Video: Keeping up with AI: Modernizing AP and AR | Duration: 3604s | Summary: Keeping up with AI: Modernizing AP and AR | Chapters: Welcome and Introductions (9.76s), Agenda and Audience Poll (145.53s), Platform Overview (243.315s), Why AI Fails (355.74s), Hybrid Agent Approach (635.37s), Agent Teams (793.145s), Least Agency Principles (872.765s), Automation Comfort Levels (1210.505s), AP Agent Team Structure (1306.595s), Agent Implementation Examples (1515.71s), AR Agent Team (1802.655s), Live Demo Walkthrough (2071.16s), Agent Governance (2437.09s), Q&A and Next Steps (2636.355s), Team Structure Strategy (3476.36s), Closing Remarks (3550.195s)
Transcript for "Keeping up with AI: Modernizing AP and AR":
Hey, everyone. Thanks for joining. We'll give it around a minute just so we reach a a quorum, but we're excited to get started, and we appreciate you tuning in. And, yeah, drop in the chat where you're tuning in from. I know, Thomas is is bright and early in San Francisco, but... Okay. We have Jersey, India. Cool. Awesome. London, Arizona, come. from all over. Nice. We have all over all over the continents. Well, awesome. Thank you again for taking some time out of your day today. We have a really exciting topic. We'll basically be speaking more to, like, tangible and actionable ways where you can implement agentic automation within your finance department through AP and AR. Just do a quick round of intros. So, if you go to the next slide. My name is Chris Ferraro. I'm on the product marketing team. I work primarily on our agentic features and launches, and I'll let Thomas from the Sligo team introduce himself as well. Yep. So I'm Thomas Neil. I'm the head of a genetic transformation here at Zaligo. By background, I'm a CPA. I spent my career running accounting and finance operations before moving into building the AI agents for those same teams. Today I'm going to share a lot of what we've learned from building AR and AP agent teams including what's worked, what didn't work, and how we've decided how much an agent should be trusted to do. Alright. And as we get into our agenda, we're first going to cover the orchestration layer and really how the Soligo intelligent automation platform fits into the picture of agentic automation and finance. Then Thomas will take over, and we'll talk about agent teams, the level of autonomy you should be giving agents, specific use cases across accounts payable and accounts receivable. And then we'll actually share a demo from Sandeep on our internal AI team and finish with actionable steps on how you can get started and really put your best foot forward. We'll end with a q and a as well. So if you do have any questions, feel free to drop them in the chat. We'll make sure to get to them at the end. But first, it's always helpful to get a sense for what everyone's current state is when it comes to their journey. So we'd be curious to hear where is your team at with AI agents in AP and AR? Are you already running agents in production? Are you sort of in the pilot stage where you're experimenting? Are you before the pilot stage where you're evaluating your strategy and options, or have you not started yet? And, again, there's no wrong answers. Answer it as truthfully and as honestly as you can. And we'll give you thirty seconds to a minute to see these results start rolling in. Okay. I'm I'm appreciating the honesty so far. It looks like most teams out there are either evaluating options or haven't started yet. That's totally okay. And I think it's honestly a a good place to be at, especially as you take in this presentation because we'll give you actionable places that you can you can start your journey and really what to consider as well when it comes to AI governance and least agency. Before we get into the more specific use cases and how you can build agents in Soligo, first, we're gonna take a step back for those of you who aren't familiar and really walk you through what Twiligo does on one slide. We do a lot, so this is our best attempt to consolidate into one slide. The main point here is that we're a connectivity and integration platform, and we're allowing you to do all of your integration work from this one single unified experience. So across the top, you can see all of our solutions. We do EDI. We do app to app integration, DI, AI agents, MCP management. My point here is that it's all in one platform. Security and governance are built in. They're not bolted on. We we started in the back office in finance departments where our system couldn't afford to go down. So the platform runs on enterprise grade architecture, 99.97% uptime, and scalability to match. And that reliability is ultimately why thousands of finance teams trust us with their data and processes. Now we can start to talk about sort of why AI pilots stall, what we've seen in the market, and what what these pilots need and what these agents need to reach production. So first, why are they failing? So we've seen in the market is... And I... This this stat was from roughly a year ago, so maybe it's gone up slightly. But we're still at a point where a wide, wide, wide majority of organizations are reporting no ROI when it comes to their AI initiatives. So why is that? Well, one, workflows are brittle. They're built maybe beside the process instead of inside it. They don't learn. They don't keep feedback or improve with use, and they fit poorly with daily work. So people are are starting to route around them. What works in the AP and AR space is kind of the opposite of that. Right? You have hybrid flows where they're a combination of deterministic and agentic. Rules carry the structure and agents take only the judgment steps. A learning loop where humans can get in the loop and provide data to the AI so the inputs... Or so the outputs get better over time, and then agents that work inside the systems and approvals your team already uses. So ROI really follows that adoption. Another critical reason why, if we go to the next slide, a lot of these initiatives never reach production comes down to integration and connectivity. Right? A lot of AI orchestration is an integration problem because an agent is only as good as the data that it can reach, and that data has to be specific. It has to arrive in the right format with only what the agent needs to do its job and nothing more. If you give an agent too much information, one, it creates security risks, but, two, it's also gonna lead to hallucination. Right? Even when you feed an LLM too much information, you see that the quality of outputs gets worse over time. We wanna give agents specifically what they need to execute the job and nothing more. And that data could be spread across multiple systems. Right? It could be across ERP, your CRM, your email, your Slack. So every... Without a integration layer, every agent has to wire its own way into each one of those systems. But with an orchestration layer, we're handling all of that. Right? This is what we know best. Connectors, off, model routing, guardrails, limits. That's why a platform like Soligo is the missing piece. We're connecting your agent to all of the enterprise systems that it needs access to to do its job, and we're writing... We're routing, each task to the right model as well. If we go to the next slide, you'll see here that we acknowledge not everything's agentic. Right? There are rules based use cases. If you're just syncing data from one system to to another. There's also purely agentic use cases where it's a task or an assignment that requires human judgment. But a lot of what we've seen when it comes to agentic use cases is that they sit somewhere in the middle, and that's what we call hybrid. So that means that part of it might be rules based, might be logical, and then there also might be part of that workflow that requires that you want to use judgment and add in agentic steps. The point here is that you're doing it all in one platform. So you have centralized visibility. You can understand what the agent did, what the human did, and it's all in one unified platform or one unified experience. And we'll talk through this combination of agentic and determinism later on and how we implement it in production. Go to the next slide. So here's where we, I'll hand things off to Thomas to get into the meat of the presentation. But really want you to take from that is a lot of pilots are stalling, one, because they're applied to the wrong use cases. They don't fit into the workflow. But, also, a lot of times, they lack that that unified integration layer, ability to connect into multiple systems, the ability to give agents only the data they need access to and nothing more, and really operationalize AI in the right way that's governable. So, Thomas, I will let you take it from here. Alright. Thanks, Chris. One one thing to note on the last slide we just saw, every agent that you're gonna see in this presentation is a hybrid approach and that's not an accident. So I wanted to start off by kind of telling you why because I learned that the hard way. When I built my first agent team, I built it entirely on AI. There's no rules. There's no fixed steps. Every decision was up to the model. One of the most simple agents that I first built was just to run our AP mailbox. They read every supplier email, sorted the invoices, and decided what should happen next. And for a while, it worked pretty beautifully. Then it drifted. So on its own, it invented a new hold type that it called a business hold. We'd never defined a business hold, and it quietly put about 20 supplier invoices on it without telling anyone. So those invoices didn't get paid on time. And the first time I heard about it was when a key supplier came calling, which obviously is not an acceptable outcome. Nothing about that agent was malicious. It wasn't even actually wrong in an obvious way. It was doing exactly what I'd allowed it to do, which was anything. That's when I understood how dangerous AI can be if you build it wrong. So this chapter is about building it right. Three ideas, how agents work as a team, how much autonomy each process should get, and how an agent earns more of that autonomy over time. Most companies start where we started. Right? One agent per task with a very limited scope. So an invoice agent, an expense agent, a collections agent. Each one approves the simplest, cleanest work and then hands the rest back to a person. And that's a good start. But what each actually... Each agent actually is, it's just a really small slice of one person's job and an even smaller slice of your team's processes. So your AP team obviously doesn't just approve PO invoices. They watch the mailbox. They chase receipts. They answer supplier questions. They reconcile the subledger. Right? The agent does one of those things, and the steam... Team is still doing the rest. Plus now, your team has to manage the agent. So when you look at the bottom of the slide, with single agents, no one reviews the agent's work before it reaches the person, and the handoff between the agents wait on all of the people, the people on your team. So every agent reports to someone directly and nobody sees the process end to end. Now watch the same agents become a team. The gold cards are newer versions of the agents you just saw. What changed is that each one is now a team member that owns an entire role, not just one individual task. That's the power of an agent team. A single agent wakes up, does one task, and then just forgets it. An agent team carries a process from start to finish the same way your team does. And every team has a manager. In this case, Omar Outflow runs accounts payable. Amani Inflow runs accounts receivable. The manager agent reviews the team's work, reports daily and weekly KPIs, makes suggestions about how to improve the agent team, and escalates to a person. That person is the one who's accountable. So I'm sure people are wondering right now. One thing about the naming. Obviously, this is a choice, but for me, every agent has a human first name and a pun surname. And everyone says AI in its title. So So when a supplier gets an email from Ingrid Inbox, they know they're talking to an agent, which I think is extremely important. But here's the catch. A team of agents that can each do anything is just another way of creating my same business hold problem, but multiplied across an entire team. So the real question isn't how to build the team, it's how much a team member should be allowed to do. The first principle is least agency. If you work in security, you'll recognize the idea from least privilege. Give every user only the access they need. Least agency is the same idea for AI. Give every process only the autonomy it needs to do the job and nothing more. The scale runs from rules only on the far left to acting within limits on the right. Every APNA or process sits somewhere on it, and the question that decides where isn't how clever the AI model is. It's what does a mistake cost. So let's walk through it real quick. Payment remittances sits on the far left. They go out after every payment run. Same way every time. No judgment involved. So there's no agent involved. A rule does it perfectly, and it does it for a fraction of the cost of a model call. And that's worth saying out loud. Every step you hand to a rules is a step you're not paying AI to think about. Collections email said it suggests. The agent researches the account, drafts the note, but every one of those emails lands in a customer's inbox. So on my team, I still have a team member review every email before they get sent out to a customer. Invoice processing sits at accent flags. The agent codes and approves the clean invoices and holds all exceptions for the AP team to review. The agent does not attempt to handle the exceptions, though they will still give the APT notes on what they found. Cash application sits at Accent checked. The agent applies the payments and resolves as many of the unmatched payments as it can, as many of the exceptions as it can handle. Then people spot check the sample. And at the very far right, we have small balance write offs, which is acts within limits. When a customer's payment is short 15 to 20 in in whatever currency, it almost always is just a short pay... It's not a short payment. It's almost always just a bank fee that's been, you know, taken out. And we would always approve that write off anyways. So we allow that agent to clear that full amount, but that's the extent of its authority. Anything larger than that $20 threshold or or €20 threshold then has to go to a person for approval. Two lines to remember from this slide. Rules where it can be and agentic where it pays, and consequence sets the ceiling. Payments, credits, and contracts keep a person in the loop no matter how good the agent gets. Now go back to my business hold. Under Leased Agency, hold types are a fixed list that's enforced by a rule. The agent can't choose from that... Has to choose from that list. It can't create a new one. The entire incident becomes impossible, not unlikely, but impossible. And that's the difference between a limit in the prompt and a limit in the actual platform. So let's talk a little bit about earned autonomy. Leased agency tells you where an agent starts. Earned autonomy tells you how it moves. Think about how you treat a new hire in AP. Right? On day one, what what are you gonna let them do? Just draft stuff. And then someone else is gonna review it and approve. You know? Nobody gives a new hire signing authority on week one. Right? They have to earn it with evidence and a manager decides when. That's the same way that agents should work. An agent starts at suggest. It drafts the work and a person decides. So notice the counter on on the left. Right? The test case is passed. It starts at zero. When it has passed enough test cases, and you can choose how many that is, 50 is just as an example here, a person moves it up to accent flag. So now it can handle the clean work and flag the rest. It actually starts doing work. Then as more evidence gets added, more lessons get learned, then it goes to acts and checks. So now it handles the clean work, handles the known exceptions it's already learned about, and flags the rest. And this is also where I have my team start reviewing a sample instead of everything. And then finally, once it's passed enough test cases, it can act within limits where it decides on its own up to a limit that the platform enforces. So with the counter on the right, every one of those wrongs is earned with test cases. Where does the test case come from? That's from your team's corrections. Every time a person fixes the agent's work, that fix becomes a lesson and a test the agent has to pass before it can move up to the rung. It has to get through enough transactions cleanly to be able to move up a rung like that. And four rules sit along the bottom. You know, autonomy is earned with evidence. It's granted by people, never the agents themselves. It's enforced by the platform, and it's lowered if the quality drops. And that last one matters. If an agent starts to drift or starts to not be able to perform a process, it drops down a row. This isn't a guaranteed place it can always sit. It can actually move backwards. And now look at the dotted line at the top. That's the ceiling we just talked about. Some work never crosses it no matter how many test cases an agent passed. This is where your team pays off. In a team, the manager agent reviews every member's work before it reaches you. That review is a constant source of evidence. It's how you know which data, which agent is ready to move up and which one needs to come down. So now we'll move on to our second poll, which is how... Based on what I just told you, I'm curious how much would you let an AP or AR AR agent do on its own today? What would be your comfort level? Sorry. I mean, I don't see the new poll. Okay. I guess we can just drop it in the chat. Just curious. So I I see a d, which isn't too surprising. Seeing seeing a lot of b's and d's, and and I'm not too surprised by that. I mean, I think that a lot of a lot of people would be comfortable with b. I think that is kind of like the first place that... And, well, first place you should start should be a, of course. But I think, like, getting to b is a very easy place to get to as long as you have a clean process with what I normally like to refer to as the happy path. Right? If you have 80% of your work going through a nice clean path, then b is obviously handling 80% of your work. And that's obviously a very good, process to go... To look for for automation as well, especially agentic automation. So now let's get into more specifics of accounts payable. So our AP team is five... Or AP agent team, I should say, is five agents and a manager agent. And before anything else, one line that never changes is a person still approves every payment batch. Agents can do a lot in AP, but at least for now, and and maybe this will change someday though I kinda doubt it, I don't let them move cash out the door, and I don't really see that changing anytime soon. So this is how most of us, at least a a nerdy controller like myself, would draw out accounts payable. Right? There's 10 processes here from mailbox maintenance through month end and AP analysis and about 30 tasks underneath them. It's the standard way to decompose a process and it's how we draw ours. And here's the mistake I made at first. I tried to put an agent at every single one of those bottom level four task boxes. So an agent for mailbox, an agent for PO matching, an agent for coding. So that gives you 30 tiny agents and 30 handoffs for your team to review, and you're back to single agents working alone. So I... When I had the opportunity to rethink this for AI, you'll notice that a lot of these activities end up shifting around. So we redrew drew the roles and tasks. So now these are the same tasks, but regrouped in a way... The the way that the AP team actually divides their work. So Ingrid inbox owns the AP mailbox and the supplier questions. Own Voice owns invoice processing, PO and non PO, plus cutoff testing at month end. Thomas Talley owns T and E and the company cards. Rupert Rulebox owns AP policy and answers the... Everyone's favorite question of, am I allowed to expense this? And FIMI Forecast owns the disbursements forecast, payment reconciliations, and he sends all that information to our treasury team, which we're not gonna get too much into here. But that's another very interesting part of the team concepts is that they can actually communicate to other teams within your processes that make them even more powerful. But note that every agent has a communication task. Right? That is the agent sending a daily report to the AP team member about what the agent has completed that day and any items that need the team member's attention. And so as you can imagine, if you had an agent in all 30 of those boxes, once again, someone's having to review the work of 30 agents. In this case, I have a total of six agents, but really five worker agents that each report into one of my AP specialists. So my AP specialists get a more fixed amount of information to review. And so it's everything the agent did plus any exceptions the agent found, and they get an email every day that outlines all of this. Then you have Omar Outflow who sits on the very top. So he orchestrates the team, owns the close, and reports KPIs and escalations for the entire team to my human AP manager. So my AP manager can get the reports from all these individual workers and can review them in detail, but they also get one summarized report that over... That that explains everything that each of the agents has done for that day, and then a special report at the end of the week that contains all the KPIs for the week. And when you look at the very far right, that's where we still have the human in the loop. So a person, once again, still approves the payment batch. A person reviews what the manager agent communicates, and a person decides when the process itself should change. The agent can recommend a change based on what it's seen the worker agents do, but it can never actually make a change. So let's look at the details of one of these... Well, we'll do two of them, but two of these agents. So this is Envoy, which once again is our invoice processing agent. This agent is actually going live later today. Somewhere in there it says in build, but just finished the build, so just about to launch the newest version of this agent. So if you look at the colors, teal is rules, gold is the agent, and peach is people. So once again, this is a hybrid agent agent where you have just a small amount that the agent's actually doing. Most things are gonna be rules based. A rule finds the PO and matches it to the supplier, same way every time. Only then does the agent do the one thing that a rule can't do, which is read the invoice and look at the historical coding for that supplier. Then rules checks it again to the PO and to the limits. If everything's clean, the invoice is approved, no human has to touch it. Anything else that the agent doesn't know then goes to the AP team and then the AP team can review it and that's what then teaches or I guess the agent learns. And as you can see, as a exception goes to the AP team, they correct it, they teach the agent, and you can see the lessons brings down the exception rate that you see for that agent. So that is how your agent learns. So, you know, here's the kind of invoice, for example, that would break most pilots. If you picture an invoice from a supplier in Frankfurt, you have three currencies on one page, a PO number that doesn't match in what's in NetSuite, and no clean line to code it to. Right? So back to that, if you're using a pure AI agent, it's gonna try it anyways. Right? Because you can't... No matter how you've explained the guardrails, it's gonna do its best to figure it out. But here, the agent's gonna read it, see the rules that the PO doesn't match, and then it's gonna go straight to the AP team member to handle the exception. And then that AP team member is gonna fix it once, and then that's when the agent learns the lesson so that correction becomes a test case. So the next time you get an invoice from that same supplier with the different currencies, with the some different p o naming mechanism, the agent has learned how to handle that and then now can handle it. And that's how it's gonna end up moving up the ladder of the earned autonomy because it's gonna have less exceptions and have a higher correct run rate. And I'll say, you know, I know we have, a few people... I think actually the majority of the team member coming from EMEA here. You know, this picture is changing anyways with the structured e invoices that are becoming business requirements in your region. But here's why that matters for leased agency. When an invoice arrives as structured data, reading it becomes a rule, not a judgment. So the agent's job here actually shrinks to just the case that's genuinely needed. So that's not a threat to your, you know, plans to, put an AI agent in this process. It's just actually the cheapest upgrade you're gonna get because now it can be completely rules based and you actually don't even need that reading code step or maybe you need less of that reading code step. Let me go to our next agent. This one's live. It's illegal right now. This is Thomas Tally and it checks every expense report against our T and E policy. An employee submits in Coupa, which is just our system, but you're system agnostic. So you can use any system for this. Rules checks the hard limits, the amounts, the deadlines, the required receipts. Then the agent is what reads the actual receipts and checks them against all the lines that are in Coupa and compared to the policy, which obviously can evolve. If the report passes every rule, it's approved. If it breaks one, then it's held and a person decides. Due to this agent, our manual expense report review is down over 70% now. Right? Just from this one simple agent. So that's a lot of safe time for the team to where they can be focusing on more strategic activities, instead of, you know, reviewing everyone's expense reports. And look at the agency today on the left. Thomas sits at the acts and flags. So he approves what's clean and he never decides an exception. Every one of those decisions becomes a test case, which is how he'll learn and get to the next rung. And that's just because that's where this agent is today. Once we have enough test cases, once we've trained the agent enough on all the different exceptions it'll see, the manager agent will start identifying and grouping those exceptions so that then the... Then they can make a recommendation for new rules we can add to this agent so that then the agent can then handle those exceptions. And that's that earned autonomy, happening in real life. So let me go to accounts receivable now. Very similar. Accounts receivable has six agents and a manager agent. Again, the ceiling first. People still approve refunds, return authorizations, and bad debt provision. Those touch customer relationships in the balance sheet, so they stay with people. Same as we saw before, you know, here's the process decomposition for accounts receivable and here it is rethought as an AR, in this case, agent team. So very similar, you have Reza Reply who runs the AR mailbox. They triage every email and routes it to the right agent. That routing is what makes this a team. A contract question goes to Chidi. A payment question goes to Mila, and no person has to forward anything. Then you have Bianca Billings who owns invoicing and keeps billing contracts current, Shidi Credit who owns contract changes, swaps, and credit memos, Mila Match owns cash application, Dar Dunning owns collections, and Nikhil Numbers owns AR reporting and sends the cash forecast once again to our treasury agent. Amani Inflow, in this case, manages the team and owns the close. And to the far right, that's where you see the human in the loop again. The bad debt provision, approving AR refunds, and reviewing the the manager agent communicates. Notice it's a pretty short list. Right? And it's short on purpose, but it's exactly the work where a mistake is going to be very expensive. And that's how we always decide where we're gonna keep a human in the loop. So if we look at this for Reza Reply, which is another one that's live at Salego, you know, customer writes, you know, an email to the AR inbox. A rule matches them to their account in the ERP every single time, so there's the rule. Then the agent actually reads the email, pulls the invoice, maybe the contract if it's needed for the question, and drafts a reply. Then a rule checks the scope. Routine requests go out without a human touch. Like, think a request for, like, a copy of an invoice or or a w nine, and we'll actually see a few of these here in a second in the demo. But disputes, credits, refunds, those things always go to a person to review. They're too important to have an agent out there just kind of freely communicating with your customers. You know, and here's how we chose where to start, which I think is a great way actually to start on your organic journey, is we actually looked at what arrives at the inbox. The first thing I did was connect AI to my mailbox and then just have it go through all of the last six months of inquiries and categorize them and tell me what I was getting asked the most. And what we found was that four types of questions made up about 35% of the entire AR inbox traffic. These were, routine requests. They were low risk, obviously high volume. So that's what Reza was originally designed to handle and to be able to do on their own. Everything else, they route, at least for now, until we have enough, once again, real life examples of other exceptions that can then become the next most common exception that then will add and allow the the agent to handle as well. That's back to the earned autonomy and least agency really in practice. So let's look at collections. So once again, same slide, but this last one before we jump into the demo. So Dara Dunning, our collections agent, also live, and this is a good example of a long running agent. Every week, she works every single overdue account. We call an overdue account an account that's greater than thirty days overdue that does not have a special checkbox that is saying it's being manually handled by the team because certain customers we have manually handled by the team still. So those rules pick the accounts and the escalation step based on how age they are, and then Dar researches each one, scores the rift and drafts... Scores the risk and drafts the note. So in the very first time she sees an overdue account, she pulls together a lot of information about that customer from all of our systems to defy... To decide the best way to be able to communicate with this customer, to try to drive collections. But where she sits on the ladder, right, of earned autonomy is suggests because a person reviews every email before it's sent, because every one of those touch a customer. And the email caps are enforced by the platform, not by the prompt. So no matter how persuasive the email may be, no matter how great AI thinks it wrote an email, it's still... That cap still holds. It still has to be reviewed by the team. You know, and the value is not that Dara sends the email, it's that she does all the research related to that overdue account. She gives you the first draft of that email, but she's compiled all the information the team's gonna need to to to make a really good collection attempt. And as the team alters that email and sends it out, Dara reviews what they changed and then she'll actually learn more about how to write more with the team's voice and how our company, you know, drives her collections. So I'm gonna jump into the demo now. I guess let me go to the demo slide, pull it up. Here's a quick demo of our AI inbox, inquiries, automation, triple hit X and COVID. All of our customers send, air inquiries to an email box, and our AI specialists look at the emails and go to different systems and actually try and get the contacts and reply to them. So we built them an automation around that using AI. So for the demo part, these are sandbox environments, and we have a customer called Serigo Networks. APH serigo.net is their primary billing contact, and, John Miller is one contact. And, this contact, sent us a few emails. So if you see here, it's ar-debo@silvo.com is a mailbox, and, John Miller from silvo dot sent, emails to this. So firstly, he's asking for, their month end, account statement requirement. And one query around, say, low revenue and details. One is around getting their, open invoices, and, the last is around, updating their contact details. So he's saying, API and incoming@silico.net is the primary billing contact that we want to change. So he has, sent these four request to our AR demo. Yes. And then now I'll go ahead and run the flow. So this is our AR automation flow. We're just running it manually now, but in production, just setting a scheduled basis. It takes a minute to run, and if you walk through the flow. So what's happening here is, like, this automation is reading the emails on a delta basis and then extracting the email details. And once we have the email details, it's trying to get all the threaded, context around that email, and then it's trying to identify the customer domain. And once we have the customer domain, it's trying to find the customer in its way. And, once that's done, this, step is very crucial, the classifying quite it. This one is determining what the customer is asking for. Is he asking for statement? Is he asking for invoices or bank statement or contact update or remittance or whatever? So this is doing the classification part. And once this... This is the a bubble here. And once, the classification is done, so we have a batch here to go into different parts. So one batch is going into send back details where you're trying to get the bank details and draft response. The next one is updating contact details, send invoices, account statement, rate number and all that stuff. So based on different different tasks, we have different, tools or, internalistic steps here. And once that step is done, we are drafting an email and then moving that email to a process label so that, the AR specialist can be through there, from that label and send it, directly from there. So flow is just completed. So let's go to the mailbox and see. So here, if you see the mailbox is refreshed now, there is... There are... The processed emails are no more in the inbox. So the moment this process inquiry is labeled. So the AF, specialist comes into this, opens the email draft here. So here, a has generated the message, and then the automation has brought in the invoice... Sorry, the statement, from NetSuite and attached here. So the airperson will just review the statement and send the email from here. Same with WLAN details as well. So here WLAN details and JBM, bank details are attached. Invoices, so we have an open invoice here, and then the invoice, PDF is attached to the customer. And lastly, the contact details update. So we have the message here. And if you go into NetSuite and, refresh the customer record, so the primary billing from a p... At silico dot net will move to, APFA incoming at silico dot net because the customer asked. So we can refresh this and, see the change. So pretty much we are doing right operation. And, using the concept of, least agency where a is needed for determination of the, request type. And once we determine it, it's all, deterministic steps that take forward granting. The contact has been updated here. Yeah. He has a good demo. Thank you. Yeah. So as we can see there, right, that's everything we talked about in action. As Sandeep pointed out, almost all of those steps were deterministic steps. There's only a couple little bubbles there where you actually had to have the agent do any work. Least agency in action, also very important to point out, that also saves a lot of money. Right? Each of those deterministic steps is a step that you're not sending to AI. It's one less AI call that you're doing. So that workflow works great. And as we mentioned, it covers the... I think there's actually five or six now, but it covers the most common inquiries that we get to the AR team. So you can imagine instead of the AR team having to comb through all of those emails every single day, repetitive questions and filling them out, Now they can just go to the drafts folder and they can see, hey, what this customer asked? Did the agent respond right? And they can approve them. And I I would feel at at some point very soon I will have enough comfort and I think the team will have enough comfort in that agent to be able to actually have that one start sending those that... Those inquiry reply emails automatically. You know, it is a customer so we always wanna be a little bit careful there, but I think that we've seen enough use cases work well that, without any issues, without any errors, that we'll be very comfortable into moving that up to the next rung. So that's kind of, you know, that's the pattern we've seen. Now the question I think I get asked most in any demo or any AI conversation is, alright. It looks good, but how do I explain this to my auditors? So let's talk about governance. So two things in this last section. The questions you'll be asked by your auditors, your board, and probably your CFO, and how to prove your a... Your first agent in thirty days. So these four questions will come up, and it's better to have the answers built in before someone asks. The first is, you know, who approved this invoice, a person or an agent? So every agent action is logged with who or what decided. You can see the log on the right and shows you that OnVoice approved the invoice at 04:47, that res reply updated a billing contact at 04:49. Two minutes later, OnVoice tried to approve an invoice over its limit and the platform held it for AP to review. So that's once again the beauty of having the orchestration layer is that all of these agent actions, every action within every single flow that you run within the platform is going to be logged. So you have that appropriate observability, auditability, and governance. So then the second question, could an AI change a supplier's bank details? Only if the orchestration layer has the tool that allows it, and our AP agent and assistants don't get that. If you look at the last line of that log, someone in AP asked an AI assistant to change the supplier's bank account. The call was blocked because bank changes aren't in the toolset and it wasn't logged to a person's... And it was logged to a person's name... To a named person. Every other flow kept running. For anyone who's dealt with supplier bank detail fraud, that's probably one that gets your attention. Right? We have to be very careful about updating people's bank details without, you know, performing a callback procedure. So then the third one is, hey, a result was wrong. Right? So what changed? So all prompts and flows are versioned with a rollback to a known good state. That's how you'd fix a drifting agent in minutes, not in weeks. And you can see where it went wrong and what has changed and what caused it to go wrong. And the fourth is, you know, where does AI stop and the person approve? Well at checkpoints built into the flow with limits the platform enforces. So I mean I'd say once again we have people in the EU in the audience and so... Or me I guess in the audience as well. The, you know, e... EU AI act's strictest rules are for high risk systems. Right? And the digital omnibus was pushed back pushed back those obligations to December 2027. So there's still some time before that goes live. But I would actually say that most AR and AP agents are not likely to be classed as a high risk agent or high risk at all, though that obviously would be a question for your own counsel. But I'd still design the axe human oversight test, right, within any agents that you're building if you're in that region. You know, can a person understand what the agent's doing, monitor it, override it, and stop it? And if you can answer yes to all four of those, you're ready for your regulators, and frankly, you're probably ready for your auditors as well. So how do you start? Probably oversimplifying it here, but you start with one agent in in a month, in thirty days. Right? In week one, you pick one workflow with clear inputs and a number you can measure. Expense approvals, AR inboxes, those are great candidates that we talked about. They're how we started. Right? Week two, build the hybrid agent. Rules for the structure, one, two agent steps. Keep it as limited as you can and a person at the checkpoint. Week three, you run old and new side by side and you and you test everything. Right? Every difference becomes a test case, and this is where the agent starts earning its autonomy. And week four, that's when you can actually see your time saved. You know, you can measure the results, error rate, cost, and then decide which rung that agent should be started on. And line at the bottom is one I'd hold on to. Right? If the smallest version can't prove value, a bigger one never will. So you don't... I I wanna try to start with, like, hey. This is the biggest use case I can think of and it has this amazing ROI. Prove it out on something smaller first and then grow into that bigger use case. So then now we'll launch the last poll, and I'm curious, people's thoughts here. There we go. So poll is live in the tab. You know, what's the biggest thing you feel is holding your team back from building an agent today? I think I think I see the clear winner here already, which is clean data. Right? And I think I hear this one a lot. Right? I'd say it's the most honest answer on the board. You know, my business whole problem was partly a data problem too. You know, the agent had no fix list to choose from. The good news is that this is an integration problem, not an AI problem. And integration problems have known answers. You know, connect each system once and every agent uses the same clean path. The other thing I'd say too about the clean data in and out of systems is that's actually a good use case for agents. Most of those agents don't quite fit in these process diagrams that I showed today. They kinda sit a little bit more on the fringes. But you know as a... As an I guess x controller, you know, one of the biggest problems I always had was the data that was coming into my NetSuite environment. There'd be errors and those errors would cause my team manual cleanups, you know, non recurring month end journal entries, all kinds of issues. So what we actually did was kind of move up the process into the other systems and as data was coming through, the integration layer, we have agents sitting on top of that data doing quality reviews and checks to make sure that that data is clean. And that's something that you can use AI and, honestly, just a lot of integration, a lot of deterministic rules to do as well. But that... It is a very good way to be able to clean your data and to be able to to get ahead of it. And that's the other thing I'll say about agents and data is There's a lot of times, I know at least from an accounting perspective, we essentially were waiting till month end to do those data validations. That's when you do, like, your really big reconciliations and you call... You catch all those errors. There's no reason they'll do that anymore. There's no reason to wait that long. Now with agents and with AI, you can perform those, you know, data checks and reconciliations honestly on a daily basis. And I should have said that when I was showing the decomposition slides. One of the biggest changes you'll see with the agents was there was a whole section over there for month end tasks, and all that month end reporting task is pulled into the real time activities that the agents are doing. Where for AP and AR, you know, I have a soft close for AP and AR every single week now. Right? I don't have to wait until month end. Now I'll say the exception there is I don't do accruals. I don't do allocations on a weekly basis. There's just really no reason to do that kind of additional work, though you could. But I do have, like, the every, you know, every invoice is processed, every item is reconciled, every payment is reconciled, at the end of the week so that I have a very good picture of what is going on in my process. So what do next steps look like? Right? And here's what I do this week. You know, I'd pick one. I... Once again, AP and AR, that's what this webinar is about. There's a lot of good examples we went over here. Choose one where you can measure either the time saved or the hours avoided. Number two, you know, map your systems that it touches. List every system the process uses today. That's your integration scope. And it's usually where the real real work is. And step three, talk to us. Right? Book a working session with our team. We'll help you scope your first agent and then build the second one in half the time. Because the second agent doesn't start from scratch. It gets to join a team. So that brings us to our q and a. So, you know, thank you very much for spending this hour with us. You know, if you remember one thing, make it this. Give every process only the autonomy it needs and make every agent earn the rest. And Chris and I are happy to take any of your questions if you wanna drop them in the q and a, section of the chat. So... Alright. So I'll take the first one. How how complicated, technical would you say it is to build agent workflows on Soligo? So I would say that it has never been easier to build these workflows, period. But even on Soligo, Soligo has released two, big product innovations over this last year. One being Aura, which is a in platform agent that you can talk to that literally can help you build flows. You say, hey. I wanna integrate this system and this system. And then like magic, it starts work building the flow for you. A little more steps than that. Then the second one is the CLI, which is the command line interface. So all the agents that I was showing you in the AP and AR process, for the agent teams, I personally am building all of those using the CLI. And once again, I'm a twenty year accounting controller. I I have no technical background, no Now with the innovation of the CLI and Aura especially combined, I'm able to build these workflows and I'm able to monitor the errors in these workflows very easily. So going to the next one, how would you handle agent governance in your platform? So, I mean, I guess there's a lot of different aspects to governance. And when I really think of governance, the big one I think about is the actual observability, and who has the access to that agent itself. So I do limit who has access to the agents. Like in that example, hey, I'm I'm no longer the controller. Right? So I'm out of the actual functional controller role. The way that my my governance works is the agents will learn lessons and then they will communicate those lessons to the manager agent. The manager agent will then synthesize those lessons and recommend updates and changes to the agents. I am the only person as what we would call the automation engineer who can make those updates to the agents. Now that's a choice for me, but even if you allowed maybe your manager level to... Was technical enough and you wanted them to actually own and make those changes to the agents themselves, you could log and see within the platform any any changes that were made to any of the agents or any of the flows that those agents are made up of or any of the instructions at any of the AI steps. And you could create, you know, an audit log of that or a report that comes out to you as the controller that could see everywhere and anywhere as updating agents. That's definitely still evolving. So, really, the platform is giving us the ability to see and monitor any changes that are happening. Especially if any errors happen, we can always roll back to see what was that change. But as far as how you run that governance through the platform, I would say that that's still evolving. I'm still figuring it out myself. Yeah. And I was gonna add on, there's there's the control component. So there's only giving agent access to what it needs, the specific tools that it needs. That's a part of governance. We also have something called guardrails in the platform. So you can add those before or after an agentic step. They can make sure that the agent is adhering to your internal policies. If it's something like no PII data, no hate speech, whatever that is, you can add an extra guardrail, the flow to make sure that the agent is adhering to, what you care about internally or your own compliance policies. And then, yeah, like Thomas mentioned, observability is huge. Right? If you have one standardized place where you can see any changes that an agent has made, that's going to greatly help with your governance structure. I see the one, and I knew this one would come right. What's the cost per invoice for having it go through the system? So, unfortunately, as I've just launched the... Well, literally later today, I'm just launching the the AP invoice processing agent, which I assume we're talking about AP invoice processing here. I don't know that cost yet. And and I don't I don't wanna just kinda guess it. If we're talking on the AR side, our our AR invoice processing, is actually fully handled just through the integration with, we use NetSuite and we have Salesforce. And so that's purely deterministic, no AI involved, to get those, all that data over to NetSuite and get those invoices processed, through NetSuite. So, actually, this is a really good one, and and we didn't touch on it too much in this. But, you know, how is data and access to information protected? So one of the things I like the most about Saligo as a system that is... Was a real unlock for me in building agents is our MCP management. So I think if people know what MCP is, which is model context protocol, that's kind of the... How you can have agents communicate, with systems, the LLM, with systems, between systems, what tools they can see, what information they can access. With Saligo, you can manage the MCP to to where you can actually control exactly the data that they have access to. So I don't know if we have a lot of NetSuite users on the phone, but if you have NetSuite, NetSuite's MCP is actually excellent. It's, in my opinion, one of the better MCPs out there. It's also massive. And so if I just go and build something, let's say, in cloud code and I use my MCP credential to NetSuite and as a controller or CFO, I have access to all of the tools within the NetSuite MCP at right level. That means whatever agent I build is using my MCP and now has access to everything that I have access to, which can obviously be a big problem. So with Zaligo, you can go and actually create, I'll say... And, Chris, you probably have a better word for it, but a clone of that NetSuite MCP that then you can say, hey. I only want this... I guess the AR invoice... AR, inquiries agent. I only want this to be able to, you know, read invoice numbers, read customer contacts, and write to secondary billing contact. And that's the only three tools that it has available to it. So it can't see any other information across NetSuite, and it can't take any other actions across NetSuite either. Yep. That's a really good point. And in addition to being able to build, modify, add tools, you can also add, rate limits and IP filtering on top as additional security measures. So if you wanted to set a limit on how many calls the agent can make or if you wanted to filter out certain IP addresses, you can do that as well with our, third MCP management. And I'd say, like, the... How much is the cost of setting up agents? It's a hard one to quantify because it all depends on kind of the architecture of how you set up the agents. I mean, you know, at Allego, we have several different ways that pricing works. But the way that I I see my agent architecture going, you know, I am a little bit heavier on the flow usage for my agents, which that is, you know, a a factor that we charge for. But there's also other ways you could build these to where agents are even longer running flows than what I currently use. I kinda like mine broken down into smaller pieces. So, unfortunately, that's super variable depending on, like, how many agents we're talking about, how big those agents are, and how much of a process that they're actually gonna cover. And so the one about, handling headcount reduction. So I think that... That's a hard question. But one thing I will say is, you know, all those agents and agents teams that I built, each of those agents still essentially reports to a human. Right? And so like all of those AP agents like... So it's illegal. My AP team is actually really small. Right? I just have one AP manager and one one AP analyst. We used to have two AP analysts. Naturally, just through attrition, one of them left, and we didn't have to hire the backfill. Right? And so, I mean, I'm probably in a lucky position that a lot of time my management is or is forward thinking like that and is for, hey. We're still a growing company. Instead of hiring people, instead of looking at the f y twenty seven headcount plan and seeing, you know, hey. I need three more team members to be able to take on this additional volume. That gets heavily questioned about why can't you build AP to not have to hire... Or why can't you build AI up agents to not have to hire extra a people, AP people, AR people, whatever it is, and what your process is. So I feel like I have been very lucky in that way. But, you know, with a team of two, AP, an AP manager and an AP analyst, like, that actually is pretty much the perfect size team to run an AP agent team because once again, you have your manager that the manager is reporting to, you have your analyst that each of the individual agents is reporting to. But now that analyst can get so much more work done and we're also able to expand our AP team to, handle procurement as well. So we've been able to expand the duties of the team so they can be a little bit broader because they're supporting procurement agents as well. And that's kinda how we've been able to handle that. But, you know, it is it is a true conversation that you have to have, but it does take a while to build these agents up. The other one I'll say with team structure is I do always kind of view it that way as I still want, you know, like, a manager and and at least one analyst on every single team, but it's probably the more senior analyst that used to be on that team. And that's for two reasons. One reason is, you know, redundancy. Right? People need to take vacations. People need time off. And so you need those two people to be able to cover each other. And two, I still feel it's extremely important, at least in the accounting profession, that we're still bringing up, training up those, you know, lower level entry level employees to be able to understand the products or understand the processes, become those subject matter experts that then someday grow into those manager roles. And it's also important for that manager to have someone to develop and train because I think that person, you know, that man... Yeah. The management aspect of a, an... Of a manager, right, of training and developing a team, is an important skill set that I wanna build up within my teams as we look to expand and grow. Awesome. I think we've gone through all the questions, and we're also coming up on time here. So, yeah, without further ado, we can wrap up the session. I really appreciate everyone taking the time and and being so engaged. If you have any questions, you can feel free to reach out to me or Thomas. It's first name dot last name at siligo dot com. And I would recommend getting in touch with our team. I know we'll share resources after the call, but you can go to siligo.com, request a demo, request a strategy session, and really start to think about the best way to put your foot forward when it comes to AI. Alright. Thank you. very much, everyone. Bye.