---
title: "Transcript: How to Measure AI Impact: From Token Costs to Board-Ready ROI"
webinar: "How to Measure AI Impact: From Token Costs to Board-Ready ROI"
date: 2026-06-17T14:00:00Z
duration: "1 hour"
speakers: ["Charlie Cowan"]
topics: ["How token costs work and what your AI programme really costs to run", "Live demo: building an AI impact dashboard for your CFO", "Real examples from companies proving AI ROI to their boards", "Key questions to answer before your next board review"]
recording: https://youtu.be/ARX1zqxaT_w
canonical: https://kowalah.com/resources/wednesday-webinars/measure-ai-impact-token-costs-board-ready-roi/transcript
source: Kowalah Wednesday Webinars
---

# Transcript: How to Measure AI Impact: From Token Costs to Board-Ready ROI

**Recorded** 17 June 2026  
**Duration** 1 hour  
**Speakers** Charlie Cowan

## About this session

Every AI buyer is being asked the same question: what's the return? This session shows you how to measure AI impact, from token-level costs to executive-ready ROI reporting.

## Topics covered

- How token costs work and what your AI programme really costs to run
- Live demo: building an AI impact dashboard for your CFO
- Real examples from companies proving AI ROI to their boards
- Key questions to answer before your next board review

## Recording

https://youtu.be/ARX1zqxaT_w

## Transcript

**Charlie Cowan:** Okay, hello, hello? Checking you can see my screen, all looks good. Let me just go into present mode, there's a thumbs up, brilliant. Well, welcome to today's Kowalah Wednesday webinar. My name's Charlie, and we get together every Wednesday at the same time to spend an hour talking about something that is top of mind in the world of AI. It might be stuff that we are using or seeing in our own organization, or in some of the client projects that we are running. It's such a fast-moving space that it gives us this time to just take a little bit of a breather, take a look at something, and then hopefully you've got something really valuable to take back into your organization. If you're joining live, thank you very much. Welcome, lovely to have you here, and if you're watching on YouTube, then just the same. Thank you very much. We're glad to have you with us.

**Charlie Cowan:** We're gonna go through, as we always do, on our Wednesday webinars, three main topics. So, firstly, the topic of the day, which I'm gonna introduce you to in a minute, is how to measure AI impact, and we'll take a look at that. And then, we're gonna look at AI in the news, and there's one hot topic that I want to talk about today, which you can guess what that might be. And then we've also got a Q&A section. But I would encourage you, if you've got questions and you're watching live as we go through, feel free to ask those questions without having to wait until the end. So, when we're talking about the ROI of AI, please do dive in. You're in a Zoom event, so you've either got the chat widget, or there's a Q&A widget for a more formal question.

**Charlie Cowan:** I'm in a luxurious, spacious WeWork booth today, so I'm only on one screen, so I'll endeavour to keep an eye on what is going on. But my colleague, Caitlin, is online as well, and she will keep me on track if I miss one of your questions. With that, let's move on to the topic of today, which is how to measure AI impact. And this is coming up in every single client project that we are working on. As the adoption of AI goes through the roof, as companies roll out to more and more users in their teams, as they start applying more and more workloads to AI, the costs are going up and up and up. And as I say often to the finance teams at our clients, 6 months ago, you may not have even heard of a token, and in 6 months' time, tokens could be the biggest line item on your P&L.

**Charlie Cowan:** So you better gear up pretty quickly to figure out what they are, how to manage them, and how you might manage that cost item on your P&L, so that you are able to lead your company into the future. I saw this image on X, formerly known as Twitter, the other day, and I thought it was quite funny, and it lands very well when I'm speaking to finance teams. So, EBITDA, but with a new T in there. Earnings before interest, taxes, tokens, depreciation, and amortization. It's a little bit tongue-in-cheek, but I can absolutely imagine this is something that finance teams, CFOs especially, are starting to think about, and companies might even start reporting in this way. So let's talk a little bit about some of the top questions that are coming up from CFOs and from their finance teams.

**Charlie Cowan:** So the number one question, and we're going to go through each of these four through this session, is, how do I forecast this thing? I know this thing is coming. I know that it's something that, as a finance professional, I need to be on top of. How do I look out across the next 6 months, the next year, and forecast what is about to happen? Once I've built out that forecast, how do I account for this? I'm sort of joking there with EBITDA, but how do we account for this? Where is this token cost gonna sit on our P&L? How do I limit this line item on my P&L? We cannot be in a world where there's just unrestricted costs flowing out to whether that's Anthropic or OpenAI or anyone else. So how do I limit that?

**Charlie Cowan:** And then, finally, how do I report on this back into my company, to the board, to our investors, to shareholders? How do I report on this accurately? So before we get into those four topics that we're going to go through, this post was put out on X, just, I want to say, like, two days ago, by Satya Nadella, the CEO of Microsoft. And he talks about this new world that we are moving into, where there is both human capital, so the people that we employ, the knowledge, the experiences that they have, the ways of working that are in their brains, and then this token capital, which is the phrase that he uses for describing all of the workflows, the agents, the ways of working that have been built into your AI program.

**Charlie Cowan:** And he's making the point here that it's the combination of human capital plus token capital that is going to allow companies to dominate their sectors. It is not that just as you acquire token capital, you can remove or dramatically reduce the human capital. So I very much encourage you to read that article. We'll give you the link in the show notes. What I would point you to here, look at the number of views, 64 million. I think this may even have been his first post, or first article on X. 64, 65 million at the time of me pulling this slide, so it's a good article to look at. Well, what I think is useful from our perspective for the topic of this conversation is that every employee now is a capital allocator.

**Charlie Cowan:** Whether you have 10 tokens to your name, or whether you have a million tokens to your name, you are now going to be allocating that capital to your various different use cases. And we're going to talk a little bit about how we help our employee base, our human capital, to deploy those tokens in the right way that drives our company in the right direction. Here's another graphic that I think is very relevant right now. We've been talking about this graphic for the last couple of years, mainly to people teams, but it is now very relevant to finance teams as well. So the concept here is that, up until recently, any company in pretty much any industry could be described as a pyramid.

**Charlie Cowan:** You have a CEO up at the top, you then go down to a senior leadership team of CXOs and VPs, you then go down to your management layer, and then you go down to the individual contributors who make up the base of the pyramid. But increasingly, what we're seeing now is that companies are going to be hollowing out those bottom two corners of the pyramid and turning it into a pentagon. And there's two ways that this is happening. One is the collapsing of some of the middle management layers. We've seen posts recently from Block. They're the makers of Square and a few other financial products. You've seen this from companies like Coinbase, who are saying, look, we're just getting rid of middle managers. If all you do is manage people, then there is no longer a role for you in this business, that we're gonna have player coaches.

**Charlie Cowan:** So there's this hollowing out of the middle level. And then there is also this hollowing out of the bottom corners, where we are seeing that lower-level individual contributors that would be considered administrator, associate, people that move numbers from one document to another document are no longer going to be required. And so what we're seeing is there's this kind of reallocation of human capital at some of the bottom level of these organizations to becoming token capital that is being provided to people that are further up in the organization. Now, that doesn't have to be just the CXOs up at the top. I give you a little star here to talk about a high-performing individual contributor who's been given a sizeable token budget to complement what they are doing.

**Charlie Cowan:** So, from a CFO perspective, as you're starting to think about your P&L, and you're starting to think about your payroll, and how some of that allocation of capital may be reorganized. So let's get into our first of the four questions. So, how do I forecast? As a finance professional, you are not a deep sort of machine learning expert or IT guru. You have had no need to know what a token is up until now. But now is the time to get up to speed with it. So, firstly, what is a token? What I love about doing these webinars is it's a safe place to ask some what you might think is a silly question. So, a token really is around four or five characters, or words. It's a way of chunking up sentences and text into not a single letter, but a group of letters.

**Charlie Cowan:** And without going into the technology of how these large language models work, it's about them predicting what the next token is, and we do that by breaking these words down into these four or five. So just in this sentence, what I've done is tried to carve it out, pink is a token, black letters are a token, and you can see how this one sentence might be broken down into what is something like 10 different tokens. And when it gets to charging for these AI models, we're thinking about the number of tokens that are used. So you can pretty much then say, right, well, a page of A4 that has a certain number of words on it would be a certain amount of tokens that get consumed. Okay, so what does a token cost? Well, it depends on what plan you are on.

**Charlie Cowan:** Now, if you're on Anthropic, you might have a team plan or an enterprise plan. If you're within OpenAI, it might be a business plan or an enterprise plan. And then the same thing if you're on Copilot or Gemini, they're going to have different products. But generally, there is a plan on the left, which is kind of a subscription, and then I'm going to show you a plan on the right, which is a pay-as-you-go plan. So if you're on Anthropic, you might be on an Anthropic team plan. That runs up to 150 seats, so that might be because your company has only got up to 150 people, or it might be that you're running a pilot in one team or one region, and so you're still on that team plan.

**Charlie Cowan:** Here, you pay a certain amount per seat, and that could be around the $30 a month level, if it's for a standard user, or it could be around the $130 a seat level if it's for a more advanced user. But it's that kind of level. You're paying an amount per seat, and depending on what price you're paying, you get a certain amount of usage. Now, on Anthropic, you get a 5-hour usage limit. This is the amount of tokens you can use in a 5-hour limit. And you also get a weekly limit, so those 5 hours roll up to a weekly limit. And as long as you stay within those limits, then you're paying your $30 or your $130 a month, and away you go. All good.

**Charlie Cowan:** If you hit those limits, as you occasionally might, then you can allocate a usage credit to individual users, or individual teams, or across your whole company, and then you go on to this pay-as-you-go beyond that. So, you may, as a CFO, you may be looking at this and say, right, well, we're on the team plan, everyone's paying a certain amount, but then you've also got requests for pay-as-you-go beyond that. I put in capitals here, this is an amazing deal. It is an amazing deal. And this is the same whether you're on OpenAI or whether you're on Anthropic, and indeed are on the other providers as well. They are subsidizing their subscriptions very, very significantly. OpenAI, to all accounts, even more than Anthropic. And so, this is the best deal that you can get.

**Charlie Cowan:** You can get huge volumes of usage out of these subscriptions, compared to what you would do if you were paying on the API. So what is that other side of the coin? So, on the enterprise plans, and again, varies by the provider, but you may be paying a dollar per seat, so on Anthropic, you'll pay a dollar per seat, which will be less than on the team plan, so it might be something in the order of $20 a month, depending on the conversations that you have with your Anthropic rep. But the important thing here is this comes with zero usage, so you're paying for the platform, not for the usage. And you then pay as you go for every token that is going through the platform.

**Charlie Cowan:** And so we are seeing that some clients that are moving from the left to the right are having a very quick wake-up call as they move out of the amazing deal to the pay-as-you-go. Because if you've got 150 seats on the amazing deal, and then you now move to 151 seats on the enterprise plan, you are going to be paying significantly more just for that one extra seat. Now, you get more because of the enterprise compliance and so on, but this is very much what we're seeing with companies going through this wake-up call. So what does a token cost? Okay, we're moving on to the enterprise plan, or we've got usage credits on the team plan. What do these tokens cost us? This is Anthropic's.

**Charlie Cowan:** I'm going to show you OpenAI in a minute, but on Anthropic, you've got this rate card, and let's just look at Opus 4.8 up at the top, just as one of these line items. Opus 4.8, at the time of recording, is sort of the frontier model. We'll talk about this in a minute. But Opus 4.8 is what a lot of your developers will be using, and they split it down in this way, so, input tokens, $5 per million tokens, that's what this MTOK means. So, $5 you're gonna pay to send input tokens to Anthropic. So when someone writes, hey, can you help me to analyze our budget for this year? That is being cut down into tokens. Whether that is the text that someone is writing, whether it is a spreadsheet that they're uploading, whether it is an image that they're giving, whether it's a PDF, these are these input tokens.

**Charlie Cowan:** Now, you can then start writing to the cache as well. Now, we're not going to get into too much detail here, but when you send these same things to Anthropic each time, it can write it to a cache, and this is going to be important in a minute. So it's $5 every time I send something. But, if I write to a cache, I'm gonna spend more. Maybe $6.25, or if I want it to last for 1 hour instead of 5 minutes, $10. So, why would I spend more on writing to the cache? Because when you read from that cache, it is going to be significantly less. So these right-hand two columns, if I read from the cache, it is going to be 50 cents per million tokens. But if I am just getting a normal output, $25 per million tokens. So look at that, 50 cents versus $25.

**Charlie Cowan:** Okay, I want to write to the cache. How do I write to the cache? Because this is significantly less. Now, there are things, we'll talk about this in other webinars with your sort of AI team, with your developers, about how what you send to the models can ensure that you hit the cache and you don't break the cache, and then this is going to significantly reduce the cost of what comes back from the model. So, what does a token cost? Well, it depends, and it depends on these things here. Now we need to drop down a few rows to the other pink row I've got, which is Sonnet 4.6. So, Sonnet 4.6 is Claude's other sort of daily workhorse model, and you can compare and contrast some of these things. So, I'm making this up when I say it's, like, 70% of the cost, my quick ready reckoning math.

**Charlie Cowan:** But yeah, $5, or maybe it's 60%. So $5 per million tokens for Opus. And it's $3 for Sonnet. On the output, $25 for the output tokens for Opus, and $15 for Sonnet. So you can see here, just by picking a different model, you've immediately just saved 40% of the cost. And then for certain use cases, you could go down to Haiku, which is significantly less than that. So not all models were created the same, and we'll come to this a little bit later. Just to give you some visibility on OpenAI, I won't go through too much, but GPT-5.5 is their frontier model, and so we've got $5 input, 50 cents for the cached input, and 30 for the cached output. So, 5, 50, and 30. Let's flip back. 5, 50, and 25. So, ChatGPT 5.5 is effectively the same, slightly more on the output tokens.

**Charlie Cowan:** And then they've got different models here, so their preceding model, ChatGPT 5.4, 250, 25, and 15. So you can see it's pretty consistent across the different providers. You're not going to immediately save huge amounts just by being on one model versus the others. There's give and take on each of them. So then you go, right, okay, Charlie, I've kind of worked out now how these tokens are charged. Like, how are we going to build out a budget? So let me just flip over my screen here, and I'm going to go over to my Claude. And what I've done, just to give you an example of this, is I've built an artifact in Claude that would help me, as a CFO, to do some rough forecasting and analytics. Now, your business, you're gonna know inside and out.

**Charlie Cowan:** You're gonna know how you're split up by regions. You're gonna know how you're split up by business units. You're gonna know the relative sort of use cases or adoption across your organization. So you're gonna build your own artifact, your own budgeting tool using Claude. But here's some ideas about how I have handled it. So, number one is, right, employees. So, human capital, well, how many humans have we got in the organization? Is it 1,000? Is it 10,000? Is it 30,000? Well, let's work out how many people we've got. Then across those people, you might say, well, what do we think is going to be our assumption of the usage? And I've split this out by people that are just using kind of low-level chat. Hey, can you help me summarize this email? Can you help me prepare some meeting notes?

**Charlie Cowan:** And so you might say, right, well, 50 or 60 or 70% of your people are going to be in that. So let's say 60% of our people are going to be low-level users. Then you might say, right, well, actually, we've got some teams, they're going to be a bit more connected, they're going to be using loads of connectors, they're going to be using sort of plugins, and they're going to be doing more high-intensity work. And so maybe there's 20% of those people. And then, you're going to have, like, we're all in. So these might be your developers, they're in Claude Code, maybe they're using Codex, if you're on the ChatGPT platform. Maybe you've got some people in your finance team that are going to be using large, complex spreadsheets and models, and so they're going to be driving things up. So maybe you've got 10% or so of those people.

**Charlie Cowan:** And so you can play around with these to sort of make sure that you get your 100%. Then, and this is going to be difficult, maybe you're going to have a chat with your AI team, with your IT team, are there going to be a number of agentic workflows? And you're going to have to have a chat with your CIO and work out, like, how would you come up with a number for this? Is it 5? Is it 500? Is it 5,000? You're going to have a thought about this, and come up with what are the AI workflows that are not linked directly to a human. And then maybe you've got some other parts of your business about how you're tracking either different scenarios. In here, I've just put some dormant or inactive use cases, users. And then in my artifact here, I just asked Claude to help me model it.

**Charlie Cowan:** I pointed it at the Anthropic pricing. And I said, based on this, can we come up with some assumptions of what this might be for these typical users? And here you can see, we've got, in this sort of expected, we've got a seat fee of $20 for each, and then we've done a bit of budgeting over what might each of those users expect to use across the month. So down here, monthly cost per user by scenario. Well, a chat-only user might only use, you know, $20 to $40 of tokens. A connected knowledge worker might expect to use 100 to 175 of tokens. Maybe a power user is 500 to 1.2K, something like that. And I think over the course of a few months, we'll talk about reporting in a little bit, you're going to get a better feel for this.

**Charlie Cowan:** Certainly what I'm seeing on X, and you see some of these reports of people who have blown through budgets, in some of these high-power users, you could expect a high-power user to be using the same amount of tokens as their salary. If that person is costing you 10K per month, you might expect that they're spending 10K per month on tokens, and this is where we see this reapportioning of human capital to token capital for some of these high-performing users.

**Charlie Cowan:** So, I'm not saying these are the right numbers, I'm not saying you should look at this and go, oh, Charlie said 175, but I'm just saying, in your Claude, start to build out a model that is consistent with your business, the way that you structure things, some of the numbers that you've got in mind, and then you can start to play around with this to come up with what you think your budget is going to be. Before we move off this budgeting, I want to introduce, if you've not heard it before, or remind you of a thing called Jevons Paradox. Now, what Jevons Paradox is, it says that as the cost of something decreases, the demand for that thing increases, and the demand increases higher than the cost reduced, so the actual spend goes up. So, a good example of this would be widescreen TVs.

**Charlie Cowan:** I remember when I was at school, and you would go to, I think the shop was called The Leading Edge, and you could go and see a big flat-screen TV for £25,000. And you're like, oh, wow, amazing. But bit by bit, technology's developed, the cost of those screens came down, and now, who knows, you could buy a 40 or 50-inch screen for just a couple of hundred pounds. And so, as that cost has come down, more people want to buy flat-screen TVs, and therefore the total spend on flat-screen TVs increases. Same thing for cars, same thing for computers, and it's going to be the same thing for AI. So even though you're going to see this continual decline in the cost of tokens, and it's gone down something in the order of 60 or 70% over the last year, the actual demand is going to go up.

**Charlie Cowan:** So just always bear this in mind, in terms of how you think about this. The total cost is going to go up, even as these things come down. So that was budgeting and forecasting. Now I'm going to touch a little bit on how do you account for this. Now, I am not a chartered accountant. I am not here to tell you how to account for this, but I am going to just explore a little bit about how you might think about this, because it's not as simple as saying, oh, we're going to spend $100,000 a month on tokens, and that is just one line item in our P&L. Because different tokens are going to be used for different things. And so what you might think about, and you know better than I do in terms of your accounting, but one would be, is this a cost of goods sold?

**Charlie Cowan:** So, are these tokens being used in the delivery of our service to our clients? Is it, we've got consultants and they are delivering reports to our clients, or that AI is generating those reports? Are we providing customer service using AI? And this is where these tokens are being assigned. So, this could be a line item that goes towards your cost of goods sold. Then, where a lot of this is going to end up is in your OPEX. So this is our internal human capital, using these tools to do their work. So this is going to be the bulk of where your token cost is going to go. And then, depending on what your business model is, or the business that you're in, you may be capitalizing some of these costs.

**Charlie Cowan:** So if you are building your product, and it is engineers building that product, then it may be, depending on where you are in the world, that you're able to capitalize some of that. So not all tokens are accounted for in the same way. So I'd start to think about these tokens. Well, which team used that token? Was it the engineering team? Was it the sales team? Was it the customer support team? Which workspace was this token consumed in? I'm going to talk about this in a minute, because it's going to help you when you come to your reporting. You can have multiple Anthropic workspaces, and therefore, you could account for them in a different way. Which API key was used to generate this token? Was it one of our human teams, or was it an agent that used this? And then, which model was used for this token?

**Charlie Cowan:** Again, this is going to come into our reporting. Was it a Haiku token? Was it a Sonnet token? Was it an Opus token? And this can all feed into our accounting. So now we've forecasted, we now know how we're going to assign these tokens across our P&L, and now we're starting to think, how do we limit this? How do we get this horse back in the stable somewhat, so that we've got a bit of control over it? Now, generally, this is what we see across our clients, regardless of industry, regardless of where an organization is, is that you've got this bell curve of usage across your humans that are using it. So on the left-hand side, you've got very low usage or adoption. I either don't use it at all, or I'm using it for very low-level tasks, you know, summarize my meeting notes.

**Charlie Cowan:** This specific phrase comes up from other clients, I'm not making this up. We've got people in our organization who can't spell AI. And so there are definitely people that sit in that. Then you've got the medium usage. So this is where people are starting to use it. I'm in Claude, maybe I'm in ChatGPT, I'm becoming more adventurous with Copilot or Gemini, I've connected it up to some of my systems, maybe I'm now using a bit of Claude for Excel, so now I'm using it as part of my daily work. When I come in in the morning, I open an email, and I open Claude, and I start to do work. Then you move up to the high usage. So now we've got people that would say, right, actually the first thing I open up in the morning is Claude, before my email, and I'm asking it to plan my week.

**Charlie Cowan:** I'm now using Cowork, maybe, to do some agentic work for me. And then up at the top end, you've got the extreme power users. And these might be your engineers, developers, that are just using a huge amount of tokens. Now, there's no right or wrong here. What I would say is that generally, you are trying to move people from the left to the right, and we do this through the change enablement, through the training, through building the right tools for people, giving them the right skills. You want to move people to the right, because over on the right is where high usage and high value comes from. But I would start to take a close look at the people that are in this power usage.

**Charlie Cowan:** I have this as a bell curve, not as a leaderboard, and there's been some really good discussion online about Meta, the owners of Facebook, who had a token leaderboard, and the rights and wrongs of that. Being at the top of this list is not necessarily a good thing. For me, this indicates potentially inefficient usage, and we might want to take a look at that. So generally, you want to move people to the right, and as a change enablement team, I'd be looking at the people at the extremities, the people that can't spell AI, who don't know how to use it yet, and the people, very extreme users, who may be inefficiently using tokens. Talking about that change enablement, we looked earlier at Opus and Sonnet and Haiku. Now, if you're not deep in AI, how should you know which model to use and when? Because they have very different costs.

**Charlie Cowan:** And I have to say, I struggle with this. I struggle to explain to someone in HR, to someone in finance, to someone in marketing, which model to use when. And I've got 3 analogies that sort of might help. So one might be a bottle of wine. Whether this is you buying a bottle of wine at a supermarket, or maybe it's you going to a restaurant, you go to the restaurant, you know, you can order the house wine. Look, it's me, it's my family, I just want a nice bottle of wine, and I'm not too worried about it. But then, you might be taking a client out, you might be celebrating a big anniversary, it might be to celebrate a wedding or a birth, or something like that.

**Charlie Cowan:** Well, now, actually, I do want to push the boat out, and now I'm going to get a better bottle of wine. So you, as a human, are making these decisions about when to upgrade, and so maybe this is something to think about, that Sonnet is your sort of standard model. And then you think about, well, actually, I'm going to use a better model, because I'm going to do something that's really complex, or I really want to assign more thinking to it. The other analogy that I would use might be about fuel. So here in the UK, you would get your standard unleaded, and then sometimes, depending on the car that you've got, or maybe you're going to go into a track day, not that I've done a track day, but I'm assuming this might be a scenario where you say, I'm going to use some super unleaded.

**Charlie Cowan:** So generally, I'm always unleaded, and then maybe sometimes I'm going to put some different fuel in to get some higher level of performance. And then the third analogy, which might be more sort of day-to-day, would be Uber. You could use UberX, or you could go to Lux, or whatever it might be. Based on what you're trying to do, you might try and upgrade. So just think of these sort of scenarios where your base level is not the top. It's not always Opus. The base level is we use Sonnet for our day-to-day, or maybe a lower model on other vendors. So, we're educating our team that not every request requires this frontier-level intelligence. It is not a flex to say, oh, I always use Opus, or I always use the frontier.

**Charlie Cowan:** The flex is I use the right model for the job, and then I upgrade when I've got something that needs it. So a couple of controls that you can put in place as a finance team, and then working with your IT team. And I'm going to use a finance-type sort of analogy here, and thinking about a trading desk at an investment bank, and thinking about how they might limit the risk of one of their traders taking the bank to the wall. So the number one is, well, we provide what they call fat finger protection. How do we prevent someone doing something extremely wrong and potentially destroying the bank? Well, what we could do on our AI tokens is to think about having some budgets. So, how do we provide some per-user or some per-team token budgets?

**Charlie Cowan:** So, within that, do what you want, but we've got these daily, monthly limits that are going to provide you with alerts, so there are no surprises at the end of a month or a quarter. So that's the rate limits, stopping people doing anything completely crazy. Then the second is, like, giving people visibility of what they're doing, and how what they're doing links back to the business impact to the company. So we've talked a lot so far just about tokens and spend, but what do we get for that? If a sales team is using tokens, then are they selling more? If engineers are using tokens to build, are we getting more product and high-value product in the hands of our customers more quickly? And so, I'm going to talk a little bit later about reporting and dashboards, it's not just about the spend.

**Charlie Cowan:** How do we show people what the profit side of that is, not just the loss side of that? And then the third is a bit of this governance. How do we make sure that people are not doing the wrong thing at the wrong time? In the investment bank example, we're talking about compliance and fraud. Fraud is a strong word here for what we're talking about, but we might think about making sure that people are using the right tool for the right job, that they're not using personal systems when they should be using corporate systems, and that they are using simple, well, cheaper models for simpler tasks. So that's the limiting. Hopefully that gives you a few ideas on how to limit the costs that are coming through. And then the final part of the equation is how do I report on this? Now, I showed you earlier this same chart.

**Charlie Cowan:** I said, you want to be thinking about which team used this token, which workspace used this token, which API key used this token, and then which model was used in this token. And this is important, because if you think about this up front, it allows you to report correctly. So within Anthropic, and it's the same thing on OpenAI or the other providers, you can set up different workspaces. And you might set up a workspace, or your IT team, your CIO and your AI team would set up these different workspaces. It could be by business unit, it could be by your legal entity, it could be by geography, it could be by kind of use case.

**Charlie Cowan:** You can see here, for example, that we've got different workspaces for our Claude Code usage, for the clients, for the agents that we build for our clients, for our own internal sort of Kowalah usage, and there's many more underneath this. This allows you to report differently on each of these. So have a think about how your business is structured, and it will help you with your reporting. Then how do we get some of this data that we can then start to report on? And there are kind of three tiers here, and your CIO, your IT team will be able to help you with this, but I think about it on three levels. So the first is nothing to do with Anthropic, but it is more about just usage and logging in. So, depending on how your teams are provisioned with access to tools.

**Charlie Cowan:** Depending on how you report, you might use something like Okta, which is sort of single sign-on. You can make sure that you're only provisioning licenses to real humans, and so that when someone leaves the company, then they lose their license automatically. You're not paying for ghost licenses for people who aren't in the company. And you can start to track how many times people are logging in. So when we talked about that bell curve, you can start to see who never goes to Claude, and maybe we do some training and change enablement for them. Then, Anthropic provides you with a couple of APIs that help your team to report on this. So, one of them is the Analytics API, and this provides you with usage and spend data based on individual workspaces.

**Charlie Cowan:** So this is where this workspace thing is important, because if you set them up correctly, then you can report by individual workspace. It's going to show you the spend via different models, so you can start to balance a little bit around the different models that are being used. And then also, depending on how you set up individual API keys, maybe you assign different API keys for different teams, you can start to provide some of this reporting back to individual teams. What this will not give you is user-by-user analytics, and for that, you apply for the Compliance API. Now, the Compliance API, you have to get permission, or a request, I should say, from Anthropic, because this gives you actual chat data. So up until now, anything that someone is talking about, hey Claude, can you help with X? Hey, Claude, can you help me with Y?

**Charlie Cowan:** There's no visibility of what those conversations are to the IT team. But with a Compliance API, you absolutely can. So depending on where you are in the world, this is like a sort of GDPR and privacy issue. You need to make sure that the people that have got access to the Compliance API are properly certified and have got the right security in place. But once that has been sorted out, then you can report at this user-by-user level on the tokens that they're using, on how they're using those tokens, the models that they're using, and this is going to give you some really rich data that you can use for your sort of financial reporting. If you've then got that data, you can then go through these sort of two steps.

**Charlie Cowan:** So the first, which is what we'd recommend, is just to start by doing a showback, which is, we've got this data, we can now start to build out our bell curve. We could start to build out analytics by individual user, and find out who is in individual slots, so that we can start providing some coaching and change enablement. We can find those people that aren't using it, and then we can also dig into those biggest spenders, and work out what's the ROI on that, and maybe there's a bit of coaching out there.

**Charlie Cowan:** Having done the showback, and then once you are comfortable in those numbers, and that they stand up to interrogation, then you might start looking at chargeback, and how do we cross-charge this, not just being an IT cost, but actually, we're going to start charging this back to sales, back to marketing, back to different business units or regions, because we've got confidence in these numbers. And so this brings me to my second sort of artifact that I built in Claude, and you might want to build something like this yourself. And if I go to my other tab here, which is Reporting and Tracking. Once again, with the caveat, do not pay any attention to the actual numbers, because this is just for assumption, just for demo purposes. But we've got Claude here to build a little artifact.

**Charlie Cowan:** For those of you that are not familiar with an artifact, it's like a visual tool in Claude, but what you can do with an artifact is you can make it a live artifact, and it can pull in information directly from your other systems. So you might have Snowflake as a data warehouse, you might have Salesforce, you might have your accounting platform, I don't know, NetSuite, and you're able to pull in live numbers from these tools. What this means is that once you've built out that reporting framework, you could say, right, I want to pull in the actual live spend information by team, by business unit, by region, and I also want to pull in the live return information, how much have we sold, how much have we saved, so that you can start to build out this live sort of ROI dashboard.

**Charlie Cowan:** And how you might choose to frame that is by region or team. You might choose to frame it by your accounting classification, is this COGS or OPEX or CAPEX? You might choose to frame it by the type of usage, but you can really start to build this out and have it update automatically. You can then turn that into what is this ROI, so that you've got some board-ready data points that you can refresh to your heart's content every 5 or 10 minutes. Okay, this is bringing us towards the end of this section. So we've looked at how do I budget, how do I account, how do I limit the usage, and then how do I report? And so what that should leave you with is, okay, well, what do I do now?

**Charlie Cowan:** Well, we think about it in these three ways of maturity that you might go through as a finance team. So the first being, well, let's just start tracking and reporting. This is where most clients start, especially when they are on a team plan. They haven't really hit the challenge of enterprise adoption and tracking yet. So, we get a bill, we pay for it, no one's really too concerned about that. Step number two is, right, we need to get a handle on this, so let's start building some of these showback reports, let's start building out some of these artifacts for us to budget and to track, and let's start splitting things out by team. And then the third level we're going to get to is, really, you're working in tandem with your IT team to actually train and coach and build a plan that can scale out into the future.

**Charlie Cowan:** So, I would think about where you are on that journey, and then pick one or two things that you can build with your IT team. I can see there's a little question here. Oh, well, that's very kind, Francis. One of the most dynamic and clear presentations that I've seen. I really appreciate that. It makes it worthwhile. Thank you very much. So, to wrap up this section, I thought about this, what do I say here? Let's go back to what Satya said right at the start. Human capital and token capital, the token capital, if you allocate it correctly, is hugely powerful to the success of your business moving forwards. If you look at this increasing token spend and think this is just a cost item, a cost center that I need to just limit, limit, limit, you're gonna miss out on the opportunity that this transition has. So, think about this.

**Charlie Cowan:** Think about that you're an investment bank, and you're going to apply this capital in the correct way, I think it'll frame your mind in a very different way. Well, we've had a lot of fun there. 48 minutes passed, and we've got 10 minutes or so to go. Talk about AI in the news. I've got one item on the agenda, and this is about last week's release and almost immediate un-release of the new Fable 5 model from Anthropic. So, just a little bit of a backstory here. Up until now, Anthropic has always had 3 models. Haiku, Sonnet, and Opus. And they basically go from a small model, Haiku, up to a larger model, Opus, with Opus being the frontier model. Over the past couple of months, Anthropic had been trailing that we have built, I say we, I'm not Anthropic, but speaking as Anthropic, we've built this great new model called Mythos.

**Charlie Cowan:** It is so powerful and so clever at uncovering cybersecurity issues that we cannot roll it out to the general public, because if bad actors get hold of Mythos, they're going to be able to uncover cybersecurity issues, and they could take down core systems that run our economy and civilization. And so Mythos had just been rolled out to a handful of cybersecurity companies themselves to be able to get a grip of it, and a number of the sort of leading technology companies. What happened on Friday was, maybe it was Friday or Thursday, was that Fable 5 was released. And what Fable 5 was, a distilled version of Mythos that had a number of protections built into it to allow people to have access to this model, but without the big security risks.

**Charlie Cowan:** So out that came, I think it was Wednesday, actually, that it had just come out after we did our Wednesday webinar last week. I was very excited. I was looking forward to the weekend. I said, right, the weekend, I'm going to spend time on Fable 5. But on Friday afternoon, the U.S. administration issued a directive saying that it was still so powerful, some people had jailbroken it, which means basically they were able to go around some of those protections. And the US administration said that Anthropic could only, well, they had to restrict access to Fable 5 to only US citizens, and that included Anthropic employees. So, if you were a non-US citizen at Anthropic, you couldn't use it either. And so what Anthropic said, well, the only way that we can submit to this order is to block it completely, because we can't give it just to US people, it just doesn't work.

**Charlie Cowan:** So, Fable was taken away from usage. Now, the conversation is still going on. We're sort of four days into this. We hear these stories of Anthropic representatives flying into Washington to try and convince the US government that the relevant protections are in place. So, why do I think this is interesting from a business perspective? I think two things. One is that these models are getting so powerful that they are far surpassing what a human can do. And whatever your definition of AGI is, sort of advanced general intelligence, these models are beyond that. They're beyond what Helen in HR or Frank in finance needs to do. The models are super powerful, and they're going to continue to get more and more powerful as we go.

**Charlie Cowan:** Second thing that I think is interesting is that the people that have access to these frontier models are going to have more power, more token capital than those that don't have access to these models. And, I think we're going to see more and more of this as we go through, that as a company, you want to have access to those frontier models, and you need to make sure that your company and your people are being given, whether that's Anthropic, whether it's OpenAI, whether it's leading models from Google or anyone else, it is not going to be competitive for you as an organization to say, we're not going to give our people access to these models. The moment Fable 5 comes back, or whatever the version of it is that we get, the companies that have are going to be at an advantage to those that have not. So, watch this space.

**Charlie Cowan:** We'll have to see what comes out by this time next week. So, I haven't seen any other questions. If there are any other questions, then feel free to ask. But I'm gonna bring us to a wrap-up now. Next week, we have got another one of our Spotlight webinars, which I'm really excited about. We have got one of our clients, Triple Point, who are a financial services provider here in the UK, a provider of investment opportunities and private capital, both to individual investors and to companies. And we're going to be talking about how we and they have leveraged AI, and specifically Claude and Anthropic, in replacing an HR SaaS tool to gather feedback from their employees. So, looking forward to that, we're going to be joined by Shelley Davison from the People team at Triple Point.

**Charlie Cowan:** So, same time next week, you can register on this QR code, or go to kowalah.com/resources, and you'll see the webinars in there. With that, thank you very much. I hope you feel empowered now to track your AI token budgets, and we will see you next week.

---

This transcript is published by Kowalah, a UK Anthropic Implementation Specialist.
Webinar page: https://kowalah.com/resources/wednesday-webinars/measure-ai-impact-token-costs-board-ready-roi
All transcripts: https://kowalah.com/resources/wednesday-webinars/transcripts