Ayush Sharma, Founder and CEO of Warp – Interview Series

Ayush Sharma is the Founder and CEO of Warp, which he launched in 2023 after experiencing the complexity of managing payroll across multiple U.S. states and countries. Raised in a small town in India, he studied computer science and physics at MIT. His master’s research focused on mitigating slow-running components in distributed machine learning training, exploring how large-scale systems can operate more efficiently when individual machines fall behind. Warp joined Y Combinator’s Winter 2023 batch with an initial focus on payroll tax compliance. Sharma’s approach emphasizes AI agents that execute complete workflows, with permissions, human approvals, and auditability incorporated from the outset.

Warp is an AI-native employee management platform that combines payroll, HR, tax compliance, benefits, and IT around a shared employee record. The platform uses agents to handle tasks such as state tax registration, payroll checks, tax-notice resolution, benefits enrollment, and employee onboarding. Companies can set autonomy separately for each workflow, choosing whether an agent recommends an action, waits for approval, or acts on its own. Agent activity is recorded for review, and tasks requiring human involvement can be routed to company administrators or Warp specialists. Warp also supports employee device and account provisioning, global contractor payments, and dedicated account-management and benefits support.

Your path from growing up in a small town in India to studying computer science and physics at MIT and later working on machine learning infrastructure at Yelp is fairly unconventional. What ultimately led you to found Warp in 2023, and why did you decide that payroll and employee management were the problems you wanted to rebuild around AI?

Before Warp, I worked at a startup that grew from about 10 to 40 people during COVID, hiring remotely across states, and I saw how much of payroll and compliance was still done by hand. Later, when I was hiring my first people on my own, I lost an entire day figuring out how to register with the New York Department of Labor, with a fine waiting if I missed the deadline. With fires everywhere else in the company, that day stuck with me.

HR teams deal with procedural, repetitive work every day, and it eats into the time they should be spending on people. It’s the perfect place to apply AI, but every major HR platform was built before agents existed, and you can’t easily add that capability onto old software.

When we sat down for our YC interview in 2023, we didn’t have a finished product or any revenue, just a problem we couldn’t stop thinking about. Five people from YC were on the call, and for the first ten or twelve minutes they pushed on one question: could payroll tax compliance really be a big enough wedge into everything else? It was a hypothesis, but the reasoning held up. Companies were going multi-state far earlier than they used to, so the complexity only climbed. And payroll tax is harsher than sales tax, which gives you generous thresholds before the rules kick in. With payroll, hiring one person in one state puts you under that state’s entire tax apparatus.

That’s why compliance was the obvious place to start. I’d spent years at MIT on machine learning, and when LLMs and agents started becoming real, I could see that the work nobody wanted to touch in payroll and compliance was about to become possible. The first version of Warp did one thing: you could hire someone in any state and never touch government paperwork. We figured if we could get payroll tax right, the rest would be easier.

You’ve argued that companies such as ADP and Workday face a fundamental disadvantage because their systems were designed long before modern AI. At a technical level, what does an AI-native architecture look like for HR, and what can Warp do that would be difficult to achieve by simply adding AI agents on top of a legacy platform?

Most HR software was designed for a person to sit at a screen and click through each step, so when those platforms add AI, it usually shows up as an AI agent that tells you what to do next while you’re still the one doing the work. We built Warp’s architecture the other way around, designing the API, the permissions, and the audit trail from the start so that an agent could handle the work and people could step in when they’re needed.

Say you’re the Head of People at a 350-person company, and four offers get signed, two in Oregon and two in Colorado, in states where the company has never had an employee. Before anyone can get paid, you have to register with two state revenue departments, open unemployment accounts, sign up for each state’s paid leave program, work out the local taxes for every address, and set up deductions. A typical AI assistant can tell you which forms you need, but our agent fills them out and files them, and if you’ve given it permission to act on its own, you hear about it once everything is done.

Our agents have the context, permissions, and infrastructure to do multiple steps across a company’s systems, from running payroll and managing benefits to handling tax compliance, without someone manually managing it all. Adding an agent to a legacy platform can help users navigate software, but it doesn’t change how the underlying work gets done. Warp’s agents actually change how the work is done.

Warp describes its agents as owning entire workflows, including opening state tax accounts, filing returns, resolving government notices, and managing benefits. How do you determine which actions an agent can execute autonomously and which should require human approval?

Warp sets the outer limits, and each company decides how much autonomy its agents get inside them. Some actions, like payments, carry firm limits that no customer setting can remove. Within those guardrails, companies set autonomy workflow by workflow, so the agent might only suggest an action in one area, wait for approval in another, and handle things on its own somewhere else. Most customers start by approving everything and expand autonomy as they gain confidence in the agent’s work. When a decision belongs to the customer, the agent sends it to their admin. When a task needs a person, either because the law requires one or because an agency won’t work with software, it goes to a Warp specialist.

Payroll and tax compliance leave very little room for hallucinations or probabilistic answers. Warp operates across thousands of tax jurisdictions and changing regulations. How do you ground your AI systems in authoritative information, measure reliability, and handle situations where an agent is uncertain about the correct action?

We’ve spent three years testing our controls on payroll tax compliance, which is the most unforgiving area of HR. Warp tracks more than 10,000 tax jurisdictions, and the agent never answers a tax question from memory. Every rate, form, and deadline it acts on comes from the agencies themselves, and our tax team reviews each rule change before it reaches the system. We’re on track to process over $2 billion in payroll this year, and our customers have avoided more than $200M in penalties along the way.

Before each payroll run, the agent checks for anything that looks off, like an unexpected jump in someone’s pay or a missing tax setup in a new state. If it isn’t confident about what to do, or doesn’t have permission to act, it stops and hands the task off to the admin instead of guessing.

Tax notices show how this works. Agents resolve 80% of them automatically, working from the company’s payroll data and the agency’s own guidelines, and the rest go to a Warp tax specialist, so every notice gets resolved. In payroll, a wrong answer costs a lot, so we’d always rather the agent hand off than guess.

HR systems contain some of a company’s most sensitive information, from salaries and Social Security numbers to banking and benefits data. How does the emergence of autonomous agents change the security model for this data, and are there categories of employee information you believe AI should never be given autonomous control over?

Agents change the security question because they can move through a company’s systems much faster than any person, which means a gap in permissions gets exploited faster too. The biggest risk is someone using an agent to reach data they could never access on their own. We designed around that from the start, so every Warp agent inherits the exact permissions of the person it’s acting for. So if an employee can’t see compensation data, they can’t ask the agent to show it to them either. Individual employee data stays inside that company’s own account and is never shared with other customers. Warp is SOC 2 Type II compliant, enforces multi-factor authentication, and signs Business Associate Agreements for Enterprise customers who need HIPAA compliance.

As for what AI should never control on its own, the line for us is anything hard to undo. Moving money is the clearest example, so payments carry firm limits the agent can’t exceed, no matter who is asking. The same goes for any task where the law requires a person, which always goes to a Warp specialist.

There is a significant difference between an AI recommending an action and an AI actually moving money, filing a tax return, or changing an employee’s access privileges. What safeguards, audit trails, rollback mechanisms, and escalation systems become necessary when AI agents move from advising humans to executing consequential business processes?

The moment an agent starts taking action instead of making suggestions, you need to be able to answer who it was acting for, what it was allowed to touch, why it made the choice it did, and how to reverse it if something goes wrong. Every action on Warp carries that information, so nothing the agent does happens in the dark.

Most of the agent’s work runs through routines, which are jobs the agent runs on a schedule or when something happens, like a new hire signing an offer. Each routine keeps a running record of who created it, when it was approved, and every step the agent took, which gives HR teams and auditors a clear trail to follow.

How we reverse an action depends on what it is. Changes like updating someone’s access or a benefits election carry a rollback, so they can be undone. Payroll runs and payments are the hardest to reverse, so they get the strictest controls, including firm limits the agent can’t change. Anything outside the agent’s authority goes straight to the customer’s admin for a decision.

Some situations still need a person, like a state tax office that will only talk to you over the phone. When the agent hits one of those walls, a Warp specialist finishes the job, so the work keeps moving instead of stalling in someone’s inbox.

Warp recently raised a $60 million Series B, bringing total funding to $85 million, with plans to invest further in AI agents and its tax and compliance infrastructure. Which technical problems around agentic HR automation are still unsolved, and where do you expect the greatest portion of that investment to go?

An AI agent might make someone 20% faster, which helps, but the bigger opportunity is an agent that finishes the job. The hard part is that HR doesn’t live in one place. Payroll, benefits, devices, and IT accounts usually sit in separate systems, and an agent can only act well if it can see all of them together. The other challenge is the long tail of exceptions, since every state agency, carrier, and tax office has its own quirks, and plenty of them were never built to work with software at all. That’s where we see the most opportunity and where our funding is going.

Our customers often have one or two people managing HR for hundreds of employees, so agents need to take on more responsibility as their company grows. We’re focused on bringing everything into a single system the agent can act on, and making it easier for customers to write their own routines around how their company works. We’re also growing the team that supports customers, strengthening our tax and compliance infrastructure, and expanding what our agents can handle across the employee lifecycle.

One of the promises of agentic software is that companies can operate with much smaller administrative teams, while HR professionals spend more time on strategic and people-focused work. Which parts of HR do you think AI agents will largely automate over the next few years, and which decisions will remain fundamentally human?

The parts of HR that depend on judgment will stay with people, whether that’s choosing who to hire, shaping the culture of a company, supporting someone through a difficult moment, or recognizing when an employee is ready for something bigger. Those are the reasons People teams exist, and they’re the parts of the job most HR leaders wish they had more time for.

What agents can take over is the administrative work that quietly fills most of the week, from reconciling payroll line by line and chasing down tax agencies to answering PTO questions, handling benefits enrollment, and setting up or shutting off laptops and accounts every time someone joins or leaves. At a large company, that work keeps an entire department busy, and at a small one, it tends to land on the founder late at night after everyone else has gone home.

Onboarding is a good example of how much time this frees up. By our estimate, fast-growing companies typically spend 10 to 20 hours managing each new hire during their first month, and our customers tell us that same work now takes less than two hours, which leaves HR teams with far more time for the people side of the job.

Looking five years ahead, do you expect HR software to still revolve around dashboards, forms, and workflows that employees actively interact with, or does the interface eventually disappear as AI agents handle most employee-management operations in the background?

Within the next year, we expect agents to be taking most of the actions on Warp, well beyond tax. That includes onboarding new hires, running payroll, and the everyday HR tasks people still click through today. That changes what people need from the software they use. Instead of logging in to check on a process or move it along, HR teams will mostly hear from Warp when something needs their attention, like a decision only they can make or a quick update that a new hire in another state is fully set up and ready to be paid. The rest of the work will happen quietly in the background, the way most people already expect their email to get delivered or their bank to process a payment without watching every step.

Nearly every corner of enterprise software already has a new challenger built for AI, and employee management is one of the last areas still waiting for that shift. We believe we’re building what comes after Workday.

Thank you for the great interview, readers who wish to learn more should visit Warp.