Nami Baral, CEO of Niural, on global payroll for AI agents

We've covered payroll from the COVID-era rise of contractor payroll and EOR players like Deel ($1B+ ARR in 2025), Remote & Oyster to the surge in contractor payroll with Wingspan's Anthony Mironov.

To learn how AI agents are changing global payroll, we reached out to Nami Baral, founder & CEO of Niural, who previously built AI agents for consumer financial negotiation at Harvest (acquired by Acorns).

Key points via Sacra AI:

To start, what inspired you to start this company?

When I started Niural, a lot of people asked me about this, since in my previous company I had built one of the original AI super agents in consumer financial negotiation and sold it to Acorns. I'd been in the agentic AI space since 2018, which is a very long time before ChatGPT launched in late 2022, and Niural launched right around then. So people would ask, you were in this hot, exciting AI agents ecosystem, why did you decide to build payroll instead?

My answer to that is that I didn't set out to build a payroll company. Niural is named after neural networks, and the idea was to build a unified intelligence and a true execution layer for the office of the CFO, essentially a neural network for the office of the CFO. Payroll happens to be one of the most important, sensitive, and complex fields for AI agents to get right, and it's one of the harder ones. Anyone who has built AI agents knows how difficult it generally is to get accuracy out of them.

If you truly want AI to become powerful in a way that increases productivity and meaningfully changes our lives beyond just answering questions, it needs to be able to automate away the most complex, boring, and unglamorous parts of what we do every day. For any company, payroll is one of those critical, high stakes domains, and at the same time one of the most unglamorous and antiquated. Within the office of the CFO, plenty of other areas were getting attention, but payroll wasn't getting much attention from an AI perspective. So my reasoning for starting Niural wasn't that payroll is a processing problem, it's an infrastructure problem. If you solve payroll from an AI infrastructure perspective, you can then solve a lot of other issues for the office of the CFO, and that's the larger goal Niural has been built to serve. Payroll was just a really interesting proving ground that let us build the foundation the right way.

Niural was founded in 2022 after Deel, Remote, and Oyster. What did you believe the market still misunderstood about the opportunity?

A lot of the growth in startups offering international hiring products during COVID happened because there weren't many domestic players that could provide that offering. The thesis in 2020 and 2021 was that companies needed to do international hiring, since everyone was moving back home or elsewhere, and needed to retain talent while also being able to hire anywhere in the world. Deel, Remote, and Oyster proved out the thesis that international hiring could be made easy, which was an important problem in the COVID era. The opening for that thesis existed because none of the domestic players had that capability.

I came into this after having started and scaled companies of my own. When you think about your workforce as a company, you don't think of it as having a PEO employee here, an EOR employee there, a 1099 contractor here, and a W-2 contractor in another country, you think of them as one single unit, your workforce. All of them, whether employees, contractors, or EOR employees, experience themselves as part of one company. But what's happened is that there's so much fragmentation in this ecosystem that the same company, to take care of one workforce, has to use a payroll vendor in the US, maybe a different vendor for benefits, a service like Deel or Remote if they have any EOR employees, and yet another vendor if they have international bill payments. That level of fragmentation is not normal in the age of AI. It's something people were forced to accept because nobody was providing a unified, intelligent solution. Unification alone isn't enough either, if you unify multiple components but still can't execute intelligently, that's still bad.

So the opportunity I saw wasn't about building for 2022, it was about building for the next decade and the decade after that: what does a new age payroll company look like in the age of AI? That has to be treated fundamentally differently from just building a payroll company. I think of it as building an intelligence company where the foundation has to be infrastructure, and payroll is a natural place to start because it handles your employees' payments and money movement, along with everything adjacent to payroll: benefits, 401(k)s, insurance, compliance. All of that falls under the umbrella of payroll, which is why it's one of the most critical workflows to solve first.

The opportunity is really about rethinking the current norm. Currently, when companies say "global," they often don't have a US presence at all, they outsource that to third parties, and the ones that do have a US presence rely on a network of international aggregators to offer the international version of the product. We wanted one comprehensive system of record that could convert into a true system of intelligence, so companies could experience their workforce as one unit and adapt as AI matures and the workforce itself becomes a mix of agents and humans. All of the added complexity around entities, jurisdictions, and having to use multiple vendors for different things needs to go away, because there's a whole other layer of complexity coming from how the workforce itself is structured around humans and non-humans.

Who is the core Niural customer today and why do they choose Niural?

Our product today is the only one that can handle all kinds of employment models and all kinds of money movement in one platform. That lets us serve many different kinds of clients at different stages of their life cycle.

Here's a direct example: if you go with a PEO only product in the US, the TriNets, Insperitys, and Justworks of the world, they can win a customer at a point in their life cycle where they need payroll plus benefits, and serve them domestically in the US. But the moment that client outgrows the PEO model, into something that requires an in-house ASO as well as the ability to broker benefits under their own entity rather than a pooled PEO entity, those providers have to turn the client away, because they don't have the infrastructure to accommodate that. It's the same on the earlier end: companies like Gusto serve micro SMBs. When you're really tiny, in one or two states, you can handle that complexity with just the ASO model, which is what Gusto serves. But once that company starts growing and needs access to better benefits, it needs a proper, modern PEO product.

So when you look at how companies use these vendors, they're constantly turning from one vendor to another at different stages of their life cycle. Companies with one product, or that support only a limited set of employment models, have a single wedge, so they go after a specific kind of customer: Gusto focuses on micro SMBs, for example. Similarly, on the Deel side, they look for companies that need international presence or are hiring a lot of contractors or EOR employees.

With Niural, we can meet our customers at whatever pain point they're having at that particular moment. Rather than saying we serve mid-market or SMB, we focus on solving their immediate need. High growth companies experience distribution in their workforce at some point in their life. It might be a company that comes to us after they've graduated off of an ASO player like Gusto and wants a PEO that grows with them. It might be a customer who has outgrown a PEO and needs an enterprise ASO solution. Once they come to Niural, they do not have to graduate off of Niural.

So the core Niural customer is defined more by their growth and increasing complexity than by any particular company size or stage.

Why do most companies in this space find it so challenging to expand outside their original wedge? Is your advantage partly about the post-LLM era, where it's become easier to stitch systems together and generate code faster, or is it more that there was never much reason to do this before, because there was so little global payroll happening?

AI has a lot to do with it. Part of it is that when you start building a product, you have to be intentional from day one about what you truly want to build. If you go in wanting to build a payroll product alone, you end up doing what plenty of YC companies each batch have done: integrate someone else's payroll API, call yourselves a payroll company, and go to market in two or three months.

At Niural, we built it with a very different intentionality. We didn't want a point solution, we wanted to be the intelligence and execution layer for the entire workforce. That meant we knew from day one that we had to build a system that could proprietarily support different employment models. It wasn't something we could take to market in two or three months, it took us several years. We drew on a lot of learnings from having built AI products before, which helped us build a genuinely robust stack and increase product momentum, including being able to build our own tax engines.

We had to build out all of these different models: a PEO is a different tenant model than an ASO, for instance. We also had to secure all the licenses we needed and the institutional partnerships, like our recently launched master plans with Aetna. A company like Aetna isn't going to hand out master plans easily unless they genuinely believe in the trajectory, substance, and infrastructure of the product they're looking at. So we spent a lot of time building the product because we knew that's what the next decade will require. The previous decade was about modernizing the front-end interface for slightly faster payroll. Justworks' claim to fame was having a modern interface, at least compared to the old school PEOs that existed at the time. But clients will increasingly reject having to use multiple vendors, multiple reports, and reconciliation processes, and having to close their books late every month because they have to pull information together from multiple systems.

I keep coming back to this mindset: most clients don't think of themselves as needing a PEO employee here or a US EOR employee there, they think of themselves as needing an employee to do a job, as one unit, one company. That's what the product needs to reflect. The benefit of the Niural team building a product like this is that we didn't see it as an insurmountable problem, because we'd already done much more complex things in other domains, like debt negotiations with 99% accuracy using AI agents back in 2018, before off the shelf LLMs even existed to build on. We went through a long journey of understanding what AI is capable of, building and perfecting an orchestration layer that could deploy thousands upon thousands of agents handling extremely nuanced problems, with a superagent orchestrating all of it. That's the model we leveraged to build a platform that's both as broad and as deep as any payroll company can get. AI native is in our DNA, and it's what we want to offer our customers as well.

Payroll is very operationally intense and typically involves a lot of humans. How do you use AI there, specifically, how much of it do you think of as working behind the scenes versus how much is exposed to the user, letting them do new things?

That's a good question, because payroll is definitely not an industry that adopts AI quickly. When we first started Niural, we had to build both a native AI version and a traditional dashboard version for customers, because at the time people were a little apprehensive and didn't really know what AI native meant. That entire dynamic has changed the last six months. I'd attribute a lot of it to Anthropic releasing Sonnet. People started using AI so accurately and genuinely across other parts of their lives that they began to trust it more and more. These days, the default conversation with clients isn't "is there AI," it's "I love that the AI can do this, can it also do that." That shift has been great for Niural, since we've been waiting for our moment in the AI spotlight and it has finally arrived.

AI is present across the entire Niural system, front end, back end, everywhere. People tend to think a conversational interface is how you experience AI, but there's also this idea of ambient AI: in everything you're doing, AI should understand the context, understand how to make that same job or workflow faster next time, and ultimately be able to automate the entire thing without you having to specify exactly what to do and what not to do. It takes a few runs, and it understands the full context of what you're doing, whether that's enrolling in benefits, drafting an offer letter, or figuring out compliance around an international termination. Those are all critical workflows that customers previously assumed AI would never be able to solve, and now they welcome AI executing across all of these fronts, because they've seen the proof that it can not only work, but work extremely accurately, and save them a lot of time and headspace.

AI startups look to grow revenue with far fewer employees, while Niural largely monetizes per worker. Is AI-driven headcount compression a long-term headwind, or does increased company formation and global complexity more than offset it? Is payroll for agents something you're thinking about?

There's going to be a world where a lot of the mundane, day-to-day analysis work can be entirely handled by agents. Niural's agents are unusually powerful, which is rare in this industry, even across frontier labs. Most AI agents excel at generic tasks and are most helpful where the need for accuracy isn't that high, which is why the best AI agent benchmarks tend to involve things like QA. Look at the SWE-bench benchmark that Devin initially started with: Devin's initial unassisted score on the SWE-bench benchmark was **13.86%**. Compare that to what the payroll industry requires. It's not okay for us to be right only 99% of the time, we have to be right 99.9999%, what we call “six nines” at Niural. You need that level of accuracy not just to tell people how to run payroll, but to actually run the payments associated with payroll. Once the money goes out, that's the most deterministic flow you'll ever encounter, and you have to solve it with probabilistic models.

That's where Niural has been a trailblazer: we understand high stakes domains and agentic orchestration in high stakes domains extremely well. Other payroll companies wrap a chatbot and call it an AI agent. That's not what we consider true AI agents. What we mean by AI agents is full end-to-end execution of multi-jurisdiction payroll: understanding tax laws, compliance, time off, expenses, time sheets, benefits, workers' comp, and all of the other components that matter for protecting workers, then actually moving money through the proper rails and getting it to employees on time.

You may have heard of long horizon agents. Everyone is trying to build them, because the actions that really matter aren't one-click actions, they are actions that have consequences across multiple time horizons. Payroll is one of those. You run payroll today, and there are tax liabilities due next week, the week after, a quarter later, at year end when you have to issue tax documentation, and so on. Those long horizon actions are very difficult to build from a contextual memory standpoint for AI agents.

That's where we've shined, being able to build something that reaches that level of accuracy, which is why we recently launched Niural AI Labs. In order to handle this for payroll at this level of accuracy across this many employment models and countries, we had to build a lot of critical AI infrastructure: the right environments, the right verifiers, the right kind of benchmarking that lets us do both AI verification and human expert-led verification. That system is powerful enough that we think it's useful well beyond payroll.

So our goal is to make those learnings available so we can extend this kind of AI productivity into additional high stakes domains. Insurance needs to be disrupted, payments need to be disrupted. Writing code or a better email are useful applications of AI, but ultimately, to make the entire human population meaningfully more productive, we need to reach much higher ground in terms of what AI agents can do. That's the goal Niural AI wants to push forward through Labs.

Tell us about Niural AI Labs. What is its mandate? Is it focused primarily on training or post-training models or on more applied work around solving concrete problems on the business side? How do you think about differentiating an AI lab that's a serious research lab from one that's more for marketing or recruiting purposes?

That's one of my pet peeves, honestly. Whenever a new technology comes to market, everyone tries to jump on the bandwagon. It's very difficult to actually perfect that technology over years and build something people genuinely use. We could have launched the AI Labs component many years ago, since all of our infrastructure is entirely AI native, but we wanted to prove it out first in a high stakes domain we control ourselves. That way people can see that this infrastructure isn't just something in theory, it's already been used in an industry where you cannot afford to get it wrong, an industry that requires six nines of accuracy. If you can deploy agents there, what you have is genuinely substantial. We wanted the proof to exist before the launch.

We've proven it out on payroll, and now we're bringing it to other domains. We don't worry too much about how others position themselves as AI native or not, because we feel confident enough in the infrastructure we've built, both the application layer and the infrastructure layer, that we can walk our own path on our own timeline.

Is the focus more on application layer products, or more on post-training the models themselves currently?

Environment building, verifiers, and reinforcement learning training are already part of the overall architecture we've built to deploy AI agents ourselves. Our goal is to make sure the world knows about the learnings that are specific to high stakes domains, because Niural isn't a general purpose lab, we don't want to own everything. We want to bring our learnings out into the world so other builders can apply them across additional domains.

That said, there are certain domains we will control: we already control payroll, benefits, compliance, and payments, and our overarching goal is to become the neural network for the office of the CFO. Anything pertinent to the office of the CFO, we'll build both the infrastructure through AI Labs and an applied AI function in the form of actual applications. But we want to make the infrastructure itself available to others so they can build on it too.

If everything goes right for Niural over the next five years, what does it become, and how is the world different?

If Niural does everything right, the future we're building for isn't just better payroll. That theme has probably been clear from the beginning of this conversation, because it's never just been about payroll. It's about making sure companies can evolve over the next decade, as the workforce archetype changes drastically. It won't just be humans in one place, on one surface, anymore.

It'll be humans distributed across multiple different places, where you don't care where they're based, what matters is that they can rise to the level of productivity the ecosystem demands. That means creating a world where you can take care of humans really well, not just their payroll, but their healthcare, their benefits, their insurance, in a way that ultimately allows a different generation of businesses to exist. The next generation of businesses is going to want a payroll company that can coordinate not just people, but compliance across agents, humans, and jurisdictions, and money, with much greater intelligence overall.

The old guards in this industry have reduced payroll to a mere processing problem for decades. Payroll is not a processing problem, it's the ultimate infrastructure for how you manage people, how you care for them, and how you move money. We want to do all of that with enough intelligence that a new category of companies can exist.

This transcript is for information purposes only and does not constitute advice of any type or trade recommendation and should not form the basis of any investment decision. Sacra accepts no liability for the transcript or for any errors, omissions or inaccuracies in respect of it. The views of the experts expressed in the transcript are those of the experts and they are not endorsed by, nor do they represent the opinion of Sacra. Sacra reserves all copyright, intellectual property rights in the transcript. Any modification, copying, displaying, distributing, transmitting, publishing, licensing, creating derivative works from, or selling any transcript is strictly prohibited.