‘AI isn’t a category’: Inside the mind of technical VC Jéssica Leão

Jéssica Leão spends the first two hours of her mornings building what she calls a "mental map of patterns" across the technical AI infrastructure landscape. Her process involves reading research papers on AI, tracking announcements and product launches, and keeping up with what AI leaders are saying. She synthesizes her analyses in a weekly AI blog, updating her investment thesis in real time.

"You'd be hard-pressed to find AI news I haven't heard within five to 10 minutes," she says.

This information diet, combined with time spent as a forward deployed engineer at Palantir, has given Leão a unique vantage point into venture capital.

Today, she's a partner at Decibel, an early-stage venture capital firm, where she invests in technical founders at the Seed and Series A stages building technical products. This spans infrastructure software like developer tools, data engineering platforms, and cybersecurity, as well as autonomous software. She focuses on areas where complex AI models meet real-world problems.

I sat down with Leão to understand how she evaluates companies in the fastest-moving market in tech history and what she believes makes a company win today.

Jéssica Leão

Palantir to venture: Learning the hidden infrastructure

Leão’s investing lens was shaped by her hands-on role at Palantir, where she learned the pipes and plumbing inside of companies.

“As a forward deployed engineer, you're doing everything: understanding customer use cases, making pilots work, but also debugging pipelines and making customer data useful. You end up learning SQL, Python, and working with disparate data systems,” she explains.

She also credits Palantir’s internal development team for her interest in cybersecurity. “We built our own developer tools. And because we worked with government customers, we were thoughtful about how we federated systems and kept everything private.”

Leão says it was a master class in all things infrastructure, which eventually led her to venture capital.

AI as a computing paradigm

Ask Leão about investing in AI, and she'll push back on the framing.

"I view AI as a computing paradigm rather than a category," she says. "Saying you're investing in AI right now is almost like saying 'I invest in startups.'”

Leão believes every company should be incorporating AI into its tech stack since it’s clearly a hyper-powerful technology. “If you're not adding the latest and greatest technology into your stack, you are obviously behind,” she says.

There are several use cases for AI, she notes. AI doesn't have to be a chatbot. You might build a natural language interface for codebases, or you might use LLMs behind the scenes to summarize data, route requests, or enrich information — invisible to end users but core to how your product works.

"There are many things LLMs are really good for that work best as part of the back end, not in front of users," she concludes.

But there's a crucial distinction for her between companies that slap on AI for marketing purposes and those that think critically about where the technology creates value.

Leão’s favorite use case: End-to-end agents

A use case Leão is especially excited about are agents, which are end-to-end autonomous systems capable of performing multi-step tasks toward a goal autonomously.

"When I think about applications that have a continuous human in the loop, the human is still in the driver's seat and the AI is helping to copilot them," she explains. "But that's still a one-to-one interaction."

“With autonomous agents, you're not just making someone 10x faster—you're enabling things that couldn't happen at all with humans in the loop."

A strong example of an autonomous agent comes from Strella, one of Leão’s portfolio companies. Strella has built an AI system that handles UX and market research from start to finish — the kind of work that typically drowns product managers in Zoom calls, transcripts, and presentation decks.

Strella's AI moderator maintains a calendar where research participants can self-schedule interviews at any time—no back-and-forth coordination needed. The moderator conducts in-depth conversations with users and can run multiple interviews simultaneously. “It's a real conversation—just conducted by AI,” Leão explains.

And the time savings and faster results are only part of the benefit.

"Turns out, people are actually more transparent with an AI moderator. It's easier to give real feedback on a product to AI than to the person who built the product," Leão says.

The AI analyzes the interview to pull out key information and compare the transcript to all other transcripts to assess patterns. Another software, meanwhile, creates highlight reels of the most important moments for product managers to share with their teams. It also generates presentation-ready decks summarizing all the insights.

The technology that powers agents like Strella is giving rise to a new wave of infrastructure opportunities Leão is actively exploring.

Building an autonomous agent demands far more than plugging into the ChatGPT API, she notes.

To manage conversations, Strella, for example, needs real-time voice APIs that allow for “barging in,” natural human interruptions where someone jumps in mid-sentence. The system also needs to minimize latency, or the time between an input and the AI's response.

Memory components that allow agents to adapt to individual users’ behaviors over time can enable personalization but are especially complex. For instance, an AI moderator could research interviewees beforehand, tailor questions to their background, and then remember those insights for future interactions.

Finally, there's the need for greater precision. E2B, another of Decibel’s portfolio companies, addresses this challenge with code sandboxes that let agents write and run code in secure environments. Instead of relying on a probabilistic model to approximate an answer, the sandbox lets agents compute exact results by running code.

And looming over everything: cybersecurity. Leão says that as autonomous agents begin working inside of enterprises as if they were employees, issues of privacy around data security will come to the fore.

Leão at an investment panel

What makes companies win — and what founders miss today

What do founders miss when pitching enterprise buyers? Demonstrable ROI, Leão says.

"Every vendor in history has said they'll make you 'more productive.' You'd be amazed how often founders don't actually quantify that."

So, Leão urges founders to consider: How much are you going to save that enterprise? How long is it going to take to deploy that application? Do these gains come with the off-the-shelf product or will someone need to be on site to customize the solution? Do you need to train folks? Do you require the keys to the kingdom to get started?

Even infrastructure tools built for AI site reliability engineers, which often require deep integration into production systems, need to demonstrate ROI from day one. "Show you're thoughtful with a commercial mind," she says. "If I'm a vendor, I want to know that you've thought through how to create ROI."

Another important differentiator is if the product itself is capturing valuable data. “In a world where internet data is becoming increasingly recursive with AI-generated content, and foundation model companies are entering AI applications, data becomes a competitive advantage,” Leão says.

She encourages companies to build a compounding data asset, systems that automatically capture telemetry every time users interact, approve suggestions, reject outputs, or rerun processes. That data should feed back into the product and underlying models continuously, not sit in evaluation dashboards you check once a month.

To illustrate, she contrasts the quality of a ChatGPT essay generated without any context versus one generated with a tailored prompt, instructions on user voice, and examples. To her, the difference in output is night and day.

“The more that you can train the system on the data of the user and their actions, the more powerful it becomes,” she says. “And the more the system differentiates itself from competition and the threat of foundation models coming to eat your lunch.”

What's hot right now

In recent months, Leão has observed two big themes in AI: reinforcement learning and inference optimization.

Reinforcement learning — letting models learn from their mistakes and successes as they complete tasks — has become more important as models improve at multi-step reasoning. "There's a shift from pre-training (model development) to post-training (model improvement), because scaling alone no longer delivers big capability jumps. Iterative feedback methods are where improvements increasingly come from," she explains.

And as models learn from ongoing feedback, they need rigorous ways to be tested. Leão notes that teams increasingly rely on simulation environments to validate that an agent will reliably behave as expected before it’s deployed into real systems.

Another key theme Leão observes is inference optimization, or the economics of improving the cost and latency of a machine learning model.

“Early on, everyone just turned on the OpenAI API or Gemini and ran with one model,” she says. “Now as workloads get complicated and startups think about margins, inference optimization naturally pushes them to ask: What model is better for what task? How do I adopt open source models like Kimi K2 or Qwen3?”

Beyond software, Leão keeps a close eye on AI applications in the physical world in areas such as biology, materials discovery, and robotics. These were previously considered capex-heavy, deep-tech domains. Now that AI models have improved, she sees real opportunity at the frontier.

“The next decade will see innovation in this area, and AI will allow us to build there.”

Advice for builders

“In many ways, it's easier than ever to start a company because AI makes it so much easier to build,” Leão says. So, how do companies stand out today?

• Ship fast and innovate relentlessly: With AI shifting rapidly and model launches every week, being even two months behind puts you far from state of the art. Her advice: iterate constantly, keep your product modular, and frequently launch new features to compete.
• Stay up to date: Leão suggests that, at minimum, builders should be able to answer these questions by week's end: What are OpenAI, Anthropic, and Google launching? What's happening in geopolitics and Chinese open-source models? What are the new startups, big announcements, or latest in frontier areas like AI in science or robotics?
• Use AI: “Try having Claude or ChatGPT build you a personal website or application,” Leão says. “You'll encounter questions like: Do I need a database? Is this 'batteries included,’ meaning does it provide everything from hosting to data storage, or do I need to piece together different services?" Use voice mode, custom prompts, custom GPTs. Or, use AI to walk you through the process if you don’t know where to start.

At its core, AI is a helpful, wonderful, transformational technology, she says.

“Everyone should be using it.”

You can subscribe to Leão’s Substack, and follow her on LinkedIn and X.