From AI Spend to AI Budget: Building Your FY27 Number Live
A fireside chat with Andy Woods of DualEntry on what companies really spent on AI this year, followed by a live build of the FY27 number in Aleph.

Inside the session
It's budget season, and finance teams are being asked for an AI number for FY27. The category did not exist in the last three budgets, there is no benchmark to anchor to, and the fallback is a percentage bump on whatever got spent this year. That number does not survive the first question from a CEO or a board.
Andy Woods (Head of Product, DualEntry) opens with the actuals: what mid-market finance teams really spent on AI over the past year, broken down by vendor and category, including the spend that never went through finance. Sebastian Kontarovsky (Head of Solutions, Aleph) then takes that data live into a spreadsheet and builds a driver-based FY27 forecast on top of it: seats that scale with headcount, usage costs that scale with volume, category rollups, and scenario toggles. Open Q&A throughout, no slides-heavy pitch.
Designed for finance leaders and finance teams at Series A through pre-IPO companies: CFOs, VPs of Finance, FP&A leads, and controllers who owe someone an AI number before the FY27 budget locks.
Join in on the discussion to learn:
- What comparable companies actually spent on AI this year, broken down by vendor and category
- Which AI costs behave like headcount, which behave like usage, and why that changes how you model them
- How to build a driver-based AI forecast instead of a flat percentage bump on last year
- How to structure the AI line so it holds up in front of your CEO and your board
- Where AI spend shows up outside the finance-approved stack, and how to catch it before it compounds
- How to build scenario toggles for the aggressive and conservative FY27 cases
- Where to put AI in your chart of accounts so next year's actuals are actually reportable
Plus, much more...
Watch the webinar
Register to see how finance teams are building a defensible FY27 AI budget, with a live build from Aleph.
Chapters
- 00:00 · Welcome and audience polls on AI budgets
- 06:57 · Introductions and agenda
- 10:20 · Benchmarks: what companies spend on AI
- 15:15 · The AI adoption curve
- 17:40 · How AI spend splits by vendor
- 23:35 · Budget principles and building the baseline
- 36:31 · Forecasting 2027 AI spend by department
- 43:30 · Q&A: measuring ROI and AI contracts
Read the transcript
00:06 Alright. Hello, everyone. Welcome. Welcome. We're excited to be speaking with you all today. Yeah. Welcome to a joint webinar between Aleph and DualEntry. It should be pretty exciting. I think a lot of people have been asking, how much money should I should I be spending on AI? What does being AI pill mean? Am I spending enough? Am I spending too much? How do I think about this? Right? I think it's a it's a very topic of the moment right now. Yeah. Is that a discussion with many of your users, Andy? Yeah. Yeah. Because there's always this sort of fear of being left behind. Right? Like, am I am I doing enough? Am I am I spinning in the right direction? But then there's that equal fear, I think, among accountants in particular, but people in finance of, like but where is the alpha? Where is the ROI? Am I spending effectively? Is this a good use of money? Right? You know, am I just getting a bunch of AI slop or token spend? Right? So, I think it's always a relevant topic. What about what about you? Yeah. Same.
01:14 We have more and more CFOs trying to wrap their head around how to build a 2027 budget when there's very thin baseline, and we're you know, the landscape is changing so fast beneath our feet. So, I think it's gonna be an interesting topic, and, hopefully, a few of the things we shared today can help people, build their own numbers. I see we have people from all over, San Francisco, India, Nashville, Princeton, Germany, another Germany, Vancouver, North Carolina, California, Texas. So great representation. Thank you all for joining. Yeah. Yeah. I'm excited. I my ears perked up when you said North Carolina. I grew up in North Carolina. I live in I live in New York now, but, I always I always think about the old home state. Where are you calling in from today, Sebastian? I'm actually in New York as well. Fun fact, you're probably expecting to see, Albert here, founder and CEO of Aleph. You know, he has been taken away by a last minute trip to Costa Rica for engineering retreat. So he was kind of, perked up there where Wi Fi wasn't working as expected.
02:30 So, happy to sub in last minute. Dialing in from New York as well, but originally from Argentina. So a bit further than, the North Carolina. Yeah. Yeah. Absolutely. Absolutely. But it's fun to fun to go down there to BA. I was there for a wedding a couple years ago. It's a lovely place. Yeah. Love the city. Also, I feel so bad for your for your, founder. He has to hang out in Costa Rica. That's such a terrible place for an off-site. Right? Right. Terrible place to be in, man. Costa Rican rainforest. I hope he's pouring through that. Yeah. We'll pour one out for him today. Yeah. Great. So we have our first, poll question. Do you have an AA number in your budget yet? And seems like the audience is quite split here. Most have a rough placeholder. Nothing yet. For those of you who have a rough placeholder, do you wanna share in the chat a bit about how you've gotten to those numbers? Or maybe a rough sense of what the number per employee is?
03:46 I see a lot of people doing the last year's spend plus a percentage. Right? Like, that's the that's the classic. Right? Like, put the finger in the air, kinda assume that things will contend, in this place. And I think that definitely makes sense as a starting place, but everything is changing so much. It's kinda hard to it's hard to know if that's actually predictive. Yeah. Let's see. Do we wanna launch the second poll question while we wait for people to trickle in? Second one is, where does AI spend sit in your budget today? We have people, basically logging that up as software and SaaS, others as COGS, headcount, depending on the department. So curious to see where people are. No nobody's gone all in on agents. Are people here, and our budget is there, headcount wise? Seems like that's a few months away, so far.
05:00 We'll see. Alright. So we're seeing votes come in. 36, 37 votes. Let's get everyone on this poll. We wanna hear about all of you. And okay. Seems like most people suffer in SaaS or spread across departments. Great. So, again, hopefully, a few of the tips and tricks we're gonna share in a minute might come in handy to kind of spread this out and have a more granular budgeting. Do you wanna launch a third one, please, about what the average spend per employee is? I'm curious to see where people stack on that one too. Maybe it's also if anybody wants to be brave in the chat, if you don't even know the answer to that question. Right? Because you might not. Right? Depending upon how many how many different sources of data comes from and how mature your reporting is. Yeah. A 150 to 400 seems to be the mode.
06:21 Interesting. We'll see how that stacks up to our benchmarks in a second. Yeah. It's also it's this point in time, right, which is one of the things we're gonna look at in a minute is a bunch of data over the past twelve months, but, gosh, how much has changed in the last twelve months. Right? So, we'll talk about that and why it's a useful starting place, I think, in the in this conversation. But, obviously, there's many factors, that are at play here. Alright. So I think it's a really good segue into our presentation. Should we get started, Andy? I think it's, seven past of all time. Right? Yeah. Let's jump into it. Let's give the people give people the real data. Right? Like, that's what they that's what they want. Right? Sounds good. So topic at hand today is how to build an AI budget from AI spend. I know, spend is skyrocketing across every company. It's been doing so for the past couple of years. So it might seem like last year's baseline, is not necessarily the best indicator of what forward looking spend will be.
07:29 You might feel like you don't even have a good grasp on what last year's baseline was. So, hopefully, a few of the tools we're gonna share today might come in handy, to get your head and your budget around that. Right? So quick introductions. I'm Sebastian, head of solutions at Aleph. Been working at Aleph for the past year or so. Aleph, for those of you who aren't familiar with us, is an AA native FP and A software company that helps CFOs and finance teams at large kind of, automate much of the, grunt work that used to take, weeks or months, for finance teams and kind of focus on the real value added strategic work that finance is meant to drive. Andy, you wanna introduce yourself? Yeah. Absolutely. So my name is Andy. I lead the product team here at DualEntry. DualEntry is a AI native ERP or accounting system of record here. We're really excited about, really embracing this idea of being at the center of control and compliance and AI. Right? So for us, everything has to start in, being audit ready, having the right approval process, the right controls, But we want you to have that foundation and be able to enable the AI automation across the board. So we talk a lot of times about an accounting team of two or, where preparers and approvers each have their own agents.
08:56 They're able to automate things. They're able to take on these activities, but it's because, it's rooted in those control space. Right? So, one of the things that I think is is super amazing about the time we're in is there's never been better tools available to, finance teams, to CFOs, to accountants, controllers, etcetera, than there are today. And it's a really exciting time to be building this space because there's just so much that that we can deliver, at that intersection. So I'm really excited to be here today and to talk with you all about this because I think we're seeing these trends in different angles but in the same service of the same type of, problem. Awesome. Okay. So let's dive right in. Our topics for today, number one, what companies actually spend, and no one is in a better position than Andy and DualEntry to kind of give us a few of those, bureaus numbers and kind of, stack, benchmarks for where most companies are at. Then, from those benchmarks and your actuals, we're gonna walk you through how to build the forecast and what those forecasts could look like. And then finally, an open q and a section where, you can drop any question you might have in the chat, we'll be happy to go over them, discuss them live.
10:12 And, hopefully, you can get, a few of doubts answered, in real time. So without further ado, let's dive right in. Andy, I'll, know, see the stage to you and have you walk us through your benchmarks. Yeah. That sounds great. That sounds great. So, yeah, as, Sebastian mentioned, we're gonna look at, an aggregated across, all of our customers here. So this is no individual data here, and it is over the last twelve months, and we'll see some trends over that time period. Also, it'll be a very interesting conversation, I think, as we go through this about how much of those trends continue, how do things change, what to think about for this. But I think, like everyone, we wanna be rooted in what the data we have at least tell us about what to think about here. So I think this is probably not all that surprising to you all, but when we look over the last two years, the median monthly AI spend grew 30x, right? And it's an incredible number, right? Obviously, I think everyone has seen the valuations, the gaudy revenue numbers from Anthropic and OpenAI and across the board here, but we're really seeing this ramp happen, and it's continuing to go this route.
11:27 Now, within this, there's some more nuance about who is spending, how are they spending, how does that spending pattern changes that we'll dive in here, but certainly, it's a wild takeoff over the over the last couple of years here, which is, I think something that everybody has been feeling, but, certainly, the data backs that up as well. Yeah. Definitely seems like, sort of an arms race where nobody wanna be once we, left behind. Yeah. Absolutely. Absolutely. It's interesting because what we're seeing across the board is sort of actually smaller number per employee of this, like, 39 in the process here. That's the median number across all of the dataset. You'll see that it actually changes by company size, it changes by adoption, it changes across a variety of factors, but it's actually a lot smaller than what the audience had talked about here today. I think, to sort of bury the lead a bit, I think part of this is because the adoption has accelerated so fast in the conversation here, and then I think part of it is also because as we get into this, the real thinking about what is the right way to measure this and how to measure this is also evolving as we go. So all companies here, you could see this chart, we're gonna load the rest of the company size here in a minute, but I thought it'd be useful to be able to see the first piece before the rest of the data is on here.
12:46 There's the median number, which is that $39, the P75, P90, etcetera, in the process here. This helps us sort of aggregate the data, take a range of the data, kind of understand where we stand. Right? I think median is the most useful metric because it's the sort of one that lets us account for the outliers in the setup here. But it's also useful, I think, to understand the outliers because where the outliers is are really where the most AI pilled people are. And I think we'll have some stories about this as we continue on where some of that AI pilled usage is extremely virtuous, right, is a is a high ROI, right, and then some maybe not as much. But let's look at the other company size to see how that that plays out here. Yeah. So you see there's smaller companies with much more variability, and then as companies grow larger, it kind of converges. Right? Absolutely. Absolutely. And I think, intuitively this kind of makes sense. Right? Because small companies, like people are just trying to figure stuff out. They have like real pressing problems of like, hey, how do I actually get to my next funding round? How do I collect revenue today? How do I, go about taking these activities in the process here? But as they start to get a little bit larger, people are like, Oh, okay, well, how do we be good stewards of our money and think about this?
13:58 And then as they get much larger in the dataset, there's more sort of controls, approval process, maybe that somebody couldn't just try out one of these new tools as they as they come around the corner. Right? And I think that our data kind of reflects that that distribution here. Do you also think that, younger stage companies are also experimenting more heavily into replacing headcount with AI and that's, probably reflected here. Oh, yeah. Absolutely. Right? And I think, one of the one of the most exciting things about being at a early stage company, I think, as probably anybody who has made a transition from a larger company is, is that you just don't have as much process. Right? And so that can be really exciting because you don't have all loopholes and things to do, and it can also be challenging because process usually starts from a virtuous place. Right? So I think being able to be nimble and redesign it from the ground up and think, okay, now that this tool is in place, how do I actually use it? What does the old playbook look like versus what I should actually do given the constraints of the new tools? It's like, I think one of the most interesting conversations I have on our product team, as I'm building out DualEntry itself is just like, okay. The old playbook was like, you hire you have certain number of engineers, you hire a certain number of PMs.
15:06 Right? But does AI change that math? And I think the answer is undoubtedly yes. Right? But even those conversations themselves are quite interesting, I think. Awesome. Let's move to the to the next trend. How does the adoption curve behave? Yeah. I think this is probably not all that surprising, but you see this massive spin in the beginning because people get so excited. I think everybody has had that moment where they sort of just had the oh my gosh. It can actually do something meaningful. Like, I can use it here. Look at this thing I'm doing. Like, I've seen that so many times. Right? And then what you start to see in month three to five is maybe people go slightly too AI pills. Maybe the line between virtuous AI spend and AI slop starts to starts to kinda come into place, and people say, hey. Stop sending that AI slap to me. I don't I don't be a I think one of my favorite blog posts is, don't be a meat puppet. Right? Like, don't just be a repeater of, well, Claude told me this, so blah blah blah. Right? Like, use your brain. Be engaged in the in the process here. Right? And I think that that clawback sort of happens as people start to realize how much they're spending, how much they're using it, but then they sort of find this sort of elevated, outcome, which is like, okay. I see the value in it.
16:15 I see how it's useful. I see the places where I should sort of, be more careful about it, but then the spin kinda ramps at that point because other people start to see how valuable it is, and how many different places it can be applied within the company. Yeah. So this is probably the place around which finance controls come into place. Right? And we start seeing limits applied to spend and, maybe even model routing to make sure that you're not going to the most expensive model every time or a bunch of those things. Yeah. Absolutely. And I think, we'll talk more about this later as the as the presentation continues. But what also happens is the models themselves are constantly being developed, constantly being improved. You know, staying on the frontier is actually kind of a challenge in the process here of thinking about where to spend and where it happens. And, depending on which model you're using or your ability to switch between the models, can really impact you what your cost profile looks like here. Yeah. So I'm curious about our audience's take here. Have you seen this adoption curve play out in your company as well, or does it does it sound a bit foreign to you? Wanna comment in the chat with any insights?
17:30 I'm sure everybody has at least that that one person who got so AI pilled that, the work exploded from them, and then there was, a little bit of a clawback. Right? Yeah. We have Steven kind of jumping on the bandwagon. Seems like, this resonates. Alright. Let's go into the next one. How is AI spend, split? So this is one in particular where I think the past might not be a good predictor for the future because right now, all the models all the spend is around the models and the APIs and the coding tools. Right? I think almost everybody would argue we're at product market fit on the coding tools. They're quite powerful. Maybe not for every single task within a company, but many, many tasks, they're phenomenal. Right? In fact, I remember reading this piece about somebody had done a productivity test between people who used AI and people who didn't use AI as engineers, and they found that it was kind of a mixed result, and this was maybe six or seven months ago. They tried to repeat this test recently, and they couldn't find enough engineers who wouldn't use AI. Like, people are so adopting in this space, in order to actually run the test here.
18:41 I think what you'll start to see, is that more and more products are now being built from the ground up with AI in mind. They're being built to, help think about and, automate activities. I think they're being built to take certain tasks that are the monotonous, laborious task and replace those. So I'm very curious to see if this trend of the model APIs and the coding tools will still be the dominant one a year from now, or if we'll start to see this influence SaaS tools, other tools in more capacities. Yeah. It'd be interesting to see the trend play out. Right? I do expect, every SaaS company kind of embedding, this more heavily into the workflows. So, as usage based pricing comes into play, we could probably see, some token spend come from other tools which are, independent from the model providers. That's really gonna be an interesting trend over time. Yeah. Absolutely. Absolutely. Think that ties in nicely with the next piece of information here, which is that, you see the spend go up as the number of vendors goes up. And then intuitively, this kind of makes some sense. Right? Like, more places you can buy AI, the more budgets and more bills that you're gonna be attached in the process here.
19:52 But I think it also speaks to the idea that we've just talked about here, which is that there are more places to buy AI from now. There are more places to use AI in the process here. And so, inevitably, that number is gonna go up. To the poll at the beginning of this conversation, it's very interesting to think about how to allocate the budget here. Right? Because, it is a software spend, but if it replaces labor in an activity or means that you don't have to hire as fast. Right? Like, there could be a great ROI for your business here where you spend more in software, but you spend a lot less in net new headcount. Right? And so it'll be very interesting to see how the this trend continues. Yeah. Another interesting corollary is tying it back to what the average spend per company was. It seems most companies are in this camp right here, right, with only one or two vendors for AI spend as of now. And these are probably your kind of AI build early adopters. Right? Yeah. Yeah. Absolutely. Absolutely. And in fact, we don't have it tracked in this chart over time, but I think also probably unsurprising would be that over time, the number of vendors goes up, right? As people get more adopted, see the value, move this throughout the process here.
20:57 Yeah, for sure. This one I think is also interesting. It's funny because I think if we'd done this a year ago, people would be shocked at the result that Anthropic is the number one spend, right? Like at the time, it's hard to remember because things move so fast, but there's always this sort of leapfrogging back and forth of the different model providers and the process here. Anthropic, it seems, has really turned the corner in terms of market share and has established themselves well, but spin in cursor has gone up quite a bit. We're starting to see all of these tools. These are the sort of dominant ones that you would expect in a process, but there's some fast growing numbers that don't quite make the overall share on this chart, but are on the slope axis quite interesting to see coming down the pipeline here. Yeah. I think about a year ago is when, one of the, OPUS three models came out, and that's where actually On top of me, they jumped forward. Right? So it kind of resonates that over the course of the last year, it has led in terms of market share. But I agree that, probably eighteen months ago, we would have asked OpenAI to be at the forefront of it all.
22:05 Yeah. It's also probably interesting too, I think, on this that just probably a year ago is about when people started getting serious about setting a budget for AI, setting a contract, going from a, ad hoc spend to trying to secure an enterprise discount and work this direction. So I would be curious in the chat or otherwise, but I would expect a lot of this yearly contracts to be sort of coming up in Q3 and Q4. Bet a lot of these are sort of like in mind for people right now as they as they're starting to negotiate their spend. Yeah. For sure. Many of them might be coming up for renewal, and it's really an interesting time to kind of, get a sense of what you're planning on spending over the course of the next twelve months even though it's, naturally hard to do so. Yeah. And then maybe the last thought here too, which is that we're measuring spend, which is based on these activities here, but and that's very correlated with usage, but it doesn't account for some of the open source models. Right? And we're seeing a lot of people start to think about and use open source models as well. So I think it'll be quite interesting to pay attention to the usage versus spins access here as these open source models get more accessible to people, get easier to be hosted by your third party tools, more available from the neo clouds and the current cloud service providers.
23:19 It'll be quite interesting to layer on the open source conversation here moving forward as well. Yeah. For sure. Seems like probably a bit harder to track, but interesting to see if that actually excludes over time and how it stacks against these vendors. Yeah. Great. So let's jump into, how this applies to every individual company. All these benchmarks are extremely useful to kind of stack yourself up against, the broader market and kind of get a get a better sense of where each company stands. But, of course, it's hard to do that if, as a company, you don't really have the right baseline or the right forecast for your AA spend. So how should we go about building your 2027 budget? Right? There's first a few principles that we wanna anchor this to. First, of course, it's very hard to build a budget if you don't have a baseline. So we should always start with the actuals. Right? Even though that baseline might be a bit muddy or might not represent the future in the best possible way, having that baseline and knowing where to start from is is always helpful.
24:30 Then second, most companies are or many companies, I would say, are kind of modeling everything in one big AI budget. We think or part of, position here is splitting that by department, by budget, by owner, by project is actually very helpful in trying to get a more granular approach to budgeting and trying to peg ROI metrics to individual pieces of spend. More on that later, but, food for thought as sort of a general principle. Usage in grows independent of headcount. This is another interesting part. As we saw before, much of usage and spend is pegged to, API spend or product spend. So we'll go a bit into the drivers of this, but, having that, be forecasted completely independently of what your head count looks like is sort of a good principle to have as well. And allowance for experimentation, that's something that we've, kind of felt, firsthand. At Aleph, we, shared access to these tools with pretty much all of our team.
25:42 And, at the beginning, it's always useful, and people are kind of experimenting with it and seeing how they can, actually revamp their daily, their daily work with these new tools. Obviously, setting caps to what they can do with it and checking on that on an ongoing basis will be helpful because otherwise, you'll kind of run into wasteful spend as people who, aren't as familiar might be going down unproductive rabbit holes with it. Right? And finally, this is obviously an iterative process. It's very hard to get the forecast right the first time out. So reviewing this on a monthly basis kind of reforecasting and updating what your, outlook looks like, is is something we heavily recommend and something that's changing as fast as the AA landscape. Right? Curious what your thoughts are, Andy, and whether you'd add anything to this. Yeah. No. I think that's the right thing to think about. And, of course, there's the forecast in terms of how good are the models and how this works, but, I don't know if anybody else has been caught by this, but the models are changing their pricing plans pretty frequently, right?
26:51 Like, what's included in the Pro plan versus the Enterprise plan versus this? There's a usage gap here, right? So, you can think that you have a good thing set up here and then, be kinda surprised about the pricing and packaging changes that come down the line. It used to be, you know, you could reliably expect it's, know, two or three years before a pricing and packaging change, and now it feels like, if not every month, certainly very frequently, these things come out and can surprise you in this direction. Yeah. Good consideration to have. So let's start with the actuals. Right? We mentioned how having a baseline is always the first step. So how do we go about building that that baseline, and how do we get to split it across, different departments or use cases? Because, data that comes in from your ERP most likely is booked as a single line item. So the first step is actually grabbing these numbers in green here, which you can get from your ERP and, being able to split them across departments. So let's start with both costs, over the last couple of close months.
27:58 We should go into any of the, AI vendors you're using. So Tropic, OpenAI, Cursor, and kind of download the usage reports that they, and spend reports that they have to be able to reconcile these bulk, spend amounts to individual users and individual projects. Right? So that's gonna be a second step that we're gonna go over in a second. Hopefully, that gets you closer to this department based view. Then once you get this vendor export and match it to every user, you also need to match users to departments using your roster data. And, yeah, once you do that, then you can hopefully put this piece together. Right? So that's going to be the first step, getting your actuals, mapping roster to departments, reconciling the monthly totals to every user and department's spend. And once you have that, you can go into the second workbook, which is building your forecast. Right? So grabbing all of your department actuals, setting your assumptions for how heavily you're planning on hiring under the department, what you're planning on adoption to be, in terms of seats provision to every department, usage rates, that come through API, and finally, grab those, monthly department budgets.
29:26 And, in a sense, try to peg those to the ROI that they should be they should be returning. Right? So in a sense, when you're spending on AI for engineering, for example, might be harder to track ROI for that. You could probably build a few metrics around, PRs per engineer, but that's never a good metric. On the other hand, when you look at sales, for sales spend, you should definitely expect any AI spend in sales to have a impact in lower cap, for example. Right? So might be easier to kind of track those department specific spending amounts to the ROI you expect from each of them. So let me dive into each of the sheets, and make sure that we go over all the detail, if that makes sense. Unless, there's anything else you wanna comment on, Andy, I can I can switch and No? Can actually go to Excel. We've tortured these poor people long enough without showing them any kind of Excel or spreadsheet or real data, so let's give it to them.
30:34 Sounds good. So this is spreadsheet one. Right? And here, we're building this baseline. This is exactly what you were seeing in the previous sheet. And, how does this work? We have individual, tabs that you can see here that we've downloaded from each of the vendors. By the way, this is built off of off of disguised Aleph data. So many of the numbers that you're gonna see here has have some correlation with what we have seen in terms of actual spend. Of course, we've changed the numbers slightly, and we changed the names slightly. It's not like we hire, the entire Bennett family. But, some of these do have a real relationship to the trends at least or, high level evolution of what we see at Aleph. So here we have a single tab for every vendor. In this case, Claude, Jai GPT, Cursor. All of these should be available in each of your vendor portals. And here, I'm gonna be, putting in a shameless plug. We have developed sort of a skill internally at Aleph that you can use that will scrape all of these vendor portals for you, will download all these exports that have the, detailed spend or, detailed reconciliation by seat and, produce then the same worksheet for you.
32:00 So you'll be able to map, your bulk spend from your ERP to individual seed based or usage based spend from each of the vendors. Right? I love the, I love the sharing of the skill here. I feel like that's the sort of like if you bring it back to engineering, this open source idea or stand on the shoulders of others, or why should I have to figure out how to do this if someone's already figured it out here? I love this type of thing, and I think it's really great that we can share this in the conversation for folks because you know, we all get better when these skills and these tools come in and improve things for each of us. Yeah. This is gonna be, made available in the webinar notes after. We're gonna share it with everyone. So you're gonna be able to you're gonna be able to download this and use it freely to, gather your spend from each of the individual vendors. Hopefully, it'll be helpful to y'all. We have David in the chat saying, okay. Think it's approached by Lucas. Good. So once you have these actuals from each of your vendors, you, are probably able to map those to individual, departments by gathering your roster, right, and making sure that every name gets mapped to the right department, the right function, at whatever planning level you're gonna need it.
33:26 K? So once you do that, you can probably produce this baseline. And what's interesting is being able to decompose this baseline, into its drivers. So let's pick one department. Right? Let's pick engineering. We have 24 engineers on our team. Many of them have what we call the assistant seats, which are sort of the lower tiered version that many of these providers offer. Let's say the pro seat when it comes to the Andropic example. And then others have what we call the most advanced seats, right, which are typically more expensive. They have a bigger token allowance and so on. So this helps provide, or rebuild what the assistance subscription cost is and what the cost is for the coding subscriptions. Now on top of these subscriptions, as we mentioned before, we're gonna have, consumption usage. Right? So, coding seeds also come with, usage based pricing that you can consume, once you exceed the token limit for, the standard seat.
34:38 So that's also, laid out here. Right? And this kind of summarizes everything that's pegged to individual users, either through seats or through excess spend. Now everything that's further away from users starts, piling up here, right, which is everything that you consume via API for your product. And this can be either internal APIs For example, for the development environment for your product where engineers are actually building the features, and production APIs. So whatever your customers are actually consuming, through the production environment that's already live with those features that you've already shipped. Right? So this kind of brings together every piece of spend that's happening through your product. And together with everything that's back to your employees, you can have a pretty good view of what operating spend for a department looks like. Yeah. That's super cool. I feel like in particular, the ability to be able to tie that extra spiky usage place in is where I think a lot of people, go from v one to v two here.
35:52 Yeah. Exactly. And, of course, the pattern of how many peep how many, of your employees get, each type of seat, how much API spend you're seeing, and all of these things will vary greatly by department. Right? Obviously, usage in engineering is significantly higher than the one you can expect in other areas of the business. Right? So being able to have the right patterns, for every stakeholder here is what will hopefully drive this forecast to be a bit more accurate. All right. So once you're able to put together this baseline, now the question is, okay, this is how I'm doing in July. What can I expect going forward? Right? And this is where this second spreadsheet comes into play, where we're actually starting with what these baseline numbers are. So once you have the other sheet, you can simply port those over, paste them here. And we've, intentionally split this in two separate sheets so that it can be modular in a sense.
37:05 If you only want to build a baseline, you can get that sheet and do that. If you only wanna build your forecast because you've already built a baseline before, then you can use this sheet and layer in your baseline by basing it here. Right? So in the same way that we discussed before, the idea is going department by department and kind of layering in what we can expect in terms of trends to, build this forecast going forward. Right? So here are our actuals, and here we flip to the assumptions tab. We can see how these numbers change over time. Right? And this is obviously a simplification. We have January numbers and December numbers, and what this forecast does is extrapolate and sort of or interpolate on a on a linear basis. That's always never true. But, once you build out your monthly forecast per head count, you can have a more precise view on what this will look like. So we have, for every department, how this headcount number evolves over the course of, 2027. And we also have these planning assumptions. Right?
38:15 So for engineering, how many engineers do we expect will have a pro seat, in any of these vendors? Right? In this case, a 100%. How many do we expect will have a coding seat? Well, seems like 80%. What do we expect coding usage to be in excess of, the standard use that comes with every with every seat? And then finally, what do we expect internal and production API usage to be on a on a monthly basis? Right? I don't think this is anywhere near accurate. I would expect this number to be, significantly higher than our January number, but I leave that up to each one of you. And then finally, what we're layering on here is this sort of reserve amount. Right? Yeah. We mentioned before, an experimentation budget is always useful because you want people kind of working with this tool and kind of trying to use it for novel use cases. So setting aside that number that you can actually, deploy freely across teams without tying it to any sort of expected return, I think, is is good practice.
39:31 Yeah. Absolutely. Absolutely. It's very hard to predict where all of the value is going to be even just within the current state today, much less as the models advance. So I think having a having a little bit of idea around that already ahead of time will serve you well for being flexible and nimble here. K. So once we put together all of these inputs, then we can start to have sort of this comprehensive forecast, right, at the department level for all of 2027. We can see how, run rate is expected to increase. We're at a 100 per month now, and this is obviously expected to almost or 2.5 x over the course of the next year, including this reserve we were just discussing. So now is when the interesting discussions start. Right? As we were mentioning before, much of this 2,700,000 spend over the course of the next year, you could try to tie back to some ROI metrics. So for example, we're expected to spend a 100 k in sales. Right?
40:39 Can we reasonably expect this a 100 k to have some sort of impact and reduce CAC where this starts to be a positive ROI investment. Right? Do we have the right projects, that we're tackling, that make us, reasonably believe that this hypothesis makes sense? Is that same is that the same case in product, or does that sound more farfetched? So this is where, many of the interesting discussions with your budget owners are probably going to happen. Yeah. Absolutely. Just having data to have that conversation is a jumping off place, I think, is is gonna be quite powerful. Yeah. And finally, once you come up with this 2.7 number, which sounds like, sounds steep, you can start building, a few of the more granular numbers that also help peg you against the benchmarks that we were discussing earlier with Andy. Right? So, this 2.7 might sound difficult to track, but then once you peg it back to average spend by user is where, the benchmarks start to tell you a story.
41:54 So, for example, if we use the 51 to 200 employee range that we're just going over before, peer benchmark is $49 per employee as a median. What our last twelve months spend, was 313. So that kind of places us at, 77% percentile. Not surprised given how heavily we're leaning into these tools. But then when we go into full year '27, we're expecting this number to kind of skyrocket to, 1.5 k per employee. Right? Obviously, the whole industry will move, so the benchmark won't be flat. But at current rates, that would place us at, over ninetieth percentile. So, again, kind of triggers an interesting discussion. Right? Does that sound reasonable? Is this where you would expect your company to be, or do you think, you're going into, wasteful spend a little bit? Right? So having all this data together in a single page can at least give you the instruments to have a productive discussion around it.
43:07 Yeah. Absolutely. Absolutely. I think, none of us quite know what is is around the corner here, right, whether or not you're AI pilled all the way in or a tumor or, somewhere in between. Right? But, I think having data, being able to back it up, being able to tie it, being able to look at the last year and see the ROI, see how this trend changes over the next six months, this will give everybody a much more informed place to have a productive conversation. Yep. So let's answer a few questions that we've trained in the chat. Are we getting access to the spreadsheet? Yes. We'll, post this together with the webinar notes on the on the web page later for you to download together with, the skill that allows you to get all of your actuals from your AI vendors. So, hopefully, this will come in handy. Didn't quite get the mapping to ROI part. Can you clarify? Yes. So, what we're going over is once you have these numbers split by department, it starts to, make discussions around ROI easier.
44:15 Right? So for example, in sales, going back to the example I was just discussing, do we expect this 100 k spend over the course of the year to impact CAC that we have for customer acquisition in a meaningful way. Right? And if so, what do we expect that impact to be? And is that a positive ROI investment? Obviously, it's, projection into the future, and we don't really know for sure what the ROI will be on that. But at least having a productive discussion with your head of sales around, hey. You know, this is the impact we expect to see. And if six months in or, four months in, whatever it is, we are not seeing that number trend in the right direction, then we'll scale down spend in AI for your department because, that that investment is is not paying out the in the way we expect it. Right? So that same discussion can be had at the department level or even more granular going into teams and individual projects where you're actually deploying AI.
45:23 Yeah. And I think, if you if you marry this back to the adoption curve we looked at earlier, one of the things that is probably a valuable way to frame this is it's not just that you're sort of clawing back budget, but that you're helping the head of sales justify their budget, helping them pressure test, helping them be good financial stewards of the company, and will probably result in actually more spend over a long term here because they have to go back and make sure that all the AI spend is useful and delivering value and showcase those receipts to you, so to speak, which then lets them, have the appropriate checks and controls on the business itself. Steven is asking, how do you estimate those numbers with poor visibility into pricing models? And that's definitely a tough one. Not sure if you have any ideas about that one, Andy, but I think, pricing is one aspect of it, and usage is also, extremely challenging to forecast. Right? All these use cases are exploding, so it's very hard to see, what we can expect, consumption based usage to be. So both pieces of the puzzle are kind of moving under your fit your feet a little bit. I'd say in those scenarios, kind of playing some sensibility around them and seeing where you can expect to be in different scenarios is probably a good approach.
46:45 But curious what your thoughts are on the Yeah. I think that makes sense. I also think the industry as a whole is gonna put more pressure on the model service providers, right, both from these renewals we were talking about earlier, budgeting, forecasting, trying to build up a more predictable business from an unpredictable spend. And then I think the natural pressure as the open source models come out, they become available in other places, is also gonna put some of that desire on the model providers to give you better tools to really justify that Right? They don't wanna help you make the case internally about why their spend is useful, and so being able to allocate that, talk about this, plan around this. I think all of these types of things are what happens as a as enterprise products mature. Yeah. Fully agree with that one. Alright. So wrapping up, each department, as we mentioned, has its own drivers, and trying to get these right is sort of the crux of building a reasonable forecast. How many people do you expect to have an assistant subscription? How many people do you expect to have a coding subscription?
47:49 And what the excess usage for those is expected to be? What your APIs are expected to consume both internally for feature building and testing and externally for what users, are actually drawing. And, hopefully, that helps you build quite a comprehensive picture and, put that, the budget together. Caitlin has a question around whether APIs yeah. The answer is yes. API usage, you would allocate to COGS. On the external spend, at least, for sure, a 100% because you would, that's, customers using your product. So definitely COGS. On the internal usage part, you could argue you could also be, CapEx because it's your engineers building and testing the features they're building. So it could go either way. Yeah. I've seen people do a percentage or, try to allocate, API usage from the dev environment, those types of things.
48:55 Yeah. So, what changes the number? Hiring, adoption, usage, a few other things we've already discussed. Hopefully, having a clearer view of what each of these look by department can help you narrow down on the budget number and feel like you have a bit more control over what AI spend looks like as opposed to, a big blob of unnamed spend. Yeah. I think I think those all make a ton of sense. There's a there's one other idea that I wanna put everyone's head too, which is this, like, Pareto frontier idea, which is, this curve along two attributes about where are the best set of choices. So you can have a more capable model, but it's more expensive, a less capable model, it's less expensive, but you would never wanna use a model that is less capable than another, but is more expensive. So you wanna be on the edge of that curve, and then depending upon the use case, go this direction. I think what we're seeing right now is all the model providers are really competing to be at some point on that curve, right? So one of the things that you can easily do is make sure that you're actually using the right model on the right point of the curve.
50:04 For the activities that really need more advanced thinking and capabilities, use the best version of those at the price point. For the ones where the activities need less, move it back in the process. And I think that's part of what makes this budgeting and forecasting hard is that that Pareto frontier is changing so fast. What is true today is not necessarily true a month from now or two months from now. And so having some able ability to be flexible as you switch vendors, to be flexible as you switch model providers within a vendor, to even just think about for the task, the activity, where things should go. These are all questions that I think parts of your company should be able to answer for their particular activities, but that you should be armed as a finance professional to be asking are we really using the right the right combination of ideas as we're picking the vendor, as we're picking the models, or picking the tool here as we as we think about things. Yep. That's all great food for thought. So let's jump into q and a. I'm sure people in the chat have other questions they would like answered, so let's hear them. And, hopefully, we can cover a few of them in the couple of minutes that we have left. Bill is asking, how have you seen firms attempt to measure ROI for coding AI spend using historical data, Merge request volume, take a completion time, etcetera. Any good ideas on that one, Andy?
51:27 Man, this is this is always the hard one. Right? Because people talk about lines of code or PR is merged or features shipped, or all of these things, right? And all of those things are true, like they're all productivity metrics that you can attach here. But like what really everybody wants to be thinking about is outcomes for our customers, right? So it's like whether or not you ship one line of code, or a thousand, or one feature, or 10, but it's like did you move the needle for the customer problem? Did you solve it in this direction? We think about this a ton. I think it's hard to measure, but I think you can measure though in the aggregate is basically, did your business metrics that you pegged yourself to, did they improve as the spin goes up? Right? You're not gonna get causation. I know all of us want that causation in the relationship here, but you can get correlation, I think, from these activities, and I think you can sort of peg questions around this. It's also One of the things that is kind of true in the model providers is that they tend to produce more chatty code. So there's more lines produced, more activities produced, these types of things. And so it really skews these productivity metrics that you would maybe have used historically to measure here, which is why you kinda have to tie it back to that that sort of outcome oriented view. You can even create the wrong incentives. Right?
52:34 Because, you wanna incentivize your engineers to kind of, default to simpler, shorter solutions as opposed to bloated, solutions that have, thousands of lines of code. So actually making that the productivity metric could, orient your whole organization in the wrong direction. Yeah. I think we've seen that happen a bunch. Right? Like, Meta had a leaderboard, I think, at one point of how much token spend people had done. Like every place that like has one of these immediately walks it back because, you could just literally put a cycle of a loop of spending tokens here. Right? Yeah. Exactly. So I think trying to tie that back to like business metrics and user outcomes as opposed to productivity metrics is the way to go. It's very difficult to accomplish, but, it's pretty much the holy grail of engineering productivity. Oh, yeah. Oh, yeah. Abby has, if I'm using Aleph, would that count as one of my AI spend line item? Right now, I would say not because, the way Aleph prices is we are a SaaS subscription, so we generally not count towards AI spend.
53:45 We're reserving the right to change that in the future as our AI features take off and people start using them more aggressively, but that's the lay of the land right now. And I think that makes sense. Right? Like, you see this across various tools where, there's a software, which is, you could kinda think of it as a platform fee or access or capabilities. Then there's usage based pricing. I think, lots of things have historically been sold as credits and drawdowns and things like this, and now, of course, that model is expanding as people try to get into the token flow and wrap the tokens themselves and go this direction, and there's actually good reasons why you might wanna separate these things out as a business or pricing model because then they can be attached to different budgets. So if the AI model starts to take off really well within Aleph and you're able to provide real value, having those as a separate line item could then attack the labor budget as well as the software budget. Right? So I do think you'll see more complex pricing and packaging models, and it'll be because they're trying to deliver value, but also capture value against different budgets in the process here, and the activities relate back to them in that capacity. So we think a lot about this of like, we wanna automate stuff and the activities to free up our accountants to do more productive and more insightful oriented things.
55:01 And the idea is like, if we can really drive that down, then they can spend their time on the more value generating activities here. But part of that comes from really truly proving that we deliver on automating away a lot of those activities. Yeah. That makes a lot of sense. And I think it actually, tethers, pricing more closely to value for users. Right? Probably the traditional seat based model is kind of breaking down in this, day and age where more and more of the work is actually being, basically offloaded to agents as opposed to people. So who cares if you have, more seats or more people involved in your platform? What you really care about is, are people using it and therefore, deriving value from it. Right? So really breaking down that seat based model into something else is better for both, vendors and, end users. Yeah. And seed based models are always kinda funny because they sort of encourage you to have fewer people using your product. Right? And of course that's not what I want. I want more people using the product. Right? So I think aligning those incentives is actually an opportunity for both the vendor as well as the consumer. Yep.
56:08 What types of AI tool contracts are you seeing companies sign more of these days? Pay as you go, annual contracts, spend commitments? I think it depends on the company size, it would be my first answer to this, and maturity. Right? There's a lot of value early on in being flexible. Like, even if you know you're paying more on a monthly basis, having the flexibility to pick up and go somewhere else quite valuable. I'm trying to figure out your runway, your fundraise, this kind of stuff, your series A, your B, I think there's a lot more direction this way. As you start to really adopt these tools, as your company gets bigger, as the things start to happen here, there's real savings that come with annual contracts, with spend commitments, and escalated support lines, and all the kinds of things that tie these together. I think what people get nervous about is if I sign this big contract with Anthropic and OpenAI leapfrogs them, am I stuck spending on a thing I don't wanna spend on? And I think the answer is that's probably The things are changing so fast that that's probably not that likely to happen that you're stuck on a thing you don't wanna be on, but I think also, even if the model providers got like not a single percent better than they are right now today, we all have so many places where we could get value out of them that we're not realizing them yet because it's still so early in the adoption and usage that like you could easily spend that money in value creating activities.
57:30 So I tend to think that what we'll see a lot more of is that transition from the usage based to the spend with commitments and the setup here to get that cost control, get that savings locked in, just because we know there's so much value here. Yeah. That makes a lot of sense, so don't think I have much left that one. Do we have any other questions in the in the thread? Otherwise, I think we're pretty much on time, so, we could get back two minutes ahead of your next meeting, so that you can, grab a glass of water. This was great, Andy. Thanks for, thanks for being a part of it. I enjoyed it a lot. Think I it was a productive conversation. So, hopefully, all of our attendees got some useful nuggets of insight here. And, yeah, we very much look forward to having you in our next webinar as well. Yeah. Thank you. Thank you for having me. I always love to have these types of conversations, hear how people are using things, what they're doing. It's great to see being able to give back things to the community to help people advance. We all kinda can learn from each other here, and I don't know. I'm just really excited about the direction that we're all taking here in this direction.
58:41 So thank you for having me. Yeah. Exciting times and our pleasure. Alright. Thanks, everyone.

