Four traits came up repeatedly:

Extreme intellectual honesty about product-market fit. Enterprise AI is drowning in false positives. Companies run pilots, show usage dashboards, and claim traction. But pilot revenue is not real revenue. Usage metrics without retention are vanity. The founders who get funded are the ones who can look at their own numbers and say “this is not real PMF yet” – and explain what real PMF would look like and how they plan to get there. Seeing through your own vanity metrics is the skill.

Ability to attract top AI talent. The talent war in AI is brutal. The best researchers and engineers have their pick of OpenAI, Anthropic, Google DeepMind, or a well-funded startup. Founders who can recruit exceptional technical talent signal something about their vision, their culture, and their ability to build. If you cannot hire, you cannot build.

High resilience and willingness to pivot. The “OpenAI Dev Day kills a thousand startups” joke is only half a joke. Every major model release reshuffles what is possible and what is commoditized. Founders who survive are the ones who treat model capability shifts as opportunities rather than extinction events. The product you build today may need to be fundamentally different in six months. That is not a bug – it is the environment.

Deep domain insight and a unique data strategy. Generic AI applications are commodity. What is not commodity is a founder who spent a decade in insurance underwriting and knows exactly which workflows are broken, combined with a strategy to capture new interaction data that current models have never seen. The data you generate through your product becomes your moat – but only if you knew which data to capture in the first place.

3. Where AI Is Working Now

The panel identified clear markets where AI is delivering real value today:

Chat interfaces as the new front door. ChatGPT, Claude, and chat-centric applications have become the primary way many knowledge workers interact with systems. The chat interface is not a feature – it is a platform shift. The way the browser became the front door to the internet, chat is becoming the front door to enterprise systems.

Coding tools. Cursor, Cognition, and the broader AI-assisted development ecosystem are working. Code generation is real. But the panel noted a second-order effect: AI-generated code is breaking other enterprise systems – CI pipelines, QA processes, security scanners. The code works, but the systems around the code were not designed for the volume and velocity of AI-generated changes. This creates its own investment category.

Customer support and interviews. These are early categories where AI is doing real work – resolving tickets, conducting screening interviews – rather than just enabling humans to work faster. The distinction matters: “AI helps a human do the task” is incremental. “AI does the task” is transformational.

Where the big upside remains:

Deep vertical AI. Full-stack solutions that own an entire workflow in a specific industry – insurance, healthcare, manufacturing. The key insight: the best vertical AI does not speed up each step in a process (A to B to C to D to E). It skips steps entirely (A to E). When AI can eliminate intermediate steps that only existed because humans needed them, the value creation is enormous.

AI-enabled marketplaces. Marketplaces where AI does high-value matching or workflow automation – hiring platforms, specialized vertical marketplaces. The AI is not the marketplace. The AI makes the marketplace dramatically better at its core job of matching supply and demand.

Agent tooling. “Agents for everything” is the direction. But agents are nondeterministic – they behave unpredictably, they fail in novel ways, and debugging them is not like debugging code. The tooling layer to observe, control, and debug swarms of agents – monitoring, safety, behavior tuning – is a major investment category.

The example they loved: Quilter, which uses reinforcement learning to auto-design PCB (printed circuit board) layouts. This is a hard, manual, expert-dependent task that AI can now do. It is not a wrapper. It is not a chatbot. It is AI doing real engineering work that previously required years of specialized training.

4. What Is Not Working (Yet)

Consumer AI. Most consumer AI plays are struggling because incumbents – the big consumer platforms – still own distribution and attention. Building a consumer AI product means fighting Google, Apple, Meta, and OpenAI for users. The panel expects winners to emerge eventually, but right now the distribution advantage of incumbents is overwhelming.

Generic knowledge-worker tools with light AI. “We took the existing SaaS product and added an AI copilot” is not a fundable thesis. Without a data moat or a deeply embedded workflow, these products are fragile. The moment the underlying model gets better (or cheaper, or integrated into the platform the customer already uses), the thin AI layer gets commoditized.

AI-layer-on-top-of-X business models. Many current AI startups are generic SaaS subscriptions for AI features. Margins will get squeezed as it becomes trivial to build software. Building software was already getting easier before AI – now it is approaching near-zero marginal cost for many categories. The panel expects deeper verticalization and new business models to matter more than “AI layer on top of X.”

5. Incumbents vs AI-Native Startups

This was the most debated topic on the panel. Two views:

The incumbent advantage thesis. Vertical SaaS companies like Procore (construction), ServiceTitan (home services), and Toast (restaurants) will benefit from AI as a tailwind. They have the data, the distribution, the customer relationships, and increasingly the marketplace and partner programs to stay dominant. AI makes their existing products better. Their customers are not switching to a startup – they want AI features in the tool they already use.

The AI-native disruption thesis. Many incumbents will not make the AI transition fast enough. Their codebases are old. Their teams are not AI-native. Their organizational incentives are to protect existing revenue, not cannibalize it. AI-native teams with strong product and research DNA can out-innovate them. The panel cited Granola as an example – a startup out-producting incumbents by building AI-native from day one.

The synthesis: The most powerful disruption comes from new business models enabled by AI, not just adding AI to existing products. The analogy: Google did not beat Microsoft by building a better word processor. Google used ads to make search free, which was a business model innovation that made per-seat software pricing obsolete. The AI equivalent is a startup that uses AI to deliver outcomes at a fraction of the cost of the incumbent’s seat-based pricing – making the incumbent’s entire business model the vulnerability, not just their product.

6. Pricing and Business Models for AI

Seat-based pricing is dead for AI. The panel was direct: charging per user for AI products does not make sense when the AI does the work that users used to do. If your product replaces 10 customer support agents, charging per remaining human seat is backwards. Pricing will revolve around work or outcomes, not users.

Outcomes-based pricing is attractive but tricky. “Pay when a ticket is resolved” sounds clean. But attribution is hard – did the AI resolve it, or did the customer give up? Incentive alignment is complex – does the AI optimize for resolution speed at the cost of quality? The mapping between cost (GPUs, tokens) and customer value (outcomes, results) is still fuzzy. Nobody has fully solved this.

What they still like:

• SaaS for predictability. Recurring revenue is still the gold standard for investor confidence, even if per-seat pricing is dying.
• Hybrid models. Software bundled with hardware – robots, compute systems, devices – often sold with subscriptions. The hardware creates switching costs. The software captures ongoing value.
• Marketplaces with strong flywheels. Take rates on transactions, powered by AI matching and workflow automation. The more transactions, the better the AI gets, the more transactions flow through. Classic flywheel with AI acceleration.

The margin question. Many early AI startups raised large rounds that were effectively “X million for compute, Y million for people.” Investors are still figuring out sustainable margins. If 60% of your revenue goes to API costs, your margin profile looks more like a services business than a software business. Public-market valuations will eventually reflect this.

7. Non-AI and “Around-AI” Themes

The panel also discussed investment categories that are not pure AI but benefit from or surround the AI wave:

Pen-and-paper to digital in blue-collar/SMB industries. This classic digitization play is now supercharged by AI’s ability to parse unstructured documents – invoices, permits, inspection reports, handwritten notes – and reduce manual data entry. Industries that were too fragmented or low-tech for traditional SaaS are now addressable because AI lowers the cost of onboarding messy, unstructured data.

Infrastructure for AI-generated software. If AI generates 10x more code, every system around code – security scanning, storage, CI/CD, app infrastructure, cloud and on-prem systems – needs to scale. The explosion of AI-generated software creates demand for infrastructure that handles the volume, velocity, and quality variance of machine-written code.

Physical AI and robotics. Vision-language-action models are enabling robots to do meaningful work in the physical world. The panel thinks “this might finally be robotics’ moment” – previous waves of robotics investment failed because the AI was not good enough. Now it might be. Investment in robots, sensors, actuators, and the software that orchestrates physical-world agents.

Onshoring of critical materials. Magnets for robotics, rare earth processing, semiconductor materials – physical-world bets that are “uncorrelated” to software cycles. If robotics takes off, the supply chain for robot components becomes a bottleneck and an investment opportunity.

Healthcare and holistic health. Long-term, outcome-priced, high-alpha verticals. Human bodies still exist, still deteriorate, and still need care. AI-powered diagnostics, treatment planning, and health monitoring – priced on outcomes rather than visits – represent one of the largest addressable markets on the planet.

8. Moats, IP, and Fundraising Strategy

On moats and defensibility:

“AI feature” alone is not a moat. Defensibility comes from:

• Proprietary or hard-to-access data – data that your product generates through usage and that competitors cannot replicate
• Deeply embedded workflows – when your product becomes the system of record for a critical process, switching costs are enormous
• Superior product experience – the product is simply better, faster, more reliable than alternatives
• New business models – pricing and delivery mechanisms that incumbents cannot adopt without cannibalizing their existing revenue

IP (patents, trade secrets) is nice but only one factor. Speed, execution, hiring, distribution, and problem insight matter more. A patent does not help if someone builds a better product faster.

On fundraising and stage risk:

Investors mentally track three risks: team, market, and product/tech. Each funding stage – seed, Series A, Series B, IPO – is about de-risking one or more of those:

• Seed – all three risks are high. The bet is primarily on the team.
• Series A – team risk is partially de-risked (they shipped something). Market risk is the question.
• Series B – product works, market exists. The question is scalability and unit economics.
• Late stage/IPO – most risks are de-risked. The bet is on execution and growth.

Early-stage rounds are cheaper because all three risks are present. Late-stage valuations assume those risks are significantly reduced. The panel prefers getting in early – pre-seed, seed, Series A – before the “wall of capital” at later stages drives valuations into territory that requires enormous outcomes to generate returns.

Advice for founders:

• For hardtech (e.g., robotics, physical sensing): start with design partners. Use early customers to validate the tech and collect data before scaling manufacturing. Do not build the factory before you know the product works.
• Solve a real, painful problem. “Cool tech in search of a use case” does not get funded. Start with the problem, not the technology.
• For first-time founders: articulate clearly why you are the right person to build this company. De-risk team, market, and tech in your pitch – show that you have thought about all three.
• Warm introductions still matter. The investors shared their contact information and said they read inbound, but warm intros from trusted sources get priority. Network deliberately.

9. The Founder Story That Gets Funded

The panel was clear: they want to hear why you, specifically, are the person who must build this business. Not “I saw a market opportunity.” Not “AI is hot.” They want lived experience, domain expertise, repeated exposure to the problem, and demonstrated grit.

The example that resonated most: Jordan Taylor of Viscom. Taylor went from car design at Honda to Nvidia’s experimental automotive group. Then he burned personal and family savings, lived on Costco hot dogs, and drove to Silicon Valley to build what he describes as “an industrial design Figma for the physical world.” That level of obsession – where the founder literally cannot imagine doing anything else and has sacrificed materially to prove it – is what sticks with investors.

The pattern: founder identifies a real problem through years of firsthand experience, becomes the domain expert, makes personal sacrifices to pursue the solution, and builds with a mission that goes beyond “this is a good business.” The mission is the moat, because mission-driven founders do not quit when the market shifts or the model landscape changes.

---

What This Means for WorkingAgents

Several points from this panel map directly to WorkingAgents’ positioning:

1. “Agents for everything” plus tooling to observe, control, and debug agents – the panel explicitly identified agent governance, monitoring, and safety as a major investment category. WorkingAgents is this tooling layer.

2. AI feature alone is not a moat; deeply embedded workflows are. WorkingAgents becomes the system of record for AI agent governance – the audit trail, the permission engine, the guardrail layer. Once embedded, switching costs are high.

3. Outcomes-based pricing and the death of per-seat. WorkingAgents’ pricing can align with agent actions governed, not seats. Every agent interaction that flows through the gateway is a unit of value.

4. Deep vertical AI in regulated industries. Healthcare, insurance, financial services – the verticals the panel highlighted as highest-value – are exactly where governance is non-negotiable. You cannot deploy AI agents in healthcare without HIPAA-compliant audit trails. WorkingAgents provides that.

5. Infrastructure for AI-generated software and agent swarms. As AI generates more code and more agents, the governance and security layer around those agents becomes critical infrastructure, not optional tooling.

6. The incumbent vs. AI-native question. WorkingAgents is AI-native. It was built for agents from day one, not bolted onto an existing product. That is the positioning the panel said wins against incumbents.