Overhyped Today, Underhyped Tomorrow: A Conversation with Teo Borschberg

Teo Borschberg has a simple rule for reading the AI hype cycle: everything is overhyped in the short term and underhyped in the mid term. He's been building in this space long enough to say it with a straight face.

He's the co-founder and CEO of Augment AI, focused on putting AI agents to work safely inside organizations. Before that he founded Otto AI, a voice-analysis company acquired by Unity Technologies, and he's built companies across Hong Kong, Shanghai, Lisbon, and San Francisco, where he is now. Eighteen years of entrepreneurship, several exits and pivots, and a very grounded read on what agents can and can't do yet.

Agents are powerful, and dangerous in the wild

Teo draws a clean line. An LLM is raw intelligence. An agent is a small entity that can autonomously take an action and make a decision, and that autonomy is exactly what makes it dangerous in the wild. As people hand agents credentials and API and browser access, an agent can delete a database, make a purchase you didn't intend, or send an email you never wanted sent. The horror stories are already coming from the pioneers who gave too much autonomy too fast. So an agent today needs infrastructure and guardrails to operate around risky actions and sensitive data.

His analogy for the cost of a breakthrough is aviation: the first flight lasted 30 seconds, accidents followed, and 60 years of trial and error later we routinely fly 300 people across the Atlantic. The stakes with agents are lower than death, but the pattern is the same, move fast, learn from concrete experience where the limits are, and build the guardrails around them.

Why human-in-the-loop isn't enough

The base guardrail is human-in-the-loop: classify every sensitive action as sensitive, deterministically, and let a human approve it. The problem is approval fatigue. Sit in front of a stream of approvals and you start clicking approve, approve, approve without reading, which defeats the purpose. So Augment layers stochastic rules on top of the deterministic ones: another agent checks the intent of an action against the user's guidelines. It isn't perfect, but it adds a contextual, dynamic layer of judgment on top of the hard rules, a second check on whether the action actually matches what you meant.

Use AI to learn AI

Ask Teo how a non-technical person levels up, and the honest caveat comes first: six to twelve months ago you basically had to be a developer, juggling API keys, to do anything interesting. It's only in the last two or three months that non-technical users can truly build, and he's betting that by year's end it'll be as simple as chatting with the system. For now, his blueprint is two steps: watch YouTube videos, and just ask Claude or ChatGPT (he recommends Claude for its connectors, plugins, marketplace, and skills). Install it, watch a couple of videos, and when you're stuck, tell it you're stuck and let it guide you forward. You don't need to hunt for tutorials anymore. Use AI to learn AI.

What happens to the org chart

The Valley fantasy of a three-person, $100M company is real, Teo says, but it won't be the rule, because plenty of businesses are operational in the physical world. For knowledge work, though, he already sees it: you stay the decision-maker and strategist, and below you sit 5, 10, 50 agents working in different capacities. Crucially, the leaders he talks to don't want to cut costs. They want to triple or 10x revenue with the same team, and that capacity has to come from agents.

On jobs, he pushes back on pure doom. History shows that adding efficiency tends to expand markets and employment, and we're nowhere near optimal, four-hour hospital waits, 45 minutes on hold with an airline. His favorite example is the call center, where the metric is flipping 180 degrees: from first-call resolution (never hear from the customer again) to how long can we keep a relationship going. AI answers the basic questions, and humans get pushed forward to build relationships, care, and delight. But he's candid that the transition will be painful, because companies now ask "can an agent do this?" before every hire, and roles where the answer is yes will be hard to enter until new opportunities emerge.

The skills that hold their value

Two things Teo prizes now: generalists, and critical thinking. AI makes everything sound good, and he admits being a victim of it himself, skimming a plausible answer, deciding to ship it, then realizing it's actually wrong. The people who can pierce that noise and steer strategically are the valuable ones. It's the podcast-guru problem: a commanding voice and impressive accolades make you assume someone's right without checking. AI is the same, often right, sometimes not, and never aware of your specific situation. What worked for another company at $30M might not work for you.

Hope your customers kill your idea

The founder mistake that reshaped how he operates is the fine line between optimistic and objective. He used to want something to happen so badly that he wouldn't see the facts, and he'd massage customers with leading questions until they agreed with the vision he'd already chosen. Now he walks into customer conversations hoping they'll kill his idea, because that gets him to the truth faster. Getting slapped in the face by a customer's honest answer, he says, is genuinely good for you. It's the same instinct as trying to disqualify a lead rather than force-fit it, you build trust and you find reality sooner.

The 11 times he wanted to quit

Asked how you know when to stop versus stay the course, Teo admits he hasn't cracked it. He takes notes weekly because his memory is terrible, and when he reread the 300-page journal of his previous company, he counted eleven distinct times he had deeply wanted to quit. That company exited to Unity. His conclusion is that persistence is the difference between founders who make it and those who don't, quitting is always available, so as long as you keep listening and iterating rather than hitting the same wall, you stand to live another day. He's pivoted five or six major times across his businesses.

On Europe, and where he'd start today

Having built on both sides of the Atlantic, he's blunt: Americans are pioneers and risk-takers; Europeans lean toward regulation and caution and waiting to see what happens. As a European who loves Europe, he finds it frustrating, and on AI specifically he feels strongly that the continent needs to wake up, pointing to ElevenLabs as proof the talent is there. If he were starting his career today, he'd aim at the crossroads of technology and real-world operations, because as software trends toward commodity, the value moves to whoever can wed AI to transforming physical industries. And he expects an experience economy to follow, where manual skills and human experiences become high-value again as the pendulum swings back.

Key takeaways

Agents need guardrails, not just autonomy. With credentials and access, an agent can delete, buy, or send things you never intended. Deploy them behind infrastructure.

Human-in-the-loop breaks under approval fatigue. Add a second, intent-checking layer on top of deterministic rules, because a human clicking approve without reading is no safeguard.

Use AI to learn AI. Watch videos, then let Claude or ChatGPT guide you when you're stuck. The barrier to entry just dropped hard.

Critical thinking is the durable skill. AI makes everything sound good. The value is in piercing the plausible-but-wrong and steering for your actual situation.

Persistence is the differentiator. He wanted to quit 11 times before an exit. Iterate, don't hit the same wall, and live to fight another day.

Frameworks worth stealing

Deterministic rules plus a stochastic check

Classify sensitive agent actions deterministically and require approval, but assume approval fatigue will set in. Layer a second agent that evaluates each action's intent against the user's guidelines, so you have a contextual judgment on top of the hard rules rather than relying on a human rubber-stamping a stream of prompts.

Hope they kill your idea

Go into customer conversations trying to get your idea killed rather than validated. Don't massage the questions to lead people to your vision. The honest, negative answer gets you to the truth faster and builds more trust than a polite yes you engineered.

Quotes worth keeping

The lines I wrote down.

Everything is overhyped in the short term and underhyped in the mid term.

Now I go into customer conversations hoping they'll kill my idea, because then I get to the truth faster.

I counted it. There were eleven times I deeply wanted to quit. And then we exited to Unity.

Rapid fire round

Same questions every guest. Quick questions, quick answers.

Best advice you've ever received? When you're in trouble, have a glass of wine, sleep, and wake up tomorrow with a fresh perspective.

Advice you ignored and wish you'd listened to? Nothing in particular. He listens to advice a lot.

What would you tell your younger self? Trust the process.

Ongoing challenge that keeps you up at night? The overwhelmingly fast pace of change.

Favorite spot? A restaurant in Zermatt, Switzerland, Swiss mountain food done really well, cheese and potatoes, a view of the Matterhorn, and Swiss chocolate to die for.

---

Teo Borschberg is the co-founder and CEO of Augment AI, which helps organizations deploy AI agents safely. He previously founded Otto AI, acquired by Unity Technologies, and has built companies across Hong Kong, Shanghai, Lisbon, and San Francisco. Find him on LinkedIn, and Augment's "AI in the trenches" case studies on YouTube.