Inside Avalara Aviator, the Agent Hub for Tax and Compliance

Hey everyone, welcome to our coverage of Avalara CRUSH. We are coming at you from the Omni Hotel, which is a fairly new hotel down here in downtown Fort Lauderdale. Thank you for joining us.

Kicking off our coverage with two key people from Avalara. Let me introduce you to, actually to my far left first, is Danny Fields. Danny, welcome.

Nice to be here. Thank you for having us. Thank you.

And then the man in the middle, Jayme Fishman. Hello, all. Nice to be here.

Jamie, nice to have you. Yeah. So guys, let's start...

Well, I got a lot to go over here. But we're going to start with Danny. Why don't we start with you.

Tell our audience your present title with Avalara, and give them a little sense of your journey. Okay. So I'm the Chief Technology and Customer Operations Officer at Avalara.

So I'm responsible for working with Jamie and his team to go build our products, the cool new products that hopefully you're all using. And also with support, making sure that if customers have questions or any issues, we're there to resolve those problems as quickly as possible. I've been in technology for about 32 years.

Started off in Ireland, hence the funny accent- ... the funny voice that you're hearing. But I worked in Oracle, a big enterprise company, for over 20 years.

Worked in security, and then joined Avalara about eight years ago. Eight years. Yeah.

Fantastic. Jamie, how about you? Yeah.

Jayme Fishman, I'm our Chief Strategy and Product Officer. Been with Avalara for about six years, where I run corporate strategy and the product organization, as well as M&A. And been in the industry for 25, 26 years, focusing on sales and use tax automation.

Which is- Really? pretty nerdy, but that's okay. Yeah, my very specialty.

Yeah. So guys, it begs the question, your Avalara aviators. Aviators, I think of either two movies.

One is "Officer and a Gentleman," right? Richard Gere. I've heard of that one.

Yeah. " The other one. The other one.

"Two things from Oklahoma," or what? That's a great movie. Yeah.

Or of course, the "Top Gun" franchise, right? Maverick, Goose, Iceman. Yeah, you might have noticed which one we picked.

Yeah. Yeah. I get it.

Yeah. " So there has to be more to this than the jackets, though. Tell us a little bit about how this was played out.

I think it was the looks, because they were looking for- Two good-looking guys ... two good-looking guys. Right.

Two Tom Cruise clones. Yeah. It's like he was cloned.

Yeah. You're reading my mind. Why don't you show your family that at home?

They're all going to get a kick. But- Yeah ... tell us- They're really- I hear you guys had a spectacular coming out on the stage today.

Yeah. Well, anytime you do a production with this guy, he goes all out. I don't know what his budget is for these events, but he shows up with a real motorcycle.

An orange motorcycle. An orange motorcycle. Well, because all these orange-themed things is our corporate color.

And so there's this AI video introduction, then he comes in on a motorcycle. Not driving it, mind you, because we're inside, and they frown on that. But then, they flash over to me, and I'm on the piano doing the Goose thing with the intro.

And all of this pomp and stance is to celebrate the coming out of our AI-powered agent hub that we're calling Aviator. And so the aviation theme- So there was the connection. Yeah.

It wasn't just someone's fantasy- No ... of being Tom Cruise. And we've had an AI agent called Avi for a few years now.

Really? And we were looking at a new name for our product. We looked at Avi, we looked at a whole bunch of other words.

Avi, Aviator, Aviator. Yeah. And it's because you've got this hub that's sort of like a cockpit, a command center, to get all your work done.

Uh-huh. And so we thought that's kind of a cool metaphor. Yeah.

I like that. Yeah. All right.

Let's run with it. Yeah. So I feel like we've got all that done.

But let's lay a little foundation here. A lot of our audience at home is saying, "Hey, I know Avalara. " Right?

On the run, almost probably everybody watching this video, whether they knew it or not, Avalara probably calculated the sales tax on most, if not all of their e-commerce- Probably true ... purchases, right? Yep.

And they're saying, "Okay, I get it. They're great. " What are they doing with AI?

Where to start? Yeah. Well, look, here's the thing.

For the past several decades, technology has been applied in our field. Yes. And it's been great.

It's been wonderful. It's automated 70, 80% of work that folks need to get done. Mm-hmm.

And that was the limit of technology. Because technology could connect systems, it could exchange data, it could apply business logic, and that's kind of where it ended. But now, with the advent of AI ushering in all these new and wonderful capabilities, we're able to field agents that close out the remaining work that customers are still responsible for.

And they can do it autonomously, but with human oversight to make sure that they're making business critical decisions- Sure ... one, right? So that's kind of the genesis of it.

And the technology that we've built to do all of that, it is so advanced. Behind the scenes, we can scale, we can run really fast, we're super secure. But now in this new world of AI, we're using nearly 100 of the best publicly available LLMs behind the scenes to power all of the agents that Jamie just mentioned.

And we've even trained some of our own proprietary LLMs, with the data and expertise that we've amassed over the past 20 years. So you're creating, would you really call it an LLM, or maybe more of an SLM, small language module? You're very good.

Yeah. So we're using public LLMs, and we've trained, you can call them SLMs, yes. So are you using RAG for that or something like this graph technology?

Let's get geeky. Come on. We're using both.

Oh, yeah, RAG. We're using RAG databases, and we're using graph technologies. All right.

Very cool. All the above. All the above.

And then in terms of switching among your different LLMs and SLMs, let's call it data sources and models and so forth. Yeah. Sure.

Is it like some sort of harness that you've put together here? He's good. Yeah.

Yeah. So we basically built a layer of capabilities on top of these LLMs. We call it Alpha.

And Alpha, one of the major components inside of Alpha is something called an LLM routing gateway- Sure ... that decides which LLM to go use. So, we'll choose the best LLM that's right for the task, and we'll also use the one that is optimized for cost.

So we'll optimize based on the work to be done and on the cost. And not many companies have this kind of technology. Some companies out there are just specializing in building this functionality, but we built it ourselves over the past two years.

So we get to choose which is the right LLM for the specific task. Love it. And we're switching all the time.

I love that. And what are you calling that router? It's an LLM router.

Just an LLM router. But it's part of our proprietary framework that controls all of our AI that we call Alpha. You know what?

Who knew that sales tax- ... right, would spawn this sort of AI- But what's- ... AI specialization?

But the reason why is because if you think about what's required to do all the sales tax stuff and compliance stuff behind the scenes, from an engineering point of view, it's fascinating. Because we have so many products, everything has to happen at scale. We mentioned in our keynote earlier today, in a year, we process over 54 billion tax calculations.

In a year, we file over seven million tax returns. 75 trillion. So to do all of that at scale, we have to have technology that runs at scale, which is why we need all of these LLMs.

We have to run them in the most efficient way, not do all of our work on the most expensive ones. Switch between the expensive ones and the cheaper ones, and which ones are right for the job. So we've had to develop all of this technology to do all of this work for our customers.

So it's a huge technological challenge. It is. With great engineers behind the scenes geeking out on this every day.

Got it. It's fantastic. Yeah.

So one of the big challenges with AI in general is governance, security, and governance. Yeah. When we're talking about things like taxes- Governance takes on a whole different meaning, right?

Yeah. How are you dealing with the governance issue here to make sure, because it's not just you, it's your clients- Yeah. Of course ...

who their taxes and their tax returns that are at stake here, and that's an awesome responsibility. What have you done to really make sure governance is handled correctly? Well, there's a lot of layers to governance.

Yeah. Absolutely. And part of what you're pointing out is that these are business-critical systems, and they need to be right.

So from an operational standpoint, we never confuse a task that's deterministic in nature with one that's a probabilistic kind of a technology. And so at the core, we will apply deterministic technologies when we need a calculation done and we need a sub-10 millisecond response because of the nature of all of our integrations and those 54 billion transactions are coming in at super speed. So we're handling all those things that way, but we're augmenting them with responsible AI that's helping customers onboard easier, get their questions answered, get configurations, get mappings, build integrations, get their reports, get their data, all of these things kind of surround.

And we're also having responsible AI governance in the sense that we've got governance policies. We don't train our models on customer data. We're badged in some of our partner ecosystems for responsible AI governance and all those kinds of things.

Yeah. So we have it at the technological level as far as making sure that whatever it is that we're doing is producing the correct outcome with the Alpha framework and everything that's embodied in that. And then we have it in a policy layer that sits on top of that, that dictates how we engage with customer data in our models going forward.

I love this. That's fantastic. That's exactly right.

In our Alpha framework, we have guardrails, and the agents are only allowed access to the data that they should be allowed access to, and they're only allowed to perform tasks that they should be able to perform. Right. ai and seven or eight other titles.

What's all the TV work we do? I speak to a lot of companies, a lot. 2% of where you are on this journey, right?

A lot of companies out here watching this saying, "Hey, how come we can't do this," right? I wanted to talk. I wanted to throw three letters out at you, MCP.

What's that? No, I'm just kidding. Model, context, protocol.

Oh, there you go. He's always keeping me on. There we go.

com. And developers, our third parties are using them today to interact with our systems. So you're already integrating with every...

You're letting other people plug in. Absolutely, and we've been shipping MCP servers since last year. Yeah.

Well- It only came out last year. Well, almost immediately after the protocol came out, I was like, "How does this work? " So we started building it across our portfolio, and like Danny said, it's great because partners and customers can use it to access our capabilities agentically, however they see fit.

But it also powers our agents- Yes ... right? And our new Aviator agentic hub, where all our agents are living, are all powered by these same MCP servers.

That's right. Yes. Well, and that's where I wanted to go with this, right?

Whether you call it a hub, I've heard others... So last week was Dreamforce out in San Francisco, right? Yeah.

Right. Marc Benioff, I think they use the term garage. That's where they keep their agents.

It sounds like a lower scale area. He might have his motorcycle there. Who knows?

What does he know anyway? Yeah. What was that question?

Let's talk about the hub, though, right? Yeah. Because this is an important part.

Depending on which analyst firm you speak to or listen to, we are going to have 10 agents to every person, 100 agents to every person, some exponential number. Yeah. And we need the ability, you need a hub.

You need, whether you call it a garage or whatever you want to call it, we need a central station where these agents drop off information, pick up information, get their next task, do their next task. Right. The real issue as I see it, though, is can we afford to have different hubs, or do we make our hubs using an MCP or these other things?

Do they talk to each other? Because we can't have 12 hubs in silos. So in service of that, we've got a multi-surface strategy.

Our agents are unshackled. They can go out into the world, and they can live in partner technologies or production tool, productivity tools that our clients and partners use, and it's the same agentic capabilities, but you don't have to come to us. In addition, our agents also support the agent-to-agent protocol.

So if you've got agents, our tagline is have your agent call our agent- Call our agent ... so that they can commune and create whatever business outcome you're trying to create. But by doing this, we're kind of syndicating the agentic capabilities in a way that's not duplicative, but sort of synergistic with wherever a user's trying to access them.

I get it. So if you have a hub and it's built in your e-commerce system, you just have to call, you can interact with one agent at Avalara using the H8 protocol or a different means. And this orchestration agent that we call Avi will then understand what it is you need to do, and then will pass that request onto one of the lead agents behind the scenes and go take care of the task, whether it's a reporting activity or a returns filing activity or whatever.

So you just have to integrate with one orchestration agent. Let me get geeky again with you guys. Please.

When you're talking about one agent, and then Avi talking to another agent who talks to another agent, are we truly talking about different agents here, or are they sub-agents that are being spawned and then taken back? Oh, these are different agents. So we've got an orchestration agent, and then we have specialized agents behind the scenes trained with different skills to do different things.

So a reporting agent is trained to go look at data and produce a report, answer your question. A returns agent is trained with skills on how to file a tax return in the United States, in Europe, like a VAT report or whatever. So they're different agents trained with different skills, and they exist today, and we talked about them here at Crush Today.

The orchestration agent understands what it is you want to go do, and once it understands what you want to go do, then it passes your request to one of the lead specialized agents. And behind each of those lead agents, there are other agents that will do the subtasks. So they're not agents that are spawned or created on the fly.

They're existing, and they're trained today. I love this. Very specialized.

So just between us, how many different agents have you guys developed at this point? Oh, geez. I've lost count.

Maybe you know. I would say many, many, many dozens. Probably- Many dozens.

Many, many dozens. And you're probably, if I- It's probably in the triple digits. Yeah ...

it's going to be over 100. Really? Yeah.

Wow. Yeah. And a lot of them have been in the product.

We've been building them for over two years now. So they're built inside of the products today, doing things behind the scenes. And today, we're talking about exposing the orchestration agent and the lead agents to our customers directly through Aviator, and this is something new.

But all of the other agents are working behind the scenes in our products today. I get it. I feel like I'm doing a case study for my next research report here.

Do you envision continuing spawning so many new agents, or do you think at some point these agents become sort of multi... They can do multiple things. You don't need a specialized agent for each task.

Well, that's a kind of a layered question. In the now, we're spawning additional specialized agents, given sort of the metes and bounds of the agentic capabilities at large. Mm-hmm.

But over time, to your point, if you kind of just imagine what the next wave of AI unlock is going to usher in, it's probably going to be the case that their context windows and tool selection and other things that kind of require specialization today will melt away, and you'll have, I think, more consolidation. It's probably going to happen. Yeah.

I think it's good. It's absolutely going to happen. It's just the way of things.

Yeah. Right. Speaking of the way of things, I feel obligated to ask you guys this.

Look, the last week we've had the AI doomsayers. " Right? If your company was worth five and a half trillion, you probably wouldn't worry either.

But that being said, do you worry? Do you worry about, do we have our agents under control? Well, I don't worry at all about that, because we do objectively and empirically have our agents under control.

If you're asking more in the broader like- What about the AI slowdown, if you will? The global... " Because we all don't know as we sit here today talking about, right?

Nobody does. I don't care how close you are to the latest break or anything, and there are concerning elements about that, obviously. Sure.

" Yeah. The Sky Net takes over. We don't know.

But we're not worried about it today in the now and certainly not in any of the agents that we're building. No, I'm not worried at all. There were so many stories over the past year about agents will replace certain functions and replace certain people.

You see some of it, but we have engineers, for example. We have to hire more engineers to go help build the technology. And AI writing code, the code in our products, we still need humans to review the code and make sure everything is implemented the right way.

So, yeah, I'm not worried about the AI doing bad things right now. Good. I want to return to Avi a little bit if it's okay, guys.

Sure. So let's talk about the future of Avi, right? We talked about where it is today.

We talked about you've made these two primary agents sort of front and center accessible to customers. Where does Avi go from here? Well, the evolution of Avi is a byproduct of all of the sort of lead agents that we continue to imbue with additional capabilities.

So on the one hand, the breadth of Avi's capabilities will continue to grow as we continue to build additional agentic capabilities. On the other hand, when we're talking about multi-surface, Avi goes literally to where work happens. Avi lives today in the Outlook email program.

It lives inside our partner NetSuite's ecosystem, where we won AI Partner of the Year because you can access Avi from within NetSuite to implement- That's like- ... " So I think the evolution is kind of multifaceted. On one hand, it's that functional expansion.

On the other hand, it's that technological expansion to go to these other surfaces to make it easier for customers. Do you see Avi outgrowing the sales tax issue or market? Well, yes, because Avalara is the leader in agentic tax and compliance, and that compliance word is a little broader than just tax.

Sure. So for example, we're one of the leading suppliers of e-invoicing technologies around the globe. So all these countries are passing these laws, forcing people to send their invoices electronically.

Well, we power that in over 60 countries, and we're building the world's largest B2B network. We've got over 62 million companies in our directory already. Wow.

And so as we see the next sort of evolution of this thing unfolding, it's to bring the Avi and Aviator capabilities into some of these other product suites that we have that have been outside of the transactional tax arena, and to start to supercharge the world's largest business-to-business network in the process. I love it. Powered by Avi.

Yeah. Aviator. Guys, I want to thank you.

Yeah, absolutely. This was fun. Yeah.

In the words of Ice, you can fly with me anytime. I love that. Mate, but seriously, what you're doing, man, you guys are way out, way out in front on this.

It's invigorating to hear what you're doing because I speak to a lot of people who, we've all seen the surveys, right? Yeah. 95% don't see an impact on revenue.

How many projects don't make it into production, and here you've got real live AI agents working, doing, with an Avi kind of hub to work it on. You know what? Check it out, and if you think this is just about sales tax, you missed the whole point of this.

We're going to continue live. We've got a lot more Avalara Crush coverage coming your way the next few days. But for now, you're watching Techstrong TV.