The AI economy’s unbuilt layer
AI’s first decade of commercialisation focused on capability. A handful of labs competed to build increasingly powerful foundation models, while billions flowed into the compute and infrastructure needed to support them. As those models improved, attention shifted to a second layer of the market: the applications built on top. Coding assistants, vertical software products, and enterprise copilots emerged to transform raw model capability into something businesses could actually buy and use. What’s emerging now is different in kind rather than degree.
The AI economy begins where AI systems stop supporting commercial activity and start conducting it. Agents are no longer confined to generating information or recommending actions. Increasingly, they can execute workflows, coordinate tasks, navigate software, negotiate transactions, and act on behalf of users.
When Anthropic ran Project Vend in 2025, allowing Claude to manage a small office shop in its San Francisco office, the model sold tungsten cubes below cost, regularly handed out discounts and directed customers towards payment methods that did not exist. Anthropic’s conclusion was notable. Many of the failures stemmed less from shortcomings in intelligence than from the absence of surrounding systems and controls. The agent could reason. What was missing was the infrastructure necessary to participate reliably in economic activity.
How agents are becoming participants in the economy
Fifteen months on, that is roughly what happened. Agents now have a standard way to reach live systems through Anthropic’s Model Context Protocol (MCP). They have sanctioned routes to buy through emerging commerce protocols from companies such as OpenAI, Google, and Shopify. Stripe’s Machine Payments Protocol, launched in March with Tempo, starts from the same premise: the financial system was built for humans, and agents can’t use it. New payment frameworks are beginning to acknowledge a future where a human is not necessarily present for every transaction. The direction of travel is clear. The industry is moving from proving that agents can act to figuring out how they can transact.
All of that represents a significant shift from the generation and retrieval of information that AI models and applications have largely focused on to date. Agents are increasingly capable of navigating software, executing actions and working together across complex workflows, with different systems responsible for different parts of a broader process.
Much of today’s commercial infrastructure assumes a human sits behind every transaction. Know-your-customer requirements assume a human customer. Compliance processes assume human decision making. Liability frameworks assume responsibility can ultimately be traced back to an individual person.
An organisation may be able to establish exactly what an agent was instructed to do. It may have a complete audit trail showing every action taken and every decision made. But proving authorisation is not the same thing as determining accountability. If an agent operates within its approved parameters and still creates an undesirable outcome, existing frameworks provide few clear answers about where responsibility should sit.
As agents become more active participants in economic activity, those unanswered questions become increasingly important. And they point directly towards the next generation of startup opportunities.
Where the next wave of AI startups will emerge
If agents can execute transactions and act on behalf of organisations, businesses will need entirely new ways to verify authority, govern behaviour and manage risk. The next generation of startups will build identity systems for agents, permissioning frameworks, audit trails, compliance tooling, governance platforms, and insurance products designed specifically for autonomous systems. While these categories may appear less exciting than the latest consumer AI applications, they can address a fundamental requirement: organisations will only deploy agents at scale if they can trust them.
Financial services, insurance and healthcare have long been viewed as challenging markets for startups to break into, but the transformation potential of agents may ultimately make some of them the most promising. An agent that helps draft a marketing email presents one level of risk. An agent that commits funds, negotiates a contract or executes a regulated process presents another entirely. Before organisations allow those systems to operate at scale, they need confidence that the appropriate safeguards are in place.
Founders must innovate without outpacing enterprise trust
The organisations founders most want to sell to operate in regulated environments, where trust matters as much as innovation. That creates a balancing act. Build too cautiously and you risk creating a feature rather than a company. Build too far ahead of customer readiness and adoption stalls because the limiting factor is no longer what the technology can do, but what organisations are comfortable allowing it to do.
The most valuable applications of AI are likely to emerge within complex, document heavy workflows that were previously beyond the reach of automation. Those workflows often represent the greatest commercial opportunity precisely because they are difficult, heavily regulated, and expensive to operate. But they also demand deep integration with existing systems, governance structures, and compliance requirements.
For that reason, differentiation in the next generation of agent native startups is likely to come less from reasoning capability alone and more from trust, traceability and enterprise readiness. That shift moves the interesting problems away from capability and towards the plumbing around it: identity, permissioning, oversight, cover. It is slower work, and less visible than building the agents themselves.
Building the foundations of the next AI economy
The emergence of agents capable of managing workflows, conducting transactions and participating directly in economic activity marks a significant shift in AI’s evolution. But as agents become more capable, they also expose gaps in the infrastructure that supports them.
Governance, compliance, identity, permissioning, insurance and accountability may attract less attention than the agents themselves, but they increasingly look like the foundations the agent economy will require. If the last wave of AI innovation focused on what models could do, the next may focus on what organisations can safely allow them to do. For founders, that’s where some of the most interesting opportunities may lie.
For more startup news, check out the other articles on the website, and subscribe to the magazine for free. Listen to The Cereal Entrepreneur podcast for more interviews with entrepreneurs and big-hitters in the startup ecosystem.
