Notes from NY Tech Week: What Anish Acharya and Shuo Wang Taught Us About Building AI Startups

We spent NY Tech Week doing the thing the week is actually for: sitting in rooms with people who have built the things we're trying to build, and stealing their hard-won lessons. Two sessions stuck with us enough to write down. One was Deel&a16z Masterclass: Founder's guide to building one of the world's fastest growing companies #NYTechWeek with Anish Acharya (General Partner at Andreessen Horowitz and a Deel board member) in conversation with Shuo Wang (co-founder and Chief Revenue Officer of Deel, the global payroll and HR platform now valued in the tens of billions). The other was IBM session on AI and data sovereignty with Vice Chairman Gary Cohn and investor Anthony Scaramucci, moderated by Stephanie Ruhle.

We were lucky enough to connect with Anish and Shuo directly, and we're releasing the a16z session as an episode on our YouTube channel. This piece is the companion: not a highlight reel, but the handful of ideas we think matter most for anyone building AI inside a regulated, high-stakes field — which is exactly where biomedical research lives.

Why we cared about a payroll company's advice on building AI startups

On paper, a global payroll platform and a biomedical AI company have nothing in common. In practice, Deel spent six years solving the problem we wake up thinking about: how do you build software fast in a domain where getting it wrong has real consequences? Pay someone late or misclassify them in the wrong country and you don't get a bug report — you get a compliance violation. Tell a researcher their non-compliant assay is fine and you don't get a bad review — you get a failed submission, or worse. The shape of the problem is the same. That's why we were taking notes.

Lesson 1: The defensible ground is where liability lives

The sharpest idea of the week came when the conversation turned to whether foundation models will simply swallow every software company. Acharya pushed back on the doom narrative, and his reasoning is worth quoting:

"There is no reason for the models to get interested in areas that are regulated or involve liability, because an agent or model can't be liable for a decision."

Sit with that for a second, because it inverts how most people talk about AI risk. The usual worry is "the model will get good enough to do my job." Acharya's point is that in regulated domains, raw capability was never the binding constraint. A model can be brilliant and still be unable to do the one thing the work requires: stand behind a decision and be accountable when it's wrong. Deel's moat, he argued, isn't that its software is cleverer than a frontier model — it's that the company has absorbed the regulatory and liability burden of paying people in 140-plus countries, the part no general-purpose model wants to own.

Think of it like the difference between a calculator and an auditor. A calculator can be flawless at arithmetic and still can't sign your financial statements. The value of the auditor isn't the math — it's the signature, the accountability, the willingness to be held responsible. In regulated fields, the signature is the product.

This is the ground CLYTE deliberately stands on. A general LLM can describe an FDA guidance to you. What it won't do is take responsibility for whether your specific submission clears it — and in a regulatory setting, an AI that confidently tells you what you want to hear is not an asset, it's a hazard. (We ran a test on exactly this failure mode; it's the subject of a separate article on why Sophie refused to approve a non-compliant submission that three other major AIs waved through.) The lesson we carried out of the room: don't build on the sand where models compete on cleverness. Build where the work demands accountability, structure, and a defensible audit trail.

Lesson 2: Stop assisting people. Automate the whole job.

Acharya described the most important shift he watched happen inside Deel as a move from improving individual productivity to fully automating business functions. His framing:

"You really have to be able to do one hundred percent of something, which is dramatically harder than doing ninety percent — but that's how you get the real unlock."

Shuo Wang made it concrete. Running payroll used to mean a human uploading paperwork, logging into a government portal, verifying payments, downloading and re-uploading files — slow, manual, error-prone work. Deel rebuilt it as an agentic workflow: agents that pull the payment data, verify it against the government system, run the calculations, and deliver the payslips end to end. The result she cited was roughly a thousand operational hours saved per month, with volume growing while headcount stayed flat.

The "90% versus 100%" distinction is the part worth internalizing, and it's counterintuitive. Getting an AI to do 90% of a task feels like almost done — but a workflow that still needs a human to catch the last 10% hasn't actually been automated; it's just been made faster, and someone still has to stay in the loop the whole time. The jump to 100% of a bounded task is what frees the human entirely. The trick isn't a smarter model doing more of everything; it's drawing the box tightly enough around a complete task that an agent can own all of it.

This is precisely why Sophie 4.0 is built as a system of specialized agents rather than one do-everything chatbot — separate engines for assay design, statistical analysis, visualization, and regulatory lookup, each owning a complete slice of the research workflow. Take scratch-assay analysis: instead of speeding up the manual ImageJ grind, the goal is to own the entire path from raw images to a migration rate, so the researcher is out of the pixel-counting business altogether. Deel's experience is the proof of concept from another industry — the unlock comes from completing whole jobs, not accelerating fragments of them.

Lesson 3: "Are you happy when the models get better?"

Acharya offered this as a litmus test every AI founder should be able to answer, and it's a good one:

"Are you happy or sad when the models get better?"

His warning was about a specific trap: if you build your AI startup by patching around something today's model is bad at, the next model release will leapfrog you and quietly delete your work. The businesses that survive are the ones that deliver more value to the customer as the underlying model improves, not less.

He had a generous way of describing what that looks like — being a "point of economic diffusion" for model progress into an industry that could never absorb it directly. His example was credit unions: they're never going to figure out how to take raw model intelligence and turn it into the things they actually want to do, so the opportunity is to be the company that translates frontier capability into their world and "helps them dream the dream."

That reframed something for us. A research lab, like a credit union, isn't going to wire raw model intelligence into a reproducible, regulation-aware workflow on its own. The job of a domain platform is to be the translation layer — to take each generation of better models and turn it into something a biologist can actually use at the bench, so that better models make the tool better, not obsolete. We left genuinely happy when the models get better. That's the side of the bet you want to be on.

Lesson 4: Trust is the real product — and you earn it the slow way

Asked what business Deel is really in, Wang's answer wasn't "payroll" or "HR." It was trust:

"We're actually in the business of trust, because we pay people. If they don't get paid by the end of the month, it could be a disaster for them."

Both speakers kept circling back to this, and to a point that surprised us coming from a room full of AI optimists: in an era when AI can fake almost anything, in-person relationships and a real track record matter more, not less. Wang noted that Deel still invests heavily in face-to-face events and being "a real company you can see and trust." Acharya added that in the early days your reputation is your brand — "doing good things for a small number of people tends to compound."

For a company asking scientists to route their experiments and regulatory questions through an AI, this is the whole game. Trust in our world is concrete: we don't train on your data or share it, and we offer air-gapped deployment for institutions that need their data locked down. But the deeper lesson is that those features are necessary, not sufficient. Trust compounds from being right, repeatedly, for researchers who stake real work on the answer — the same slow way Deel earned it.

Lesson 5: While the door is open, walk through it

One practical, almost tactical note for early-stage founders. On the endless "what's your moat?" debate, Acharya was blunt: right now, distribution beats moat.

"It's more important to sell now than to have some intellectual concept of a moat... while the door is open, sell them something that has value, and compound from there."

His argument: enterprises that were impossible to reach three years ago now want to buy AI, because their boards are demanding an AI story. The window is open, and the move is to get in the door with something genuinely useful rather than wait for a perfect defensive position.

Wang added the version of this that hit closest to home. CLYTE's CEO Mojtaba Javid asked how Deel grew in the early days, and Shuo pointed to content and SEO — and, in 2026 terms, GEO (getting surfaced by AI answer engines):

"We did a lot of content on our website, and SEO helped us. Today you have GEO — you need to rank yourself, build content so that when people ask, you're there. It doesn't cost millions; it's maybe ten or twenty thousand dollars and you can drive really good traffic already."

We'll admit a bias here, because this is the strategy our blog is built on — but hearing a founder who scaled to a multi-billion-dollar company name content and being findable as an early growth engine was a useful reminder that the unglamorous work compounds.

The IBM counterpoint: sovereignty, and the case for small, specialized models

If the a16z session was about where to build, the IBM session was about how to hold onto what you build — specifically, your data. Gary Cohn made a prediction that cuts against the "one giant model to rule them all" story:

"There is a whole industry of small, unique models built to serve specific problems — that you can house internally, own in your own enterprise, and flood with your own data."

His larger point was that the cloud-versus-on-premise pendulum has swung back. When he arrived at IBM, the consensus was that everything would move to the cloud and mainframes were dead; instead, organizations got "acutely concerned about sovereignty of their data — where it goes, how I can protect it, how I can keep it from being used." The fear he named precisely: if I put my data into an AI system, can I stop that system from getting smarter off my data and handing my competitor a better answer?

For biomedical research — where unpublished data, proprietary compounds, and pre-submission regulatory strategy are the entire value of the work — that question isn't abstract. It's the reason CLYTE's architecture is built around specialized engines rather than a single general model, and why no-data-sharing and air-gapped options are core, not afterthoughts. Cohn was, in effect, describing the market conditions that make a domain-specific, data-sovereign research AI the right design rather than a compromise.

We'd add one honest caveat the panel itself raised: the same sovereignty technology that protects researchers also raises hard public-policy questions about privacy and data rights that nobody has fully solved yet. We don't think anyone in that room — or this one — has the complete answer. But the direction is clear, and it favors building specialized, walled-off, accountable systems over funneling everything into one model that sees all.

What we're taking back to the lab

Strip away the payroll specifics and the macro debate, and the two sessions rhyme into one thesis we already believed but can now articulate better:

• Build where decisions carry liability — that's the ground frontier models won't and shouldn't take, because they can't be accountable. (Acharya)

• Automate whole jobs, not fragments — the unlock is owning 100% of a bounded task. (Wang)

• Be the diffusion layer — stay on the side of the bet where better models make your tool better. (Acharya)

• Earn trust the slow way — features enable it; being repeatedly right compounds it. (Wang)

• Keep the data sovereign — the future is specialized, walled-off models, not one model that sees everything. (Cohn)

That's the company we're building, and it was clarifying to hear founders and operators from completely different industries describe the same map. Our thanks to Anish Acharya and Shuo Wang for the conversation, and to the organizers of both sessions.

The full a16z masterclass is now live as a podcast episode on the CLYTE YouTube channel — worth a listen if any of the above resonated. And if you want to see what "AI built for a regulated field" actually looks like in practice, you can try Sophie.

FAQ

What is NY Tech Week? NY Tech Week is a distributed, week-long series of founder, investor, and company-hosted sessions across New York City, organized largely around the startup and venture community. CLYTE attended several sessions in 2026, including the a16z founder masterclass and an IBM session on AI and data sovereignty.

Who are Anish Acharya and Shuo Wang? Anish Acharya is a General Partner at Andreessen Horowitz (a16z) and a board member at Deel. Shuo Wang is a co-founder and the Chief Revenue Officer of Deel, a global payroll and HR platform she co-founded after a background in robotics and hardware engineering.

Why would a biomedical AI company learn from a payroll company? Because both operate in domains where errors carry regulatory and liability consequences. Deel's experience building software fast in a compliance-heavy, high-trust field maps directly onto the challenge of building AI for FDA-regulated biomedical research.

What's the core takeaway for AI in regulated industries? That defensibility comes from owning the parts of the work that involve liability, accountability, and data sovereignty — the things a general-purpose model can't or won't take responsibility for — rather than from raw model capability alone.

Where can I listen to the session? The a16z masterclass is available as a podcast episode on the CLYTE YouTube channel.