Airspeed co-founder and CTO Devang Agrawal joined Authority Magazine for its series on entrepreneurs pushing the boundaries of AI. His thesis fits in one line: insight is cheap, action is everything.
The interview, conducted by Gabriel Borden of Arrow Fund and published in July 2026, covers a lot of ground. Kanpur to Cambridge, DeepMind to a $20M Series A, and the specific engineering problem that turned a research career into a company.
The line that runs through his whole career
Asked for a favorite life lesson quote, Agrawal gave the sentence Airspeed is effectively built on: “Knowing what’s happening isn’t the same as doing something about it.”
He explains why it stuck. For roughly a decade, enterprise software got very good at telling revenue teams what was going on. Dashboards, transcripts, reports, richer CRM data. The actual work, the emails and the follow-ups and the deal updates, still fell entirely to people.
“It’s why we built Airspeed to close the loop instead of adding another dashboard,” he told Borden, “and it’s the standard I try to hold myself to. Don’t just understand the problem, do the work.”
He traces the instinct back to his family in Kanpur, where his grandfather ran a clothing manufacturer and his father ran a chemical business before pivoting into CAD design. “Watching people build was just the water I swam in.”
What did he see at DeepMind that made him leave?
A Christmas quiz. In the very early days of Gemini, well before ChatGPT and public LLM APIs, someone entered DeepMind’s internal model into the team quiz. The researchers were skeptical. The model did well.
“Seeing a language model hold its own against a room of researchers, before the rest of the world had any idea this was coming, was when I knew I had to go build a company with this technology.”
Combine that with the conversation understanding work he had done at Apple, and the shape of the problem became clear. The next thing worth building was not more insight. It was execution. He and Adam Liska left to found the company in 2022, a story Liska picks up in his own conversations with GTMfund and on Pavilion’s Topline Spotlight.
Systems of record, systems of intelligence, and the missing third
Agrawal’s clearest framing of the gap is structural: “Teams have systems of record and systems of intelligence. What they didn’t have was a system of action.”
The CRM is the system of record. Call recording and conversation intelligence tools are the system of intelligence. Both describe. Neither acts. The third layer has to understand a team’s commercial context and then do the work on top of it, which is the argument behind the commercial brain and behind why agent harnesses built for individuals break at enterprise scale.
What made it real, he says, was getting agents to act reliably on the live deal. Reading context across calls, emails, tickets, and CRM, then writing MEDDIC fields, drafting follow-ups, and scoring calls for coaching, rather than working from a stale snapshot.
Why trust is the hardest part of shipping agents
“The hardest challenge was trust,” Agrawal said. “When agents take real actions, they can’t hallucinate or act on outdated data.”
His answer was to refuse the shortcut. “Most teams retrofit AI onto legacy systems; we did the opposite and built the foundation from scratch.” That foundation has four parts: a unified understanding of commercial context, an agent runtime with guardrails, human approval before anything goes out, and rigorous evaluations so every action can be trusted.
The evaluation piece is where his research background shows. He lists intellectual honesty as the trait most instrumental to his success, and defines it concretely: asking whether something actually works rather than whether it looks impressive in a demo. “When an AI agent takes a real action, you have to be honest with yourself about how often it’s right.”
What that looks like for a revenue team
Agrawal points to Foleon, an enterprise content platform, which he says saved more than $193,000 and reclaimed roughly six hours per sales rep per week in its first 90 days on Airspeed. We wrote up how they got there in Foleon put a stopwatch on their calls.
The moment it clicked for him personally was smaller and stranger. He ran an agent inside Airspeed’s own instance that read across thousands of calls and emails, spanning hundreds of deals, and came back with a genuinely good read on where his next quarter would land. “Work that would take a team weeks, done in minutes,” he said, and accurate enough to act on.
That experiment is now a product. Airspeed forecasting builds the number from real deal signals rather than a rep’s gut feel.
His five things for shaping the future of AI
Borden’s central question asks every guest for five things they need to know. Agrawal’s, condensed:
- AI isn’t the value, execution is. Winning products close the loop. They write the follow-up, book the next call, and send the requested document once a deal stage moves.
- Build the foundation from scratch, don’t retrofit. Bolting AI onto legacy systems produces fragile results. A purpose-built context layer plus an agent runtime is what makes actions reliable.
- Trust is the product. Guardrails, human-in-the-loop approval, and rigorous evaluations are what let an enterprise actually deploy autonomous agents.
- Act on live context, not stale snapshots. The system has to work from what is true about a deal right now, across calls, emails, tickets, and CRM, which means contextual memory and cross-search over the whole data base.
- Research depth, commercial discipline, and user empathy. Category-defining AI companies combine all three. Pure research or pure hustle is not enough.
What he’d tell founders building in AI
“Solve the part everyone else skips.” For a decade the industry kept improving visibility while the work stayed manual, so the hard and valuable problem was the one nobody was taking.
His second piece of advice is about positioning, and it cuts against most AI marketing: “Don’t lead with ‘AI’ either; lead with the value and the outcome for the user, because AI is the means, not the pitch.”
The third is about sequencing. Earn trust before autonomy. Ship the guardrails and the evaluations first so people can rely on what the system does. Airspeed built its evaluation framework in the early days for exactly that reason. “That’s what earned us the right to automate the work, not just surface it.”
Read the full interview
The whole conversation, including the DeepMind Christmas quiz, what he learned from his father and grandfather, and why he would spend a lunch with Mark Zuckerberg, is on Authority Magazine: Devang Agrawal of Airspeed On Pushing the Boundaries of AI, interviewed by Gabriel Borden.
If the gap between knowing and doing is the one you are trying to close on your own team, that is the problem Airspeed was built for. Book a demo and see an agent run on a real deal.
Frequently asked questions
Who is Devang Agrawal?
Devang Agrawal is the co-founder and CTO of Airspeed. He was a research scientist at Google DeepMind, where he worked on multimodal retrieval and language models, and worked on conversation understanding at Apple before that. He co-founded Airspeed with Adam Liska, which now serves hundreds of revenue teams across more than 20 countries on a $20M Series A led by DN Capital.
What is a system of action?
A system of action is software that does the work rather than reporting on it. Agrawal's framing is that revenue teams already have a system of record in the CRM and a system of intelligence in their call and email tooling. What has been missing is a layer that understands the live commercial context and then writes the CRM fields, drafts the follow-up, and scores the call on its own.
How do you make AI agents safe enough to act on real deals?
Agrawal names four requirements: a unified understanding of commercial context, an agent runtime with guardrails, human approval before anything goes out, and rigorous evaluations of every action. His argument is that trust has to be built into the foundation rather than added to a legacy system afterward, because an agent that takes real actions cannot hallucinate or work from stale data.
