Name: Custom Agents: Live build + Q&A
Uploaded: 2026-04-01T18:16:37.948Z
Duration: 33 min 42 s
Description: Custom Agents: Live build + Q&A

Transcript for "Custom Agents: Live build + Q&A": Hello. Hello. Thank you so much for joining us today. Really excited to walk you through a live demo of what we launched last week, our custom agent platform. I'm John Glasgow, founder and CEO of Campfire, and I'm joined by a special guest. Hi. My name is Brendan. I'm a product manager at Campfire. And Brendan did an awesome job leading the custom agent initiative, and so he's just kind of from his seat and product gonna walk through some of the things today. Upfront, just got a few quick slides, and then we're gonna go straight into a demo. But here's my ask of all of you. No. This is not April fool's joke. In the chat, go ahead and throw in some things you would love to see a custom agent built, and we're gonna go ahead and pick some examples live off the queue, for today's demo and make this very real and tangible for the audience. We'll have Brendan kick it off. What makes Ember different? Perfect. Well, thanks for that, John. So I guess a lot of people have been asking why Ember compared to a standard LLM of choice chat GPT Claude. So I thought we'd dive into that and tackle that. So first thing is Ember is purpose built for accounting. It's not a generic AI. And that starts with the system prompt. What the system prompt is, the Ember system prompt is a fine tuned on accounting knowledge and deep awareness of Campfire's databases, and it's far more accurate than a general purpose tool. So it's it is trained and it understands accounting, like your chart of accounts and things like that. And it is a specialized expert as opposed to some of that's, I guess, the the LLMs which are more of a jack of all trades. That moves on to the next part is the real key value as well with using Ember as opposed to via MTP and your LLM of choice is via code execution. So using Ember, the Ember agent will be able to execute code. And the reason why you want that to happen is because it'll allow you to process thousands upon thousands of transactions. Think of the MCP a bit like a keyhole where you're only gonna get a limited sort of dataset that can be pushed to the MCP while Ember has full visibility, can write code. So you're trying to pull a 10,000 transactions and then total up a vendor spend, Ember's the place to go for that. And then last but not least is Ember provides full audibility. Of course, I think that's really important. I'm a former auditor at the big four, so I know how important that is. So what you can do is during the prompt, it's gonna show the tools used that it's calling the chart of accounts, etcetera. And the other thing is the repeatability of that. So the orders that can come in, repeat the actual prompt, get the same results, which I think is very cool and something that's gonna be hard to do as opposed to just using a generic LLM. You said it well. I think the only thing I'm gonna add is around on the code execution side, a lot of folks are saying, hey. LMs are probabilistic. You know, they're bad at math. We agree. And this is why we have made some fundamental changes into how our AI works. We've moved to a deterministic state where, like, obviously, one plus one, the answer is always the same at two. And so when you go into a probabilistic world and it's using a next best character, which is how LLMs work, it can run adrift where how we're doing it now with and we'll show you in a second. It is writing Python and SQL and against the dataset. Not only can it handle much larger loads of data, but also the the one plus one equals two. We've got a kinda continuous repeatability on that output. And that has material with the custom agent launch, that has led to a material increase in in quality of the results. Great. And just to really dive in on Ember versus a generic AI. And what I mean by this is if you've decided to use Claude code and connect to Campfire AI via your MCP, and maybe you think, well, what's the difference? I can just now I don't need to go into Campfire AI. I'll just use my MCP. Well, there will be quite a significant difference and you probably still want to be in Campfire AI using Ember. And the reason why that is, first of all, it's Claude's brain. It is a generic sort of trained model while ours so think of it as a generalist while ours is Ember thinks like an accountant. It knows your what your trial balance should look like, intercompany eliminations, how they should work, suspense it knows what a suspense account. So that's all baked in to our system prompt and that's something you just can't engineer on the fly if you're using Claude. The next thing is that data access. So as I mentioned before, the NTP is very much just like a keyhole. You can only request specific things as well. You can't have full access to all of Campfire's databases and tables. So it doesn't know what vendor record connects to AP transactions and things like that. Well, Inbound knows the complete data model and when it queries these relationships, it can do it natively while your MCP will sort of be mapping it on the fly and that goes to sort of that more non deterministic where it's just gonna run on our next words and it could hallucinate while we'll mitigate that. Another huge thing is the ability on write back. So essentially, MCP is read only, you're only pulling from Campfire. You can't actually tell the agent to do anything for you. As John Glasgow gonna demonstrate very soon, you can most definitely instruct the agent in our Ember to do whatever you want. And that includes posting journals, setting up reports and things like that. So that's another key difference. And then on large datasets, the MCP isn't built yet built for bulk data processing. So it's if you're trying to pull your whole full GL and then trying to do calculations as John spoke to, that's basically not gonna work while using Ember through code execution. We can pull those numbers, we can do the totals, etcetera. And thanks to this new update where we use it, it is code based, the accuracy and the calculations are going to be accurate and they're gonna tie back to the underlying financial transactions as well, which is super important. So I think that's sort of there's there's a large number of reasons why you should be using using Ember and why this Ember and the custom agents launch is so exciting. Amazing. And I think we got one last slide. We're gonna talk through a couple ways to work with Ember. So you're gonna see this in the demo. We're gonna focus on the far right with custom agents. Ember chat, we've had that for maybe a year and a half now, and this is prompt based. You can ask questions, have a prepared analysis. You can have a draft journal entry for you. But customers were saying, like, this is super helpful. I want it to be a member of my team. I want it to work autonomously. I don't wanna be prompting it every time I need something. I just wanna get a Slack ping when something is done or when I need to be notified. And so that's why we launched agents. And we have prebuilt agents for a lot of the core workflows, as Brendan noted, built on top of our own foundation model, very fine tuned and honed, for accounting workflows. In custom agents, the focus of today, really for the long tail of tasks. I think people all the time are saying, hey. I wish Campfire AI did x, or, hey. I wish this one cost allocation or accrual or this custom cash flow report that the board's always asking for, and it takes me a ton of time. I gotta click download, and I gotta do tons of pivot tables and analysis in Excel to get in their exact format. We're done with all of that. Custom agents is picking up these workloads, and they're doing them on a cadence, and they're asynchronously preparing the output at a cadence of your choosing, sending it to you in email, Slack, the place of your choosing as well. So we're gonna we're we're gonna, like, give a super quick overview, and then I'm gonna dive right right into the demo here. You essentially create a custom workflow, and then we're gonna schedule and automate it. And then as things change, like, maybe that cost allocation, we wanna use different methodology. Maybe it's on a three month trailing average. On on the math, we wanna do a five month trailing average. Then we can go ahead and modify, refine, update, and build build custom agents for with your coworkers. But, with that, we will hop in. And we got one one question so far. Great one from Bert. Reading and interpreting contracts, creating rev rec and invoicing schedules therefrom. Bert, we hear you loud and clear even here at Campfire AI. And so we actually already have a feature for that that it'll GenAI will read and interpret the contract, and it'll generate, the downstream, revenue accounting and invoicing schedules off of that. We won't be demoing that today, but that is a great great callout, and it's something that, we we felt that pain, and and it is live now. So here is, where we start the work. And the the the chat tab here is actually gonna include the existing chat interface that we've had for a while as noted there on kind of that one, the chat based. And you can also chat with the to build custom agents. If you wanna use the prebuilt agents, then, that will continue to reside here. And so the green ones are the continuous agents we've built. We've got auditing agents that are constantly monitoring the work. It'll review your, all of your accounting policies here. It's like a fixed asset policy that it's, like, looking for materiality thresholds. You can see the policy we up we added. AR, AP, it's constantly matching them to cash as payments are are made. And then we we still have the on demand agents, our accrual subledger, auto drafting accruals for you. Quick example, if a contract expires and it thinks there's a renewal, but no invoice has been sent to us yet and we wanna draft, you know, it it will say, hey. It looks like we should accrue for the so then flux commentary, you know, for those that don't know, it's auto generating all of the flux. The the AI is writing flux at any dimension of your choosing, whether it's geo code, department. But, again, we're gonna focus on this this new experience. And. Go for it, please. Yeah, John. So a common question we get is, and that I hear a lot is, what's the difference between a custom agent and a pre built agent? And the way to think about it is a pre built agent is we've used specialized tooling and really embedded the accounting knowledge that's applicable to say, raising an accrual, which would be very hard to do it through natural language. So we've, in the back end, has really skilled up this agent and this pre built agent. And we know it's also using standard workflows like accruals, like flux, etcetera, that can just really straight out of the box achieve the goals you need. And it's just something I just wanted to point out to because it is a common question when I should use a custom agent versus what what's the difference. Yeah. If you if you wanna do a task that you see a prebuilt agents do not cover so here would be an example. We're doing a, and just to just to show the group, how I kick this off here was you can chat with it and ask it to kick off a custom agent here, or we have some, ideas for you. And so we're just gonna go ahead, and we'll do journal entries. And I'm just gonna click a reclass, And this is gonna kick off the prompt for me to start building a custom agent for reclasses. And so we'll we'll ask it to generate some questions. And and so we can see here we wanna do we'll just do a single transaction, and we'll send something from, like, above the line, below the line, and we're gonna run it for April, from from, above gross margin down into OpEx. Vendor, and it'll start building the task with us. And so just to show you one that I did complete give me one second. Okay. So here we did, a cost allocation, and we reclassed part of the philanthropic bill. And I just kinda came up with a random methodology, but just to show how it's more human like and not rule based that, it's looking at COGS percentage of revenue and then doing a reallocation down to marketing because we have a free tier, and so we're gonna send some of it to into OpEx. And so it's generating this journal entry for us off of looking at, you know, some some math in our account. It's gonna walk through how we got there, and then we can go ahead here's here's kind of the full steps to Brendan's point from an auditability perspective. And then I asked it to to create an agent. I can schedule it for later. We can always edit agents, as as noted there. And so here is, like, the full agent that got built, and we can edit them. But let's say, run it now. And I wanna get a Slack ping every time it runs, but I don't wanna approve it. This is really a member of my team. And so it's just gonna auto post this, but it's gonna send it to me on Slack because I wanna just know that it happened. And you can see here, no approval required, but you can always have it send it to you for approval if that's so if you wanna be in, like, a continuous autonomous state or someone that's just sending everything for your review, you're in full control. Great. Let me know in Slack every time this runs. And I'll just start getting us And I also it. sorry. Sorry, John. I just I think this is a very good example of why you wanna use Ember as opposed to just connecting via an MCP. Here we are literally posting journal entries, setting up an agent to do this on a repetitive task. So this is the the true value that you're seeing now rather than just connecting to OpenAI or Anthropic, etcetera. So just hit the proofs in the pudding, so to speak. Yeah. Exactly. I'm you know? And and so some a lot of folks are saying, like, what is the art of the possible? We're having customers. Everyone's a builder now, and so customers are are actually showing us new ways. Someone said, we weren't quite sure when to, buy sales tax software. We don't we don't know when we're gonna hit Nexus. And so they went and built a custom agent that when they hit Nexus in in a new region globally, a state, a county, or a country, they get a Slack ping. And and so it's it's really picking up a lot of things that folks weren't aware of, and that wasn't one that we were fully aware of that's now running on our own account that we get Slack pings every time we hit Nexis now. But, here's just a a range of things. So financial analysis on historical data and cash flow, forecasting, 2026 financial plan, revenue builds, you know, autonomous agents, forensic and audit. So maybe we wanna actually really put rigor into our process. We're a very lean team. We've got tight closed windows, but we want some help to make sure things are correct. You know, for the tech companies out there, revenue intelligence is always the big focus. And then, like, whether it's prepared by client list for an audit, you can have it, you know, generate larger financial packages or board materials. Or, you know, when we raised our series b, I dropped in the the diligence list in here, and it was it was just generating, it was just generating items, for for diligence. So the full range of activities. But we're going to and so here's an example as it works. But we're gonna take one off the queue from Chow and kinda I create a custom made to the categories of GL to expense category, for example, wages, commissions, which are bucket to employee related? So we wanna do an overhead allocation. And while that's running, we're gonna look at another use case here. Thirteen week cash flow. This you know, I've done many in my day. And so we can go in here, and we can kick off a thirteen week cash flow build. The good thing is it has such rich historical data that it's gonna go through geo code. It's gonna look for that quarterly $8 bill that you have and and look at when that that's gonna be paid. It's going below materiality thresholds that previously were just untenable, because it is AI. And so the work of 10,000 financial analysts are now at your disposal that we want to have a look at the historical data, build a thirteen week cash. It's gonna walk through, you know, our cash forecast here. Here's a walk. You know, it's going through the AR subledger, but it's beyond just, like, actual due date. It's saying when are these people actually paying us. Does Adobe pay early? You know, does that does that restaurant pay late? And it's gonna it's gonna go through your AP. You can give a guidance. Like, you know, we're actually gonna start to to pay AP late. And, again, it's all chat based. So I think folks are saying, like, we really need something that's that's easy for us to work with. And what I did here was I I gave it some external context. So I actually gave it our our headcount plan. And to Chow's comment there, I said, also look at overhead and employee level. And as we look at hiring, you know, I would like you to incorporate overhead, off this headcount plan. And now refreshed the cash flow forecast, and it's running fully loaded analysis, for us so I can make sure things are looking right. And then it updated our thirteen week cache. Now maybe you're only doing thirteen week cache monthly because it's a lot of work. If you wanna get it daily, hourly, literally any cadence of your desire, you can have this refreshing pinging you or just hop in here, at any time. And so I think, like, folks are saying, it's helping us rethink the financial insights, for our company. Yeah. And to add to that, Sean, I think the the sky really is the limit here. You can get as creative as you want. My background, as I mentioned before, I'm a former auditor, so I remember going out and doing the search for unrecorded liabilities to make sure everything's been raised correctly for the right period. That's something now where you can set up an agent yourself just before the auditors come to check that you've correctly accrued for all the expenses, etcetera, that you need to. You can do your own journal entries testing, flagging journals that are posted by say, not the usual people that should be posted, journals, round numbers, weekends, all of these things I used to do in an auditor. You can now literally have a bunch of agents. So you could have literally and it's gonna pull from the best best practice of the auditing standards, etcetera, as well where you can really be prepared before the auditors come. And I think they'll be very impressed with everything you you have ready now, thanks to Amber. Yes. And so we're gonna have it do a full overhead allocation on our q one income statement. It's gonna go through every vendor spend. Like, if it's a Zoom bill, it's gonna make a call that this is for everybody. You know, if it's a sales tool, it's gonna just, do an attribution for it over to, over to the sales team. So it's doing that deeper work. We're gonna go ahead and check on, another task. We looked at the the Anthropic one. Can also help you normalize data. So for those of us that operate globally or maybe you invoice in other countries, maybe at the board level, there's been a lot of FX volatility, and we you know, that's kinda showing up in our revenue number a bit. And so I'm gonna ask a term of currency impact on the financials so that, you know, now we can look like a public company now reporting to the board on, like, you know, a normalized income statement and then, like, a GAAP income statement. And you can see here, it went through and to Brendan point earlier, it it's going through, and it it is writing code. And so here's the actual kinda code that that is being written from a prompt perspective, and it's going through and running all the tables. And I think the keyhole comment Brendan calls it a keyhole. I call it a straw, but it's like a fire hose versus straw of data that, it now did a a FX impact for us and did a constant currency, income statement, but also wrote us some commentary. You know? So we've got a little bit of movement, and it's all immaterial. But but this is like, we can go to the you know, you get that random question at the board. Hey. I know there's been some, you know, euro volatility. How much of our revenue impact was related? Boom. You got the answer right here, and it looks you know, took it one minute to generate. Here's here's a totally different topic for folks that are audited, for folks that are in a SOX environment. Of course, there are rapidly growing number of public companies on Campfire AI that I wanna do a SOX test. And so it's gonna go through your account. Because we have the full context window, it's gonna look at segregation of duties. It's gonna go ahead and look for a documentation. Is there audit support? And it's gonna start flagging things for us. And maybe we wanna run this weekly, maybe just for the close or maybe just before audit season. These are the types of things that it's not just about closing faster. I think the bigger unlocks for for customers have actually been the things they didn't think about, which is, like, closing with a higher degree of of accuracy, going particularly those audited or public, but then also, the financial insights. And so we're going to hop back into this workflow we've got. And it did a proposed overhead allocation by department. It showed us the the updates here. And when categorized everything for us, it's doing head count as a percent for you can see some items and then actually doing some more custom math for others and giving us some, how it came to this conclusion for us to go ahead and have our updated yeah. So we just did a full overhead allocation together, based on but it's a great callout. This is the type of stuff that folks are folks are doing. Any other questions or prompts? We can do one from Eric. Splitting financial transactions into tags that align with operational structure. So, again, we're just gonna kick off a new one here. And this time, we'll actually create a custom major. Okay. So, again, it's not just the prompt, but it's a new member of your team where it's gonna be continuously doing this workload, in the background. Angel asked us, does Campfire have a limit on the number of lines a journal can contain? For example, is it possible to record 50,000 transaction lines? You'll be thrilled to hear we do have customers posting 50,000 transaction lines on a single journal entry today. We had to rebuild our approval workflow, actually, because it's a pretty large approval window. But but, yes, we are we are incredibly scalable. And so, please, we welcome 50,000 row JEs. Alright. While this is going, when we build an agent I'll go to one that we just built here. You can see here that we at the end, it created the new agent. And so these are all gonna be saved in the in the agent platform. And so we can search for the ones we've created, but here's a lot of examples, very custom ARR waterfall that the board wants to see daily cash flow updates. So it's like maybe we wanted to update cash at the end of q two daily and keep updating the number as we get closer because we're very closely watching cash for for the end of the quarter for some sort of debt covenant we have, for example. So I got my own analyst, John Glasgow analyst, doing doing tasks for me, things I I previously didn't wanna do. Overdue bills. So let me just, give me one second. I'll just show you what it looks like coming into Slack. Let's say I just wanna get, like, constant pings and updates that here would be so here's an example of that Anthropic one that I did. It's, like, I had it pinging every time it runs. And this was a different example. I actually have it going to the approval queue. But maybe there's a sales tax nexus alert. So here's one that looks like we just hit sales tax nexus in a new in a new state. And you can see just real time pings are coming to me again as as a member of my team. Yeah. And I think this overall what you're highlighting and you mentioned before, John, is thanks to the agents. What you're able to do now in a day is exponential. You were sort of I think accounts have been limited just through this year workload and things can fall through the cracks. Maybe you need to bring in consultants. Now with what we're seeing with Ember, there's a lot more you can take in house and still complete your month end and do a lot quicker with less with with a higher accuracy, etcetera. Thanks to these agents, whether it's preparing for an audit, whether it's getting ready for the board report, next release coming where Amber will actually be able to generate slides for you. So something that you may have spent hours, maybe days preparing for your board, your quarterly board report, Amber can now do in a very well curated prompts for you. Yeah. And so I'm, we're just building and you can see here, it's gonna it's gonna fill in all the gaps for you. We really ask it not to just make assumptions on your behalf. And so here, it's asking a lot of questions for it to fill in, gaps that it needs. And you can always chat with it, if it's something very custom you wanna provide. But what I'm gonna do here is I'll just do in app. We don't need to go to Slack for everything. And it's now and you can see the the tasks. It filled out a checklist of work. You know, again, it's running code, and we're kicking this off. One question would be possible, like, a customer port from a budget perhaps targeting a specific department, have a breakdown. I would like to see if Ember can see on the budgets, like, different breakdowns. Absolutely. And Campfire does have a full budget versus actual product built into it. So you can run it off of any dimension, including custom dimensions. Like, you have a very custom business unit that you like to budget off of. And Ember can generate the budgets on your behalf, and then you we can load them into Campfire BVA, for for the roll forwards. Alright. So here's all the steps it's taking, and we'll we'll give it a second to run. I know we had planned on kind of bottom of the hour, but so you can see here at the end, I can have it create an agent, schedule for later, run it now. Again, we can always edit agents. One other question around, can we create an agent that notifies us when a vendor contract renewal date is approaching? Yes. We obviously would need to know when the the service period is for the contract. The easiest way to probably do that is off of assuming you load it into prepaids, then we can always let you know when, like, a prepaid is expiring. And if it's just a monthly on, like, an annual monthly payment on an annual, then you can create we we can essentially pull in all of the the the service period data onto the onto the vendor, And then I guess you would create a custom agent off of that. Are there usage limits for Ember? Can you share agents between users? No usage limits. And, yes, you can share agents within the account with within the the customer account, not, like, with other people that don't work at your company. But that all all these are great questions, so would love would love to, keep the questions coming. Got a special guest, a secret CFO. Honored. Honored to have a secret CFO join us today. Thank you for being here. Okay. So we've got, while these are running, I'm I can go ahead and just show you. We can look at previous chats. And so it's all gonna be here on previous chats. You can see there's been a lot of different things from AR digest to, weekly executions, talking through our financial data, going into a board meeting, looking at tax code. This is probably for probably for sales tax nexus that was being run. Chocolate chip cookie cost variance, getting hungry, reading about, this this bakery example. But, ultimately, from an audit perspective, you can always go back. And so we can just, like, we can just grab, a prior one that was done until there's a dependency audit that someone was doing. But, again, all the agents, if you create one, it's gonna live in the custom. Otherwise, you can just perform a one off task and not bring it in. Okay. So we were going to check-in on I think that's all that we had for today. I'm just looking for more questions. Yeah. We'll go ahead and give it another few seconds. Otherwise, we're gonna we're gonna wrap up here, if there's any final questions. But, you know, my ask of all of you is is everybody's a builder. People are bringing us new ways of of of using custom agents in Campfire. It could even be done annually at $10.99 agent, monthly, daily. Really all sorts of workloads are being offloaded. The the prebuilt agents are for, like, call it the first 80%, like, cash matching and and accruals. The long tail, the last, call it 20%, or just call it the everything that previously was not feasible, whether it's, performing accounting, reviewing accounting work, or financial insights and analysis, or now seeing everybody move over to the custom agent platform. So with that, I wanna thank everybody for joining today, and, hopefully, you found that valuable.