Andreas Cleve likes to think of a startup as a book. The metaphor is tidy enough to sound decorative, until he gets to the useful part: a founder must know which chapter the company is writing now. Each chapter needs one governing idea. Wander into too many subplots and the whole thing becomes expensive fan fiction.
Corti, the company Cleve co-founded with machine-learning researcher Lars Maaløe in 2016, has written several distinct chapters. The first took place in emergency dispatch, where software listened to urgent calls and looked for patterns while a human professional stayed in charge. Another moved into documentation, coding and the sprawling administrative work surrounding a consultation. The current chapter is infrastructure: specialized models, APIs, speech tools and controlled agents that other software companies can build into their products.
The plot has widened, but the protagonist is still a conversation. Cleve's wager is that a great deal of institutional knowledge arrives through speech, then gets thinned out as it passes into forms, codes and software systems. Corti tries to preserve more of that context without turning the professional's day into an extended dialogue with a robot. The technology is meant to be present, useful and mostly out of the way.
A founder taught by workflows
Cleve grew up outside Copenhagen and studied leadership at Copenhagen Business School. He is candid that business school did not provide a straight line into his career. Technology had been a long-running curiosity, though. In one early job, he persuaded an established commerce company that he could build a mobile-friendly interface. His front-end skills, by his own later account, were rudimentary. The project worked well enough to pull him toward running the company's e-commerce operation.
His first startup at 27 was Ovivo, a workforce-planning platform for hospitals. It expanded across Denmark and was acquired in 2013, after which Cleve led a business unit at the software company EG. He later co-founded Hyvi, a research initiative concerned with context-aware language models and real-time conversation. That inquiry evolved into Corti, formed with Maaløe, whose research background gave the partnership its technical counterweight.
There is a clean progression in retrospect: commerce taught Cleve interfaces, Ovivo taught him institutional workflow, Hyvi explored language, and Corti placed language inside a demanding workflow. Retrospect, of course, is a shameless editor. It removes the abandoned drafts, awkward transitions and alarming bank statements. Yet the continuity is real. Cleve has spent much of his career studying how information moves through organizations and what happens when software arrives at the handoff.
Start where failure is expensive
Corti's opening market was a wonderfully inconvenient choice. Emergency calls contain accents, interruptions, poor connections, heightened emotion and little patience for a loading spinner. The product needed to operate in real time, fit the call taker's process and make its contribution without pretending to replace professional judgment. It was a setting that punished the usual startup indulgences.
The Seattle Fire Department began using Corti for triage in 2019. By then, Cleve had been named to MIT Technology Review's 2018 Innovators Under 35 Europe list for the company's work on emergency-call analysis. Recognition brought attention, but the more valuable result was an education in production. A model could look persuasive in isolation and still fail the larger system because it was slow, awkward, unauditable or poorly matched to the user's sequence of decisions.
“Every founder could really benefit from having a key idea of what chapter in the book they're at right now.”Andreas Cleve
This explains why Cleve now speaks about AI with the vocabulary of infrastructure. His public arguments dwell on model cards, audit trails, data residency, private endpoints and deployment options. These are not the phrases that make a keynote audience levitate. They are the terms that determine whether software crosses the distance between a pilot and an operating environment.
Corti's $60 million Series B in 2023, co-led by Prosus Ventures and Atomico, gave the company room to broaden that work. In 2024 it reported that its U.S. revenue had doubled and that it supported more than 100,000 patient interactions a day across all 50 states. The company established its U.S. headquarters in New York, where Cleve is now based, while keeping Copenhagen as its headquarters and research center. London joined the map too.
Corti's chapters
Emergency conversations
Real-time support
Models and APIs
Agent infrastructure
The case for going narrow
The arrival of general-purpose large language models might have made a specialized AI company look quaint, like opening a bespoke umbrella shop during a drought. Cleve drew the opposite conclusion. General models demonstrated appetite. Domain-specific work would determine what could actually be served.
His case rests on depth. Clinical language carries specialized terms, local documentation customs, institutional rules and consequences that extend well beyond a fluent paragraph. A company building for that environment must evaluate its systems against the work, not merely admire their prose. Cleve has compressed the argument into a line: “It's not a prompting problem; it's a research problem.”
In early 2025, Corti introduced specialized foundation models and APIs for speech, text generation and coding. Instead of trying to own every finished application, it positioned those components beneath other companies' software. This is a less visible place to operate and potentially a more durable one. An application can win attention with an elegant screen. Infrastructure wins when many screens quietly depend on it.
Cleve has also rejected the complaint that healthcare already has too many AI apps. His view is that the category will support many products because the workflows are numerous and particular. The strategic challenge is not to crown one universal interface. It is to give builders reliable components so they do not each have to become a machine-learning lab and compliance department before lunch.
“Healthcare doesn't need smarter agents - it needs safer autonomy.”Andreas Cleve, 2026
Governance enters the product
By February 2026, Corti had launched an Agentic Framework and Agent Library. The announcement said more than 100 companies were building on its infrastructure across over 50 use cases. Its core idea was controlled execution: validate what an agent is allowed to do, keep a record of what happened and prevent one plausible mistake from cascading through connected systems.
This is the point where Cleve's work in policy and community becomes relevant. He co-founded Nordic.ai, a nonprofit intended to connect machine-learning practitioners across the region, and chaired AI Copenhagen. He has advised DIGITALEUROPE and served on Denmark's National Digitization Council. In 2026 he also joined the board of the Danish Centre for AI Innovation. The pattern is broader than corporate lobbying. He has treated the surrounding ecosystem - talent, rules, public understanding and compute - as part of the building environment.
That position gives him a useful double vision. Regulation can create needless delay when it ignores how systems are updated and monitored. It can also describe the product customers need: traceable versions, clear accountability, regional control and evidence that survives scrutiny. Cleve's pragmatic answer is to place those requirements inside the architecture early, when they are still design choices rather than expensive apologies.
Corti's 2026 product rhythm followed the thesis. It released an agent framework in February, medical-coding technology in April and speech-to-text tooling in May. In August, Corti and the Danish Centre for AI Innovation announced a sovereign AI control layer for European enterprises. The original listening company was now selling the means to control how many kinds of AI act.
A quiet ambition
Cleve is direct about scale. He wants medical expertise to become more widely available, across languages and geographies, without asking professionals to surrender judgment to a black box. It is an enormous aspiration phrased through decidedly unromantic machinery: evaluation, routing, latency, governance and APIs.
His personality in interviews combines mission with a founder's appetite for a memorable analogy. He will talk about a company as a book, describe himself as a Swedish-Danish cocktail and then return to a narrow operational question. He values focus partly because he knows ambition generates distractions. A large mission can make every adjacent project sound compulsory.
The transferable lesson in Corti's story is not “start a healthcare AI lab.” It is to choose a first problem that teaches. Emergency calls forced the company to learn what conversational AI looks like under pressure. Those constraints generated capabilities that could travel: speech recognition, contextual understanding, workflow integration and, eventually, controlled execution. The wedge became a curriculum.
A decade into the work, Cleve is still writing around the same observation. Conversation is where context gathers. Software becomes valuable when it carries that context forward without becoming the loudest participant. The coming wave of AI may be full of agents eager to speak and act. Corti's bet is that the enduring layer will know how to listen, when to stop and how to leave a receipt.