The Agentic Work Unit (AWU)

Live from Schenectady, New York, it's the Metrics Brothers, Growth and CAC.

And I'm Growth, better known as Ray Reich, founder and CEO of Benchmarkit.

And I'm CAC, better known as Dave Kellogg, EIR at Balderton Capital, independent consultant, and the author of Kellogg.

And together we are the Metrix Brothers.

And we go together like tokens and turnstiles.

How New York of you, Dave, but I'm not sure anyone still uses tokens for the New York subway.

They do not, Ray, although I grew up with these little copper tokens you would use, but they do not unless of course they're vibe coding on the subway.

Ah, sounds like the good old AI token that everyone is talking about.

Indeed. And look, it's kind of a $2 word because when we're talking about tokens in the context of AI, really the dumbed-down word is just word or word fragment, right? That's what a token is. It's basically a word or a piece of a word, or sometimes a piece of punctuation processed by a model. But that's all a token is.

It is text in and text out.

Exactly. And more tokens generally means more compute, More GPU time, more cost.

So tokens correlate loosely with compute consumption?

They do.

They're correlated to compute consumptions. I think in general that the total compute required by a model is kind of the model size times the context length times the output tokens if you wanna do math. But see, you're speaking about compute consumption. The actual unit there would be the FLOP, right? The floating point operation per second, 'cause large language models are effectively giant matrix multiplication engines. And that's what burns all those FLOPs. FLOPS are technically the consumption metric. Gigawatts, as we talked about in another episode, is effectively the measure of the power to provide those FLOPS.

Well, one thing I know is tokens are not business metrics or even financial metrics.

Yeah, they're just plain not. A lot of people seem to want to hedge that and say they're kind of business metrics or not really business metrics. Like, they're just flat not business metrics. They are There are text processing metrics, you know, that's probably the clearest definition of what they are.

Kind of the units of language handled by the system.

Yes, exactly. Look, nobody walks into a board meeting and goes, great quarter, we processed 14 billion word fragments.

No.

Well, one guy might, and that's Jensen Huang, because, you know, he sells GPUs that process these tokens. And he even says that every $500,000 engineer should burn $250,000 worth of tokens per year.

He does say that. And does that mean that tokens are effectively a pricing metric?

Ah, that's a slippery slope. I would say kind of. You know, we talked about they are a text processing unit, a measure, and they correlate reasonably well with compute consumption. And that's what makes them convenient for pricing. So you see Claude and OpenAI are pricing per token right now, but That creates some of the issues in and of itself.

And here today, we're going to talk about another company's effort to find kind of an AI pricing metric in my mind, which is Salesforce's attempt to create a new higher-level metric, kind of a level up from the token called the Agentic Work Unit or the difficult-to-pronounce acronym AWU.

Well, you know, Once again, I need to get a better hobby, but I can't wait to dive into this particular metric. But before we do, let's hear a word from our sponsor. The Metrics Brothers is proudly sponsored by Benchmarkit, the B2B AI and software industry's largest, broadest, and most contextualized benchmarking index. Dave Kellogg and Ray Groth Wright have put the Metrics Brothers podcast on hiatus for the remainder of 2026. All episodes we have ever recorded are still available on your favorite podcasting app. You can reach Dave @kellogg and Ray on LinkedIn @rayreich. Thank you so much for listening to the Metrics Brothers podcast.

Okay, so we have established kind of a history of kind of how tokens have led to the need or desire to have higher level I want to call them pricing metrics. And Salesforce in February 2026, as part of its broader Agentic Enterprise push, announced the AWU. They defined the AWU— God help me in saying this thing over and over— as one discrete task accomplished by an AI agent. Superficially, that sounds like a rigorous definition until you say, hey, wait a minute, what's a discrete task? And here you can be all over the map. Some of the examples might be routing a lead or updating a CRM record or invoking a workflow or calling an API or orchestrating multi-step actions across systems. Look, conceptually, this metric is clear. They're trying to kind of get a better measure for kind of work performed by AI software that's better than kind of words or tokens processed.

Well, I must admit, Dave, when I first saw And this was February of '26 when they announced this term, the AWU. And my initial reaction was, well, that's kind of cool. They're pricing upon a unit of work. As I started to think about it more deeply, I'm a little mixed on the value of the metric, and that'll come out during today's session. But let me back it up first. So Salesforce has always sold business applications, not foundation models or compute resources, right? Because tokens make sense when you're selling compute resources utilized like large language model companies do. But the buyers, especially enterprise application buyers, the Salesforce buyers, I don't think they are thinking about the number of words processed or a unit measure for language, right? So I think they had to come out something a little more commercially closer to business activity or even outcomes rather than a compute resource consumption unit. Like tokens.

Yeah, you could argue that tokens are basically for consumption-based pricing. And if you want to move to outcome-based pricing, you need to find outcomes. And while outcomes work really well for certain examples, like customer cases resolved, like certain apps just naturally lend themselves to outcome-based pricing, right? I know what a case resolved is. I know what I'm willing to pay for it because I know how much it costs me to process it with a human instead. Right? Unfortunately for software vendors, that one example works really well. Maybe invoices processed works really well, but there's a whole lot of things that don't work so well and that don't so cleanly lead to an outcome, which creates this desire to create some sort of intermediate outcome metric. So I think, look, I think the goal for this is to price on it. Right now, they kind of report on it, at least in a macro sense. Here's how many AWUs all users used last quarter. Right? And I feel like they remind me of a cell phone company trying to get people used to gigabytes, right? Because for years, when you had a cell phone, you thought about it in terms of minutes. But now we have data capabilities, and they need to measure consumption of gigabytes, right? But I don't know what a gigabyte means from Adam. Like, how much is a movie? How much is a phone call? How much is an image, right? You just didn't know. So they started talking about them. It was kind of a soft rollout. To try and get people to understand, really develop some intuitive sense for what this thing is. So, even if I can't precisely define an AWU, I might be able to know I used a million of them last quarter, and I used 900,000 the quarter before that. So, it starts to get me comfortable, have a pricing discussion based on that.

Yeah. You know, this reminds me of when I used to sell EDI, Electronic Data Interchange products. We would charge on the kilocharacter basis, 1,000 characters. And CFOs are like, I don't want to count all these kilocharacters. I'm sending a purchase order or I'm receiving an invoice. So I think the same issue is here for AI software vendors, including Salesforce. You know, their buyers, their CFOs, you know, they don't care about a token or the number of words that I'm using. They want to know a little bit about what's the business value that I'm receiving. Things like, I got that case resolved, et cetera. And I think Salesforce is a— they've personified the challenge of how to price AI solutions because they've done per-conversation pricing with Agentforce, you know, $2 per customer service interaction. Then they started with Flex Credits, which was really kind of a consumption-based credit tied to AI usage. They did some action-oriented pricing, charging more directly for tasks completed. And now they're bundling AI usage inside a subscription agreement. So I mentioned all that, Dave, because I think this AI business application business is still trying to find that right output or outcome metric to price on. And I don't think we'll— I don't think we're here yet. And the AWU definitely is not a perfect pricing mechanism. I know it's not pricing. But I think there's maybe this vision that it could be.

Yeah, I think it's clearly the vision. They'd like it to be pricing. And I disagree. I think it might be a reasonable metric. There's never going to be a great metric for this because you're trying to capture a generic task or generic work unit. So the question isn't going to be, is it great? It's going to be, does it kind of suck less than all the others? And can you get, at least for Salesforce's case, can they get their customer base having some intuitive sense of how much work is done in an AWU? that they're willing to buy credit packs worth 50,000 of them or a million of them? And that's the question. By the way, you heard it here first. Ray, what a groundbreaker. He was selling software in the days before time began by the killer character. Now, your average token is 4 characters, Ray.

Right.

So you were selling software by the token, in the 250-token pack. You were selling that back in the days before time. Impressive.

Bundled killer characters. And I will tell you the reason I had to highlight that at the beginning, people loved it. Oh, I can easily measure. I know how many characters I'm sending, but CFOs hated it because when those killer characters went up unexpectedly, it killed budgets and it was not received well. Tokens are gonna have the same challenges.

Yeah, this is always the problem with either consumption-based pricing or outcome-based pricing. To the extent it's not predictable, customers won't like it. So superficially, it's very easy to say, yeah, let's just pay for outcomes. And then you go, well, what's an outcome, right? And you find all these kind of intermediate or low-level outcomes that aren't really outcomes, and I don't know how to value them. And then the wheels kind of fall off. So, and somewhere in the answer, you will find the middle ground. Like, the industry is pretty good at settling for something. I think Satya Nadella said it pretty well the other day. Now, he's primarily a per-seat business. I mean, irony alert, like all the major AI companies sell by the seat. Microsoft certainly sells by the seat. And Satya is basically saying he sees seats as just coming with a certain token entitlement or a certain AWU, if you will, the MWU worth of units that come with it. So we'll see how this all breaks, but I very much do believe this is an attempt at trying to find a good pricing model. And I'm a little bit more optimistic about it than you are. Salesforce has a lot of power in the market. And this is now their third or fourth attempt. To your point, you think they might get it right eventually.

So Dave, you know how I like to do history, and I know sometimes it may age me, but Matt Calkins from Appian—he had a great analogy. Do you mind if I share that one? Please, Ray.

So is this pre-Cold War? This is going to be pre-Cold War, isn't it?

It's the great Soviet factory analogy because. You know, the communist government was trying to measure the performance and productivity of these new government-controlled industries like factories. So, they said, okay, every factory is a little bit different, but they're producing these manufactured products that weigh something. So, they said, we're going to have a common measurement of weight. And that's how we're going to measure whether a factory is performing. But then, we had an example of lighting factories like, well, I'm going to be optimized for the weight of the lighting that I manufacture. So they started building, Dave, these huge, giant chandeliers that weighed so much when they actually were trying to put on ceilings, they actually fell down and it couldn't be used. So this is an example where a measurement becomes very impractical and useless in application.

Yeah. And look, I don't think anyone's proposing AWUs to be a management metric, right? I think it's being proposed to be a pricing metric. And look, customers do contort themselves to minimize their pricing. So, these things do matter. They're going to try and find ways to use fewer AWUs once they know what an AWU is and if Salesforce prices by it. So, I think that happens. And I'm going to go even before your Soviet example. Never forget that Charles Dickens was paid by the word. So, he wrote, I think this might be an urban myth, but he was rumored to have been paid by the word and therefore he wrote fairly verbose texts.

Yeah.

Right? You kind of get what you pay for and conversely.

You know, I still, I remember one more thing and I'm going to get out of my history kind of lesson here. But I remember when I was first selling packaged enterprise application software, we were charged based upon how many CPUs was on a box. You know, you had the Sun Micro, I think it was the E10000. It was an 8-way box. We charged more When our software was deployed on a larger multiprocessor box, we did for a one-way or two-way. And that worked really well until we saw things like virtualization in the cloud come into play.

Yeah. And a lot of people, I mean, look, on that specific example, the first move of the software vendors when virtualization came along was that each virtual machine needed its own license, right? That, which is a pretty aggressive posture for how to price on a virtual machine basis. The other thing I'd say, Ray, this is like Kellogg's first rule of pricing handed down to us by the ancients is always hook your pricing to things that go up. And so, the first pricing models I remember were by the MIPS, the million instructions per second, and the price of the software is a function of how big, how many MIPS the processor had you were running it on. Now, that was Moore's Law made that an absolutely fantastic thing. to base the pricing on, right? Because the machines got more and more powerful every year. But in this case, it is actually more outcome-oriented. I would say that, like, it's not inherently going up. That may be one of the flaws of the AWUs. Like, it's not obvious to me that AWUs go up every year, just like MIPS do. Or in your case, last year's 4-way box was an 8-way box, which turned into a 16-way box, right? So, go from 4 virtual machines to 8 to 16 if you're selling VMware. So I don't know, Ray. I think that could actually be a flaw in the AWU that it seems to be breaking Kellogg's first rule of pricing, which is always hook the things that go up. And you could argue people do more work every year, but I don't know. I don't know.

Yeah. Well, I think AWU, God, you can say that easily, is directionally right, but I still think of it as an activity, not a true outcome like increased revenue or a closed case. So, I think enterprise buyers are going to see a metric for pricing be tied more to the economic value. So, maybe it's activity to case close to ultimately reduce expenses. So, I think it's a good entry level, but I think it needs to be linked to economic value over time.

Yeah, that's— I don't know. I'm less optimistic about that than you are because Customers say they want that. And then you say, cut me in for 10% of the profits. And they say, absolutely no, right? You're just a software vendor. I'll buy my own computer, right? So, it's gonna go back and forth. It's a tug of war, right? This is just welcome to capitalism. And while customers will say, I want to see it hooked to outcomes, do the vendors want to take that risk for question one? And sometimes they will. Like I said, AI SDRs, AI customer service agents, that's pretty easy. But especially in infrastructure software and a lot of the more just where the work it's doing is indirectly, is inherently indirect from a business outcome, I think you're going to need to find some intermediate metric that's not going to be a business outcome. And I think you'll be reluctant because you don't know how much inference it takes to generate that outcome. You might be kind of opening Pandora's box to say, hey, I'll charge you this many dollars for an outcome without having a real good idea of how many tokens or how much processing it's going to take to deliver it, how much cost in effect.

Dave, one of the things I know you started talking about this maybe a year and a half ago, and we talked a little bit about it prepping for today's episode. And a lot of AI software is looking to attack that target addressable market of labor, and it starts looking more like digital labor. How does that factor into this discussion?

So, Look, I think overall that the whole outcome-based pricing movement is trying to increase the TAM for software basically beyond its current TAM, which largely helps work get done, to the future TAM of replacement for labor. So, software vendors and venture capitalists love this because it's like a 10x to 100x TAM increase. There's just much more money up for grabs. If you can start replacing people. And this is part of that. Like I said, I mean, like that ultimately, you know, one company I work with wants to handle your renewals and they'll just take a cut. They'll say, hey, you've got a long tail of renewals, you're not doing them, give us a cut. And that could work, right? Like the value proposition is what drives the pricing. If I say, hey, you've got a long tail of little customers and you don't really do much with them and there's a pretty high churn rate in there, Give it to me, my AI, and maybe a small call center complementing it. And I'll go get this many more renewals than you're getting and give me a third of the win. And like, you're ahead by two-thirds, I'm ahead a third, you know, and there'll be a tug of war of like, we created 100 units of value, which we can prove pretty easily. How do we split that? Well, that'll be the competitive market. There might be 3 guys like me trying to bid on that. So, some stuff just lends itself better than others to replacing digital labor. Or proving economic value. Because in that case, you're not replacing digital labor. There was nobody working on them. But what you're saying, you've got this kind of fallow field of renewals that nobody's working on because it was never worth it with people, but now it's worth it. So this is all part of that trend. And to me, it's back to what I just call an intermediate outcome, Ray. I talked about this at your show, SaaS Ventures Palooza. Like if you're a dating service, what's the outcome? And arguably it's marriage, right? So, am I going to price by the date? Am I going to price by the second date? Am I going to price by the relationship? Am I going to price by the engagement? Right? I know the conversion funnel. It's not hard to build a mathematical model that prices equivalent value across all those metrics. But the issue is which one will the market want me to price by? And I want to call the other things intermediate outcomes. And there are cases, look, dating services don't price by the marriage, right? They price by the month or maybe by the date, right?

Yeah.

So, things don't always end up directly at their economic value, like, because I think there's many reasons why, but this is one example.

Yeah. Well, one of the things I was thinking about was this AWU, that's a vendor measurement. Boy, it would be amazing if we could actually convert that AWU to the customer's measurement, whether that's they want to do resolved case or they want to do accounts qualified. I'm just making those things up. But I think it's too difficult to have a pricing unit that is custom to each particular customer. But boy, wouldn't that be a great ultimate result?

Yeah, I don't like that. I think there's a little pie in the sky for me, Ray, which is like you're saying, wouldn't it be great if something that's inherently impossible was possible? So I'm like, okay, yeah, I guess so. But, but to me, it just feels inherently impossible. possible for the reason you state, that certain applications you can do it. But if you're trying to say, wouldn't it be great if we had a generic unit of work across all tasks and all companies? And like, however we counted it, it one really equaled one, you know, like recruiting a candidate was worth 57 of them, and closing a deal was worth 287 of them. And that was actually a really good proxy. I just think it's not going to happen. And it's a cool idea, but—

That's That's one of the reasons I like vertical AI. And even though customer support's not a vertical industry, it's kind of a vertical application because a ticket is a ticket, a case is a case. So I love pricing models based upon something that's really easily understood by an entire category, like contracts or claims processed in insurance, or tickets closed in customer support. Ultimately, I think those are going to be much easier to price.

Yeah.

on that outcome that's much closer to the actual economic value to the customer.

Yeah, vertical apps lend themselves well to this. I can't remember the name of the company exactly. I think it's EvenUp, who focuses on percent of the insurance total value claim you get, right? So if the guy's insured for a million bucks and you get $800,000, that's 80%. Can't remember the name of the metric. But yes, when you get vertical, when you get in deep, you can find outcome-based orientation, outcome-based metrics. And by the way, it kind of eliminates the need for these intermediate metrics. Like, you know, if I have that to base my price on, I don't need AWUs. I can just say, hey, you're on average getting 50% of the insured value. With my software, you can get 80%. So give me a cut of that increase. You don't need some abstract unit of work.

Yeah. Kind of just linking back to this AWU conversation in Salesforce. I actually think Benioff, because this came out, I think, during his February earnings call, you know, they had closed 29,000 agent-forced deals in '15. He talked about going from $1 billion of agent force revenue growing towards $2 billion. And he talked about service agent growth and employee agent growth. And I think the AWU, the fact that he could say, hey, we've had 2.4 billion agentic workforce units to date, and it was a really big amount in Q4. I think he was looking for something also to communicate to both the Wall Street analysts into its enterprise corporate buyers that was more aligned to how they think than these tokens. Ultimately, I think it's what he was doing.

Yeah. So it's a generic work unit, not workforce unit. You somehow stuck people in there, right? But yeah, he did talk about it on the earnings call. I view it almost as an AI adoption metric, right? He's trying to say, hey, look how many— we have this new way to measure work done by AI. And we've, you know, to date, we've generated 2.4 billion of them. I'm just gonna call them schmumbles because it's easier to pronounce. 771 million schmumbles in Q4 alone. You know, our service agents generated 106% quarter-on-quarter growth in schmumbles, right? 62% in employee agents. So it's giving him, in my mind, a way to talk about AI adoption. And maybe I'm overreading into it as a pricing metric. Maybe it's just that simple. It's part of their AI adoption story, which, by the way, they, You know, they're a traditional SaaS company. They need to do a lot to try and prove people are using their new AI offerings. So maybe we're reading more into this than is there, but deep in my gut, I think this is a pricing— this wants to be a pricing metric.

Yeah. I'll just go here one more time, kind of as I wrap up my perspective. And that is during that same earnings call, Dave, I believe that they reported Internal use of Agentforce, that 96% of their support cases were handled by agents, and it saved them 50,000 hours of human resource time. So, I still think they knew that being able to put something that's really tangible, saving 50,000 hours of time at X dollars per hour, it's something that enterprise buyers could really appreciate. So, I still think we need to get closer to that ultimate outcome. Where and when possible. Yeah.

And yeah, that's the important part, the where and when possible, because it's just not always possible. It depends where you are on the stack. So, but I agree. I agree mostly. You know, it's a competitive market out there. People say they want like value-based pricing and shared risk pricing is something we've heard about for 20 years. The world didn't go to it overnight. Some people do it. It works in some models. But there's a lot of kind of good ideas out there in pricing that don't get taken literally. And I think outcome-based pricing right now is a little bit overstated. As mentioned, there's a lot of areas where you want to price outcomes, in which case you price outcomes. You don't need AWUs. The more I think about it, Ray, I might think— when I first saw these things, I was like, oh, this is a pricing metric coming. Maybe it's not. Maybe it's just an adoption metric, and they're just trying to show that, hey, Look at us, we're doing real AI work. Could be that simple.

Well, I look forward to following this because we're so early. I want to see how this AWU works out over time.

Needs a new name, we know that. If nothing else.

Some of the metrics that you and I have always talked about on the Metrics Brothers, things like NRR, they didn't come out in the first 2 to 3 years of the SaaS industry. I think NRR came out like almost 8 to 10 years afterwards. So, It's going to be really interesting to see the evolution of AI-specific metrics.

Yeah, that's true. Just to be clear, AWUs can't be one of them. This is a vendor-specific metric, right? There's no way in the world this will— my prediction, this will never become an industry-standard metric. It may become a Salesforce standard metric. But this is not like NRR, right? Because when you equate it to NRR, I'm thinking industry metric. And this is not that, nor is it contemplated to be that. Do these things evolve? Yes. The early metric in, you know, as you both, as we both know, churn was the metric that came before NRR. People first looked at churn and back before we had NRR, we'd talk about negative churn, right? It was kind of awkward. It was like, hey, we had 100, you know, we had 10% negative churn on the quarter because we were so wired to churn as the reference point, right? Just like when a startup went cash flow positive, we used to say it had negative cash burn. Right? Because you build your, you title the row on the spreadsheet cash burn, and then you don't put a negative sign in it. And then when you're cash flow positive, you have to say we had negative cash burn. That was kind of how people looked at churn in the early days before NRR came along.

You know, I'm going to wrap up with my final thing. And all I can say is, ah, you've got it right. AWU. I know it's a pretty bad joke, but I was trying to find something in there, Dave.

It was such a bad joke. I didn't even realize it was a joke. Okay, folks, it needs a new name. You've heard about it here first, and, uh, thank you for joining us, and we'll see you next time.

Thanks everyone. Thanks, Dave.

See ya.

The Metrics Brothers: AI, SaaS, and More is a production of The Metrics Brothers Growth and CAC and a member of the Benchmarkit Podcast Network. By accessing this podcast, you acknowledge that The Metrics Brothers make no warranty, guarantee, or representation as to the accuracy or sufficiency of the information provided or the humor content of the jokes told. Ray? Information, opinions, and recommendations presented are, according to our spouses, probably wrong and provided for information purposes only. This podcast should not be considered professional or, for that matter, unprofessional advice. We disclaim any and all liability for any direct, indirect, indirect, misdirect, incidental, special, ordinary, consequential, inconsequential, or other damages arising out of any use of or God help you reliance upon the information presented here. Ray Grothreich is based in New York City and available on X at Ray Reich. Dave Kellogg is based in Silicon Valley and available on X at Kelblog. Schedectity, which is French for unspellable, is not our actual production location. Thanks for listening.