The Broken Revenue Systems and AI Agents Rebuilding Contract to Cash with Ali Hussain

Welcome to another episode of Future Finance. Paul, I've been back to back meetings all day. I don't even know where I am. I'm going to be leaning heavily on you in this episode to really keep us on the rails and welcome our next guest and let me know if I just fade off and start rambling. And well, wait, that's what I do in every episode. This is what's changed Glenn. No, great to see you, Glenn. Excited to be here with you. And this week we have Oli who's on with us. Oli, welcome to the show. Thanks Glenn. Hey, Paul. Both of you. Thanks for having me. I'm very excited to be here. I'm glad I get both versions of you today. Well, thank you. We're glad you're joining us. I know we had hoped to record this a little earlier and had it delay it. So I'm glad we could get our calendars back together and make this work. So this will be a lot of fun. 100% post-tax season. Always a better time to roll through. We'll have more more listeners as well. All right. So a little background about Oli and then we'll jump into things. So Oli is the CEO and co-founder of tabs, the first AI native platform, automating the full contract to cash cycle for modern finance teams. Before founding CAHBS, he served as COO at LACCH, where he scaled the company from seed to IPO and experienced the inefficiencies of legacy accounts receivable systems firsthand. His background spans product leadership at Google, strategy consulting at BCG and public policy. At tabs, AI is modernizing the 125 trillion, that was that all. Global Bing-to-B revenue market by replacing fragmented finance tools with intelligent AI agents that handle billing, collections, and revenue recognition. Under his leadership, CAHBS has scaled rapidly, crossing 1 billion in annualized billing volume and serving over 300 customers. Oli is emerging as a leading voice in vertical AI and the future of finance and infrastructure. So again, welcome, love the background, love what you're doing at CAHBS and really excited to get some of your thoughts today. I appreciate the bio and thanks for introducing us and I think the cool tip it is when even when we wrote that it was close to a billion, I think now we're close to five billion. So when you just think about the velocity of how much money is moving in systems that are now leading on AI, it's been amazing to see the scale in the market. What's exciting, congratulations on closing in or getting that five billion, here's to 50, 100, 200, you know, wherever it is, right? That an old teats brilliant number, I'm not even scratching the surface yet. What being an AI company right now in the LLMs, the frontier models are getting so much better that there's this, well why couldn't I just vibe code that myself and I'm thinking about exactly what you're talking about. This wasn't an AI job, this is just some basic fractional finance work I was doing company and we were going through and trying to use an LLM to find some contract, several hundred contracts, some had one specific line for cancellation in term and anyway, it was when you're trying to chunk and vectorize and go through all the day, it was very hard to find that specific use case that we were looking for and it was interesting because we thought it would be something that an LLM would have a really easy time doing, but just the way that the chart, and it was somewhere PDF, somewhere like the image kind of be, anyway, it was a total nightmare of the situation and for you guys to solve that at scale, I always have a hard time not getting into making the sausage here and obviously a lot of that's proprietary, but this isn't something somebody's going to just go vibe code. And I think it goes to my philosophy, right? Like I think LLMs are amazing tools, particularly for reporting. Once the data is cleansed, Instructured and understood from multi-system, multi-contract, etc. Downstream, when I think about finance use cases, I was at a talk recently with the CFO of Anthropic, Chris Nen, like they're using LLMs for amazing reporting, but they're still using traditional systems to run core accounting, workflows. And so back to your point, like contracts, taking them over and putting them into operational parks of finance are very hard to do with just yourself out of the box with LLMs. And the reason is is by nature, LLMs are non-deterministic, but things like invoices and reverect are highly deterministic. And I think there's a reason why a lot of them have been very effective in legal and areas of just understanding a simple clause and trying to get an alternative to that clause in those type of workflows, but something to get to a level of taking pricing out of a contract and producing an invoice. The problem with a public LLMs is you can do it three times, you'll get three different answers. And that's very difficult. It's still cool and like you get to some insight very quickly. And so a lot of where we think the best finance systems are going to be highly defensible in the world of public LLMs is really tightly built context graphs, which use LLMs, but they also use other types of AI. So when you look at our commercial graph, there's a lot of last mile machine learning that is required to get to consistency so that if you load those 200 contracts every time an invoice comes out or every time we build your deferred revenue from a gap standpoint, it is the same over and over again in line with the memory and the direction you've given on how to infer your contracts. And so we think that billing is a really bad use case of out of the box LMs, but once we have all of your data right, we're happy to give it to you and the way you want to manipulate that to do reporting, it's an amazing use case. And so that's one of the ways many of our customers have been able to live the best of both worlds where they rely on tabs to get the context, read contracts, build invoices, do all their RevRack, but then we can give them their data to do a lot of their FPNA and a lot of their reporting in ways where they can pull data out of tabs and bill and in Gaston away that otherwise would have required a lot of tooling and a lot of human labor and Excel. As I listen to you talk, it reminds something you and I have said a few times, Glenn, you gotta have good data before you expect the LLM to give you good results. So extracting the data has to be, yeah, that's a huge part of it is pulling out the right data to feed these systems so that they can add any value. And to me, there's like four layers of where the future is going. There's a data layer, which is the most important. Think about it like that, the gasoline or the actual vitamins you put in your body. Like it's the most important, not gasoline in your body in your car, but like it's like allity of what you're feeling. And if that's broken, everything down's free, and you're going to have a poor drive or your health's going to weaken or wherever it may be. And so that context is the most important thing in the world of AI. And that's where hyper specialized tools that know the domain that can pull out contract data, usage data, key fields from your CRM, your data lake, etc. That is where you should always buy versus try to build. The actual interface is the software. I do believe is going to be the most commoditized. And that's why in certain areas, I'm just skeptical of all these AI tools that like are purely just a operational or software layer if they don't have a data mode. That I do think is a race to the bottom, but it is a necessary part because whether you or your accounting firm or your agent are working, you still need systems of record that show the work and show the audit trail. And I have a hard time believing you're going to build something that's good enough for an auditor or third part of yourself. So I still think you should buy even though it's not as it's not going to be as expensive as it was in the past. And then the final piece to your guys' point, the reporting part I do think gets more commoditized because that's so bespoke, so long as you have good data coming from these systems. The final part though that we're not yet talking about, but I can bet when we're on this show together a year from now, we'll be the specialized agents that are not humans that are also working on top. That is the other area. There's the data part and then the last mile agent. I do believe those agents are going to get hyper specialized to these roles. And that will be another area that will be incredibly hard to build and you will just buy. So everything in the middle gets commoditized, but the data mode and the actual agents that start to look and feel like team members, there's fine tuning of agentech work and context will get specialized enough where it will make sense to pay up for because of the quality of work they do and the specialty of accounting workflow. They know how to do it. Got it. Speaking of agents, obviously we keep hearing about those. It seems like first it was the LLM, it was prompting, you hear a lot of scripting, but now we're hearing a ton on agents. So how would you define an AI agent? What do you think they're good at today? What are their limitations? And where are we going with them? I think there's going to be a lot in for a little bit, a lot of noise around agents where people will look at something cool that LLM can do and be like, oh, that's an agent. And I, my bar's much higher for that. That is just a cool workflow or table stakes like automation. And so to me, there's a distinction between things that move more automated, but are still requiring a human to work on top. That to me is AI automation. It's not necessarily itchetic. Agentech to me are end to end workflows. They're being fully done, independent of a human being involved. The human can give direction, but a human's not coming in and really touching the steering wheel. The agent is getting from x to y location, totally on their own, and starting to completely displaced entire world. jobs that people do, whether it's a part of their job, all the way up to their entire job. To me, I think the bar for an agent is very high. I think we are still in very nice synth days of who's actually accomplished end-to-end, agent-like outcomes. I think there's a lot of end-to-end AI automation, but an actual ability to say, "Hey, I'm not hiring anyone else in billing, because now I get a agent from tabs, or I don't need a payroll admin anymore, because I have an agent from Deal. I think we're a little bit out from there, but we're not that far away." So to your question, Paul, where I think we're headed is in the coming 12 months, we're going to see a lot of lower-level jobs to be done, starting to get displaced. That doesn't mean necessarily you're cutting all your headcount, but you're able to continue to scale and do more with less. So my prediction for the next 12 months, and a lot of where our roadmap is headed is, "What are the entire jobs that can be done that are highly tactical?" In Billings, a great example, the amount of people that spend time, a good portion of their time each month, building invoices, sending them, following up, sharing a PO number, submitting in the portal, a W9, trying to reconcile payments from a bank account. That is an area of where we believe the technology is good enough to fully take that over and to end going into the next 12 months. That is not a hypothetical. That is an inevitable. And the quality of the data, the LMs, the tooling, then vendors like us have built, is starting to yield results where you can imagine that happening in a good amount of the market in the decent future. The harder part of agents is what I call the more intimate workflows, like replacing what someone like a revenue accountant, or an FPNA lead or a strategic finance lead, that I think is going to take a little bit longer. But that is also an area where agents are going to fully take over. And so my threshold for agent is less about, is it eating into ERP spend or software spend or services spend? Like if you're displacing modules and services from your ERP vendor or your third party admin, like that is like automation. But if you're starting to look at your finance team of 15 people and you're saying, look, I don't actually need to hire two more people in billing or in payroll or in revenue accounting or FPNA, that starts to actually-- that's where I use the AOR. Like where the agents have arrived. I'm so aligned with you on people who are calling everything an agent for a while. And I think last time I checked agents were at the absolute peak of the Gardner-Hype cycle right now. And like what you were saying about using machine learning at various steps in there, I fought this and Paul's heard my rant a million times, so I won't do it again. But when you call everything an agent, it lessens what a true agent is going to be. I eventually gave up that fight. And I would build for clients an agentic workflow that had deterministic steps along the way. You sprinkle some AI where it makes sense. But otherwise, it's like a decision tree that it just goes in it. You know, it's a deterministic workflow and all that. But I mean, I'd be curious. Like I know some of the coding apps. I don't remember which one on the benchmark had the longest running agent or whatever. It might have been Carcer. But you know, they've got coding agents that will go out and work for like nine hours on something and then come doing truly agentic work as far as I can tell. But I do think it's important to really understand. And this goes back to your probabilistic versus deterministic approach, or comment earlier, where an agent is still-- if it doesn't have those decision gates and those guard rails, it may go off and do a task and do something completely different than it did before. So it's going to be interesting to see how it shakes out and then how we rein in the agents to keep them-- I'm going to Nick Vostrom's paperclip thought experiment right now. But it's wild times in a lot-- six months. 100%. And that's why we're so fix it on this. I still think the most important part is the data context. That has to be perfect, because if the agent is working on something that is imperfect, that a human would have been able to solve without a system, like we have a big problem. So the data parts huge. But to your point, there are decision rails, there are controls, there are areas that avoid hallucinations or things that could go and send the agent array. But also, if you think about something like billing, you need preferences stored in memories. Like this is the way we want you to communicate. This is the tone. Don't go escalate, donning, to Chipotle. It's our top customer. We don't want to piss them off, like turn off, donning. Like, those type of controls are incredibly important. But they can be built. And I don't think you should build it yourself. But the way you now have vendors like tabs that have classroom class engineering teams are building off data sets that are just yours, but off of hundreds, if not thousands of companies and getting data on hundreds of thousands of invoices being sent, you can actually pattern match to a pretty good system. Like if you think about cars now autonomously driving on complex highways and in cities, it's not that unfathomable that a billing agent can take you from a contract to cash without a human basically touching anything other than once in a while touching the steering wheel to make sure you approve of invoices or who can send the invoice or of a certain size or if you're going to accept credit card, et cetera. So that's the direction that I do think we are headed. But it's going to require really thoughtful specialists like us to get it right versus you going out in on a $50 million business trying to buy code something. I literally had a client, one of our customers that were about to go onboard. Like they went from one system to another. They're not on the billing side, but on the ERP side and like something got triggered and they sent 900 invoices twice to all of their customers on a complete different ERP. And it's just like those mistakes can be incredibly costly, just using a specialist tool and not spending your end year in time on it is incredibly valid. And maybe this is 10-genial to that or maybe I'm completely changing the subject now. I'll be curious to hear your response on this. You mentioned 606 and I'm thinking about sort of standard, whatever SaaS contracts and all that. But as especially with AI being more and more a factor of what people are delivering, pricing models are changing. So it's usage-based, it's seat-based, it's value-based and all that. I don't know. I mean, I guess ultimately that will impact the kinds of agents you're building. But maybe you already have it. But how are you handling increased complexity in contracts and being able to do this pricing? Because I know that's a big, especially around usage-based pricing rather than the seat price. 100%. The reason we've done so well in the last 12, 18 months, I earlier Paul asked why I started the company was because I believe revenue had been left behind for finance operators. And the second reason that I didn't get to this Paul is like, if you actually look at the 4.7 trillion accounting and finance economy, when you look at the jobs that are being gotten from a finance composer standpoint, most are tied to the controlership and accounting, whether that at the corporate level doing things under the controller around billing in RevReq. That is typically 60% to 70% of finance teams. But then if you look at the professional service side of a lot of these big top 100 firms, many of the people are an audit, revenue-related audit, revenue compliance, et cetera. So one is it's just a massive labor market. But then you're going on to something which is like the holy grail of what's going on right now, which is revenue is getting more complicated. Whether we acknowledge it or not, the prior conventional wisdom is you hire a mature finance team. They come in and they standardize your contracts. They once in a year align on pricing and packaging and they create structure and controls. The last 12, 18 months have thrown that out of completely out, meaning with AI, revenue is getting more complicated. But no one actually knows the right answer to it other than we have to move away from flat fee and even seat-based pricing on the software side. But even on the non-sort software side, the way service providers, professional service firms, even hardware manufacturing companies are moving the way they sell products and the terms and all of that. And so our general mission is internally as a company is to make complex revenue simple at scale. And the way we handle it goes back to that al-lem layer. Like we're agnostic to your pricing and packaging because we understand and structure it at the moment a contract signs. We're not relying on a price book, any type of historical structured pricing or knowledge of your business model. The same way an accountant or a human reads a contract and then types in what their understanding is. The al-lem and other deterministic machine learning is highly discerning of that and can support any type of business model. And so if someone asked me what's the most surprising thing about your business, 60 to 70% of our clients now have a usage component to it. And that's not just clients like cursor and together and others that are like AI companies. A lot of legacy platform companies are now adding some form of usage structure on top. And we're getting incredibly good at not only structuring it, but most companies then feed a usage feed into tabs. It doesn't have to be API. The elements really good at capturing unstructured CSV, uploads as well from BI tools like Looker and Tableau. We do the math and we do the invoicing, but then most of my roadmap right now is going really deep on the revrex side because to your point, when you add usage and all these complexity, doing your 606 on usage based products with tiers and minimums and pre-commitments and performance obligations, it is like a 7/11 freeze times, 100 times over. People are truly lost. How do you do SSP and revenue allocation at a time of usage and professional services at the same time? These are all breakthrough macro chaotic moments that are really allowing us to go back and build products that are helping support finance teams going through immense change right now. I love the answer, I love that you took the time to join us today, Oli, it's been a lot of fun, Chad. It's a deep pleasure, Glenn Paul. I know you both both very busy, so thanks for him or me this afternoon. I'm excited for the next time. We appreciate you joining us. It's been a lot of fun and we'll look forward to the next time as well.