Token Series Part 2 | with Finance's ChatGPT moment: MCP, Harness, and Headless Explained (simply)

0:00We're talking about AI usage. It's supposed to make us all efficient. I have never been busier. CJ, I've never felt as busy myself. I actually think finance is the most AI forward after product and engineering. It's a hot take. I like it. For lack of a better word, I think of a headless agent as one that does not have a leash on it. You want to at least have an electric fence if you have a nice yard with your dog outside. But ultimately, you need a really good knowledge graph around all of your vendors. I actually put employees in there because it is, if not your biggest, your closest to largest cost center, and then your customers.

0:33There's a lot of open questions. There's parts of our jobs where AI has made an email or a certain level of insight or maybe the prep for today's call a little bit easier. Now it's just increased the quantum of output we have to put out. And there's a lot of things that still require a lot of human thought. Is this thing on? yesterday's price is not today's price

1:05me and you were having dinner earlier this week and i was saying hey you get texts all the time from cfos are they just asking you like explain how your product works with usage-based pricing or this and that and you hit me with a zinger you said that a lot of them are actually asking about the org design first. That caught me off guard. Yeah, and look, I think my team does a better job talking about our product, but I think what's most curious for most CFOs is they're like, hey, Ali, you get to be on like six or seven chats with CFOs every day. What's top of mind for them? And beyond the technology and pricing and costs and all the standard things that have now become inevitable in the market, I think a lot of the finances I spend a lot of time with have just a lot of experience.

1:49And what they've been most challenged by now is what does the new order look like? That is tooling, but a lot of it's just the type of people you hire. I think that is probably the top conversation I have with my deeper relationships I have with different finance leaders that either are on tabs or who we're prospecting with. And I think there's a couple of things driving that. One is just there's a lot of pressure from their boards and their management teams to just rethink the type of people you hire. And often finance has to be the most selfless when it comes to being a little bit leaner. The second is I actually think finance is the most AI forward after product and engineering.

2:26It's a hot take. I like it. We see it in the data in terms of the adoption of tools like tabs, but also on top of that, when MCP came out in the second half of last year, the way I found out about it was not from talking to my own teams in non-finance or engineering or in the news, the early signals that this was going to be a catalyst in the market was from CFOs, controllers who adopted MCP into something like Claude faster than anything I've ever seen in my life. I think they are deeply thinking about org design right now. I think there are some must-haves, like I don't think anyone is saying I don't need a controller anymore or a very thoughtful FPMA or strategic finance lead.

3:10But I think what they're saying is, after that, where can I get just a really AI-pilled team member that can allow us to keep most of the team strategic, but then really forward-looking on the latest technology, using LLMs, and just figuring out how the rest of the team does more with less. And so I think that's been the top conversation is a lot of them know directly the type of person they must have, but it's the added hire that would have otherwise created a bunch of people under these roles. I think it's the biggest question of what that looks like. So I do these finance tech stack reports.

3:47And last year I was saying that finance had not had its chat GPT moment yet in the same way that call it support had already had with a Decagon or something like that. And same thing like for engineering, having a cursor or a codex. I think MCP was the equivalent of finance and accounting's chat GPT moment because now you could query using language all these numbers from disparate sources that you couldn't before. I think finance did have that moment. And now look, I think there's still a decent tale of finance leaders who are still getting comfortable with it. And frankly, a lot of the key systems they work with are still getting really good at feeding data into MCP.

4:23But I think that moment has arrived. I would debate that thing any night and day. I mean, the things we see at tabs of finance leaders, figuring out how to pull immensely powerful data out of tabs, their brecks, their net suite, their deal or rippling has just become incredible. The second order now is how do they build around that? I think what we are seeing from a lot of folks is in some cases, the GL is interesting in terms of how to pull AI, given MCP can pull a big amount of the reporting layer. The next phase is how does finance present more context to MCP? And I think that's the thing that they're really hungry for over the next 12 to 18 months.

5:06Well, something that you've shaped my mindset on is that there have been a number of companies, whether that's Rippling or Ramp or now Tabs, who lay the tracks for agents to travel on, right? And that you can have some core unit that's at the center of it that allows you to access other data points in a way that is deterministic, but you can also leverage AI on top of it. I think MCP for finance, it's great. But to your point, you still need to have some sort of knowledge graph for this to travel on. I can't just be out there wandering around. They often think that that knowledge graph is in a consolidated form in the GL or the ERP, really at the subledger level.

5:49Like ultimately, look, it differs a little business to business. But ultimately, you need a really good knowledge graph around all of your vendors. I actually put employees in there because it is, if not your biggest, your closest to largest cost center, and then your customers. Those are the three primitives, whether you're running FP &A, reporting, trying to understand your COGS basis, your revenue, your growth, your opportunity. And so getting those three graphs right, which I do think are really hard to build yourself. You need a ton of expertise to get that right, given finance and accounting is very little margin of error.

6:24But if you can get those three things in a really good spot, the power of MCP becomes almost infinite. I think the folks have really nailed that when I think of my power users who have just an incredible context layer on their customers and something like a tabs, their employees and something like a rippling. And then the route they go on the vendor side, whether they have a procurement tool or a consolidated AP and spend tool, they're pretty dangerous right now. Ollie, we were throwing around MCP. For the benefit of listeners, some people may already know what an MCP is, but we've also brought up the terms harness and headless.

7:00I'm going to ask this people a question on their behalf. Can you break it down for us? MCP is kind of, think of it as the connection protocol in which how particular, I mean, is built for Cloud connects to your systems. And so think of it as a parallel to an API, but it is flexible in the way that tools like Claude can connect into your legacy systems of record as well as other systems. And so NetSuite has an MCP into Claude. Tabs has an MCP into Claude. You can access all of that data, build agents on top, and more and more bi-directionally interact with your other tools within an ecosystem that you trust in a closed environment.

7:42The easiest way I would think about the harness is the orchestration layer between these different tools. And so a good example is we've built a lot of harnesses within tabs so that when we have an application layer within the platform, it can connect directly to our billing agent or our revenue accounting agent. And the harness effectively serves as that interconnection between the action agents and the actual controls and applications that you have within the platform. Headless is this concept that an agent can just be swarming on their own and not be brought into some form of constraint or control.

8:24And for lack of a better word, I think of a headless agent as one that does not have a leash on it. You want to at least have an electric fence if you have a nice yard with your dog outside. I've heard the terms system of record, system of action slash system of execution. depending on which word you like better. How do you frame this up in context of the CFO tech stack? Is one superior to the other? Should certain layers always be deterministic? Should everything just be completely AI? What are we looking at? I've been the worst at this. I've kind of used all three interchangeably. I have this incredibly talented product marketer on the team.

9:01And she said, Ali, we're going back to the operating system. And for a minute, I was like, whoa, that seems so 2017. and I actually think it's not a system of record or a system of action or a system of whatever anymore. I think it's an overall system. And in that, there are probably four things you need. You need that context. And I don't think one context layer can be sufficient for all of finance. I think it has to be tied to the data primitives of customer, employee, or vendor. But then you need an application layer that is deterministic, but also highly auditable and deep in the domain. And so the way we think about it, though it feels like a traditional piece of software, it is highly dynamic and being fed by a bunch of unstructured context that is at the intersection of structured and unstructured.

9:53But I don't think you can run a finance team without really good application. But then on top of that, now are seeing these harnesses that allow first and third party agents to work on top. And so that just system of action is no longer sufficient. It has to be an application layer in which it is built for both accountants and agents. And so that humans and agents can work interchangeably. And that is where at the sub ledger level is where I am uniquely pessimistic that a lot of the legacy technologies can be resilient systems of record or system of action or whatever you want to evolve to, is very hard to work backwards and build in a way that is conducive to both an agent, a first party accountant that you are building for that specific job to be done.

10:42But then on top of that, the crazy moment is then an aperture that a third party agent can come in, whether it's being built by the CFO or an outside audit firm or another vendor that has to interconnect into the OS. And so I'm calling it none of those things anymore. I'm kind of calling it these operating systems that are ultimately where I think the future goes. I like that the pendulum swings back or operating system sexy again. After she relayed that to me, it took me a couple of days and I woke up and I said, you know what? This is real. And I will say, I think finance has been a little slow to characterize the category this way.

11:18To your point, I think finance came in a little late and then went straight to the downstream insight and reporting layer with something like MCP. But I think if we were to look at our friends and peers in categories like legal AI, the dollars in that domain have just come a little bit faster. I think they've started to really think about this like an operating system. And I think the best way finance catches up is starting to think about these systems in a very similar way. In fact, I would argue finance needs it more just given the level of compliance and auditability. I want to talk about governance and we can split this into two angles because governance has the SOX compliance, the auditability piece, the I wear a belt and suspenders to work.

12:01You can trace this angle. But I have to admit, it also has a cultural angle to it of how fast you allow the business to go within the org. What is the norm? How are you thinking about governance in this new operating system, we'll call it? You just got to start. I'm not saying you go straight to your audit on this new way. I think there are certain places where you should start to rip the band-aid and push the muscle. And so what I often tell people, what's like the areas where it's kind of low lift, repeatable tasks, also low risk and often very commoditized labor. And so a great place to kind of lay back the governance and start to build trust in things like agents is something like billing, collections and payments.

12:41And more the billing and collection side because payments is just a structure within it because you're already yielding it to either a very junior person on your team, or in many cases, an outsourced firm or folks offshore. And so actually, the risk of that governance, in our view, is very low. And yeah, from time to time, you may make a mistake and upset someone, but it's usually something that's salvageable. And so I think that's the area to start. But I do think we are moving in a way where revenue accounting, et cetera, start to move in this direction. But I completely appreciate, given the stage you're in, the level of comfort you have with these tools if you want to crawl, walk, run.

13:20That said, I think the pace in which we're moving, the bar of governance is changing a lot. I think a lot of the burden is on builders like myself to ensure the right controls, access points, etc. in place. And these have been traditional finance tooling. Finance has always had to build technology that has the right rules-based access, the right compliance, the right controls within the system. And it goes back to why I think certain technologies are just going to be better served because they're going to be building these things not just for humans and their permissions, but also agent. If we flash forward a couple of years here, do you think that deterministic application layers within the CFO tech stack still exist?

13:59Yes. LMs by nature alone are insufficient to get to those types of outcomes, but also those types of traceability and auditability. I think less and less it's humans that work on top of those deterministic layers. Agents will always work on top of them. I've been the biggest pessimist of like a headless agent that takes context and just does things. I think the application or the deterministic layers create the common sense spine behind all of this that puts it all together. And so I am very bullish long term on the deterministic layers where I'm looking for are ways which humans start to be less and less an active part as a player and more become the coach.

14:42If something fires a second too late, that can still be a quarter million dollars too late. You're operating at a sector that can't really have these mistakes. How should we be squaring that up? This is where I think using tools that are specialized to the domain actually really matter. What's been really interesting is even before AI, the domain we've been in in billing collections, RevRec, Rofin, build versus buy. And one of the skepticisms with AI was more people are going to try to build versus buy things around billing or RevRec. And we've actually seen a complete reversal to that. I think people are excited to use AI, but realize that without controls, AI can do more harm than good, particularly around accuracy.

15:29And so a lot of what vendors like us do is create specialized training and non-LM based kind of ways to augment AI, such as machine learning, to get to very high levels of accuracy. Now, the most important thing, though, is like we can just say that that's no longer good enough. And so you can come into a system and actually get insight into the level of accuracy that has been trained on a closed system that has just seen this across many other examples or customers, etc. To actually give you that insight, which is very hard to do on your own. Accuracy is probably the most important part of control.

16:09I am still of the view that is very hard to build accuracy, particularly for very high accuracy required workflows. And therefore, we're actually seeing a complete change from the build attitude versus like we want these outcomes that we know can come from AI, but it's time to buy. And maybe just to touch on for a second here why your neck of the woods has historically had a lot of builders in it. And it's been hard to build a horizontal solution for it. And many have gone vertical. I think about getting paid as intimately tied to the front of the house and how a company monetizes. Like the way a performer charges for work is going to be very different than a software company.

16:48Whereas on the expense side, both of those companies still rent. They both still buy pencils. They both still pay way too much for sweet green. It looks much similar on the expense side of the world. Yeah, on the revenue side, the biggest challenge historically has been there has never been a one size fits all or a consistent palette for how you do business. Everyone has very unique commercial models, different types of pricing and packaging, contract types, discounting. And so historically trying to understand that like you could a vendor relationship in structured form has effectively proven impossible.

17:28It is why billing and revenue related work is the most laborious and most of corporate finance. And that is uniquely a great example of where AI and LLMs are amazing because they can understand very complex commercial models at a even customer level in a way where you You don't have to structure and tool something based on what a human inputted, but an LLM or an agent can literally understand it the same way an accountant would or a lawyer would and then translate it to a workflow. And so it is a truly cutting edge frontier technology that would not have been plausible until frankly, mid to late 23 or really going into 24 and 25 when the model started to get good enough.

18:11And I think that's an excellent point to leave people because when we come back, we're going going to talk about the pricing and packaging behind all of this. And more importantly, we're going to talk about who foots the bill. Run the Numbers is a mostly media production. Yelling an intro by Fat Joe. Artwork by Meg D 'Alessandro. Nothing said on this podcast is intended to be business or investment advice. It's the sole opinion of me, a guy who feeds his dog way too much ice cream and has a history of net operating losses. LOL. If you like this podcast, hit subscribe and give us five stars. It will take like two seconds and our algorithm overlords Love it.

18:45Drink water, call your mom, and have a great day. Peace.