Kyle Poyar: The State of AI Pricing in 2025

But for AI products, if you undercharge, there are very real costs associated with that. You could have negative margins, highly unprofitable customers. Some of the biggest players out there are scrambling on pricing and having to make pricing changes at a much faster cadence than we've ever been used to. We also used to watch what the biggest players were doing, assume that they had done their homework and tried to follow along. Even the biggest players haven't necessarily figured it out themselves. Buying models are evolving from this idea of owning to hiring AI products that's allowing us to tap into, for instance, hiring budgets or kind of people budgets as opposed to just traditional technology budgets. We just can't afford to offer unlimited products where we're offering one price to every customer, given that variety of usage patterns and the costs associated with it.

As a founder, I think pricing is one of the biggest mistakes I've ever made across multiple companies. As an operator, I have lived and operated in companies that have never been serious about pricing. They have never sought to make it a first-class citizen in their products. They have never sought to make it a first-class discipline that they need to be operationally excellent at. Even the top business schools in the world are not teaching pricing. Like, you go get your MBA at Stanford and you don't learn pricing. And it's amazing because it's one of the most crucial growth levers and this essential ingredient to capitalism. And all of that's been happening for a long time, but it's also sort of happening at the backdrop of like a lot of change right now. Many of you are here because you're the founders of AI businesses. And so when I think about what's changing in the world as it relates to pricing, I kind of look at it through two different lenses. The first lens is we're operating in a world where There's more and more software and AI than ever before. What does that mean? It means your customers have a lot of choice. A few years ago, you do a demo, you buy a handful of seats, deal's done. Today, your customers expect that they can buy your product however they want to buy it, that they can be billed for your product how they want to be billed, and that they can use portions of your product, not your whole product. And so this is forcing on you as builders a reality. And the reality is you've got to be incredibly flexible with your pricing and packaging. It's just table stakes. It's the new expectation. And then obviously, the second thing is the kind of the word that we're all in the room for, and that's AI. And AI is creating an enormous opportunity for us as builders, but it also introduces basically a manifest manifold change in how we price. You know, you've got so much volatility and unpredictability around inference costs. What does this mean? What does this mean for the buyer's experience when they're having to consider a purchase that may have zero predictability. How do you handle that as a vendor? How do you handle that as a buyer? And so these are some of the big changes in the world that we're all having to figure out as we think about pricing and packaging and monetization. I ran pricing for a software company as it grew from $10 to $40 million in ARR. And it was an amazing journey, and it taught me just how sincerely deficit we were as a company as it related to using pricing to drive growth. We left tens of millions of dollars on the table. That's my personal story. Kicking us off tonight, we have Kyle Poirier talking about the state of AI pricing. Kyle's calling in from Boston. I'm sure he needs no introduction. I'm sure many of you read his work, but if not, it is essential reading for anybody building companies. Kyle, thank you so much for being with us, and I'll kick it over to you.

Fynn Glover asked me to just give an update on kind of the lay of the land with AI modernization and You know, it's something I've been, you know, working on a lot personally, something I've been studying, researching, surveying companies who have a lot of findings. But let's just set the stage. I've been personally working in pricing and monetization for more than 15 years. And, you know, when I started up until honestly a couple of years ago, SaaS pricing used to be kind of boring, especially at the startup stage. So folks would undercharge. Figure out who gets the most value for their product, build products for those companies, uh, raise prices, raise and repeat. That model worked pretty well, uh, even, even though, you know, there were folks out there wanting more art and science to pricing, uh, that, that model was pretty tried and true. And that model I think was fine if your margins are 80% plus. That's not so fine with AI products. For AI products, if you undercharge, there are very real costs associated with that. You could have negative margins, highly unprofitable customers. I've actually seen in one company their top 5% of users accounting for about 75% of usage costs, but they had a flat fee pricing model. And so those 5% of users were 5% of revenue, um, and they were all highly unprofitable, right? And so that's a very, very big challenge with AI models. I also think that, uh, many products are taking off really quickly. Uh, we're seeing companies hit revenue milestones much faster than before. And so you can't necessarily afford to get pricing wrong and kind of figure it out later if a product can really be that successful that quickly. And I think another thing I would just add is many folks were betting on LLMs getting significantly cheaper. There have been talk about, you know, LLMs are getting 10x cheaper every year. And I think, you know, there is a little bit of truth in that. Uh, but at the same time, the best LLMs, the best models are still fairly expensive, right? People want access to the best technology. No one wants yesterday's newspaper. These LLMs are, are certainly, yeah, not getting 10x cheaper. And I think the, the other complexity is that token volume is just exploding as folks have longer context windows, more reasoning. More agentic capabilities. And so if anything, maybe a year or two ago, we're hoping that we could punt on monetization till later. I think the reality is folks are saying, hey, actually, this is something we need to figure out now. And I'm even seeing some of the biggest players out there are scrambling on pricing and having to make pricing changes at a much faster cadence than we've ever been used to. So Salesforce, for example, only 7 or 8 months ago that they introduced conversation pricing for their AI product, Agentforce. Well, then back in May, they added a new pricing option, a flexible credit-based model, reflecting the, you know, significant kind of differences in how folks wanted to adopt their product. OpenAI, meanwhile, is, you know, historically had a seat-based model for ChatGPT, and that seat model has extended to a lot of their products. So their enterprise capabilities, they moved from seats to a pooled credit model in June. And then I think the pricing change that Fynn Glover was alluding to in his intro that many of us have paid attention to is Cursr shift from positioning as unlimited to a more of a credit-based pricing model in June 2025. They faced real pushback from users, real significant challenges around that rollout. And so I think that we also used to watch what the biggest players were doing, assume that they had done their homework and tried to follow along. Um, but I think the challenge right now is that even the biggest players haven't necessarily figured, figured it out themselves. And so, you know, I wanted to share a few observations from kind of watching, uh, this play out so far, but also pattern matching it with some of my other experiences, um, in monetization. And so one thing that I, that I think about is that, uh, there is a, an evolving shift in monetization models, right? In software, we used to buy software on-prem, and then we paid for software on a subscription basis. Usage models started rising in popularity. Now we have outcome-based models, credit-based models, you know, a host of different pricing varieties. I think there's a lot of readiness to say, hey, these old pricing models are dead. We're all focused on the new models. But I think the reality is that it actually, it's, it's more complicated and more nuanced than that. And I think about it as kind of equivalent to more mature industries, like, you know, buying a car. So when you're thinking about a car or getting access to transportation, you could buy a car, right? That's essentially that on-prem license, license model, license maintenance model, maybe. You could lease a car, so more of a subscription model with a long-term contract. You could rent a car, more of a usage-based model. Uh, you could call an Uber, more of an on-demand usage model, or you can even hire a chauffeur, right? So you could go from owning and paying for access to something to paying for the work delivered by a combination of the product and service. And with each step, as, as more of these buying options were available, more people could access transportation, right? The upfront costs of getting access came down., but you didn't necessarily pay any less. You paid less upfront and more over time for the most successful products. And overall, the market got a lot bigger. And so I think, you know, we're seeing similarly in AI products, buying models are evolving from this idea of owning to hiring AI products. That's allowing us to tap into, for instance, hiring budgets or kind of people budgets as opposed to just traditional technology budgets. But buyers actually still want a good amount of flexibility. They might want to start in one model, move to another model for more predictability or cost control. Don't think of it as sort of one model only. Think of it as a variety of options and understanding what's best for your business. And, you know, one thing that I will share that I think is, you know, pretty interesting part of this phenomenon, and this is one maybe slide that I'll share here. As we've kind of introduced different models over time, right? Moving from on-prem to subscription models to usage models to charging for work, for charging based on success, you're actually taking on a lot more risk for the customer and you're really only charging when the customer sees value. But that means that you're able to capture a much larger share of the economic value you create. And so for on-prem models, in my experience, businesses were only able to capture about 5% the economic value, mostly because customers had to pay for all of that potential value upfront. They were on the hook for realizing it themselves, and there was a ton of uncertainty, right? With subscription models, we captured more of that economic value, but we captured it over time. We had to invest in things like customer success. We had to earn the renewal and the expansion from the customer. With usage-based models, because people could use the product on demand, they could essentially vote with their usage, whether they wanted to keep using the product. It was much more flexible, but it also meant spend would increase pretty significantly the more you used a product. I think folks look at an AWS or a Snowflake and you say, hey, it's really affordable to get started, but it actually gets quite expensive at scale. But when AI products are doing the work for you, or if you're only paying on an outcome-based model, these are products that are really delivering a combination of the technology and the service itself. Itself, and that allows these products to take on even more of the economic value. I'm also seeing the rise, and, and this is a little bit quieter in the space, but seeing the rise of companies that are actually say, hey, we're going to buy a traditional services firm, replace a lot of that services work with AI, charge for a combined service, and essentially we'll sort of take a lot more of that economic value because we're not just an AI vendor, we're a fully managed service vendor. And so how you charge is actually very integral to how much you can charge. And as I think about some of these different kind of pricing models and evolutions over time, I've bucketed it as 5 main categories of pricing that are most common for AI. So one is flat fee subscriptions. So that's when you charge maybe based on features or company size or just have one fixed price for everyone. There's seat-based pricing. That's the pricing we've known and loved from, from software. There are hybrid models that are usually some sort of combination of seek and flat fees plus usage. There are usage-based models and then there are outcome-based models. So kind of 5 core models out there. Uh, and when I've surveyed companies about what models they've adopted, um, I've actually asked what model was your primary model 12 months ago? What's your primary model right now? And what do you expect your primary model to be 3 years from now? Well, the data has been pretty interesting on that. And so that's— this is another slide that I'll share if you guys can see it. Back a year ago, pricing was pretty dominated by flat fees and seats. So about 50% of companies— these are companies that have, have AI offerings as part of their suite— about 50% had either flat fee or seat-based models, but 20% had usage-based models. 3% had outcome-based models, and 27% had hybrid pricing models. When we fast forward to today, and then looking ahead 3 years, the real rise is in hybrid models. These are so much more popular than they were before. Um, and so the hybrid models, the most common are kind of combinations of platform fees plus credits, essentially to be able to charge for usage on top. And I think part of that is driven by the value of AI products can't be tied to seats anymore, right? Products deliver more value as maybe fewer people need to log in and have access to it. So we need to find more creative ways of anchoring pricing to the value those products provide. And we just can't afford to offer unlimited products where we're offering one price to every customer, given that variety of usage patterns and the costs associated with it. So seeing big decreases in flat fees and seat-based pricing and big increases in hybrid pricing generally in the form of credit models. You know, looking ahead, the holy grail for folks is thinking about outcome-based models. And I asked folks what they expect their pricing model to be in 3 years from now. You can see almost no one predicted flat fees and, and seat-based model. Uh, but there's a lot of interest in exploring outcome-based models or hybrid models., some of which might have, for instance, a platform fee and an outcome-based bonus. And I think this is really where the most exciting aspects of modernization are. Um, and if I look ahead to, um, kind of the future and what my advice is for, for companies now, I'd really try to get super clear on what those outcomes are that you're able to provide for your customer and how you can communicate that so you really get as much credit and attribution for those outcomes as possible. And that's going to serve you well if you want to charge on an outcome basis, but it's also going to serve you well if you want to charge on a hybrid basis, or if you maybe just, you know, want to have a credit model. But now you can help the customer understand the value of a credit and why that lack of predictability is fine because it actually still correlates with more ROI for them. One of my favorite examples of, you know, doing something like this even goes back to the Twilio days, um, so Twilio sells SMS messages and they sell a bunch of other things, but the ability to text is a core, uh, core component. And if you're Twilio, uh, and you're selling to a customer, that customer might have a hard time understanding why they should spend tens of thousands of dollars on Twilio if they have no idea how many messages they're going to send and what's the value of that, right? Just because developers are asking to use this thing doesn't necessarily mean the economic buyer is going to sign off on it. But if you're Twilio, you're going to want to say, okay, how are you going to use Twilio? Okay, appointment reminders. Well, what's your no-show rate today? How much does a no-show cost you? If we're able to send messages via text the day before, an hour before, what, what reduction in no-shows could we realistically get? All right, well, sending a message is less than a penny per message, right? Customers can get much more comfortable with less predictable or more value-based models. If you can actually prove to them the value that your product provides. And that requires going really deep into the customer workflow, how they're approaching that problem today, and helping, helping them attribute your product to, to their success. I just shared a lot of data on, you know, what I'm seeing and, you know, where we're headed, but happy to kind of pause, take a breath and see if there's any questions.

If Kyle can't hear them, I'll relay to Kyle. Who has questions? Can you share examples of more outcome-based pricing in the AI world? Can you share any examples of outcome-based pricing?

Yeah, so I was formerly an operating partner with a fund called OpenView. We have a portfolio company called Chargeflow that has had an outcome model for the last 3 years or so that has just done extremely well with that model. And so you can think of their model as an AI chargeback management. Solution. So customers, you know, maybe it's a small merchant, they sell a, you know, customer product, that customer might not want to return that product, maybe it's outside the return window. And so they just dispute the transaction on their credit card. Well, those happen a lot, they happen more and more. And for a small business, these are really complicated to manage. You have to have, you know, a lot of validation, you have to have documentation, you have to know the language to really get these resolved. And so for many, uh, e-commerce businesses and even anyone that really sells online, the win rate in resolving these chargebacks, particularly fraudulent chargebacks, is only 10 to 20% in the industry. And Chargeflow can say, hey, we're going to take all of your existing data sources and we're going to use AI to respond to these fraudulent chargebacks on your behalf. And we'll only charge you when we're actually able to get money back in your pocket, or we're going to take 25% of whatever incremental money we're able to get back for you. Well, that is extremely win-win. Customers don't have to pay any upfront fees. They, they see this as purely incremental because their win rate was maybe 10 to 20% before. Now it's 60% or higher. And the other thing that I think is fascinating about it is as ChargeFlow is more and more successful, both with a customer and with more customers, they're able to increase that win rate to get it higher and higher and higher. And that starts to create a moat, right? Where even if someone were to offer a cheaper product, if their win rate is lower than Chargeflow, even if they have a lower price, the customer ultimately gets less money in their, in their pocket because they won less frequently. And so I think that's actually an underdiscussed element of good outcome-based pricing is that it creates pricing power if you can really be the market leader delivering the most possible outcomes to your customer. So that's one example. I really love them as a company and their pricing, but Intercom's Fynn Glover product, I think, is a successful example of outcome-based pricing. There are other companies in the CX space like Sierra, Decagon, or Crescendo that have outcome-based pricing models. And I've been seeing it pop up more and more frequently in other spaces as well.

I like Rashik's.

How does the framework translate to consumer?

How does this translate to consumer AI products outcome?

So I think consumers, uh, there's less variability, right, than businesses, right? So that's less complex use cases. Also, the opportunity for them to, you know, rack up exorbitant usage fees is different than if you're kind of unleashing an AI product in a business environment where there can be a lot more adoption over time. So I think in general, consumer pricing is a little bit simpler, a little bit more oriented towards flat rates. Still. But I see creative models as well. I mean, I came across an AI product that helps people that are, you know, traveling and they're taking flights that get delayed, right? Or they get rebooked and they can use AI to actually dispute with the airlines and get money back because a lot of these airlines offer rewards or, you know, additional kind of compensation patterns for for any of the delays if the delay was the fault of the airline. And so there are some outcome-based models that are even in a consumer space. There are also credit models in the consumer space. And I think that one thing that people, people see credits as, and they think of it as like some sort of AI innovation. I've had credits for ClassPass for years, right? I am a Hungryroot a customer for meal delivery, they have a credit-based model. Gaming companies, particularly kind of mobile and online games, have had credit models for quite some time. Credit models are not uncommon in a consumer context, both AI and otherwise.

Yeah, one more question. Anybody else have a question?

Yes. So in terms of the outcome-based pricing, one question that I have on my mind is there are situations where value creation for customers is real and you can see the impact in dollars very quickly. And in which case, I think the outcome-based model works really well, but there are situations where it takes months or maybe years to create value for the customer. What do you, what's your philosophy on that? There's a spectrum of kind of what are the specific outcomes that you're delivering for a customer. So you can go all the way to revenue generated or cost saved. That's very hard to prove and it takes a long time. Like I was in pricing consulting, um, in beginning my career and we tried to charge our clients based on the incremental revenue generated from pricing consulting projects. And that was really challenging because they would roll out the pricing months later and it would take a really long time to see if it worked. Right. So totally have been in your shoes. But there's outcomes that are not going all the way there, but that you have more control over that often happen faster that are more micro-level outcomes. And so I, I think about it as you could have tasks that you deliver, you could have resolved conversations in an intercom basis. So I guess I think of it as there's, uh, successful work products that you can deliver that you can sort of correlate with the end outcome to the customer. And I would try to find things like that that aren't really just like usage or, you know, they feel like they're disconnected with customer value and they're just your costs, but maybe aren't all the way towards true outcomes like revenue or cost savings, but somewhere in the middle. And I think that, that does still allow you to have a conversation about outcomes with your customer while not being as beholden on, you know, waiting for that kind of final, final, um, end business outcome.

Okay. You have, we have interest in more questions. All right. Last question. And then we're going to go to the next speaker. Have you seen any penalties for outcomes not met?

There's a variety of outcome models. The one that I would say what is most common right now that I'm seeing is people having platform fees where there's sort of a base fee for the kind of software and AI and an outcome-based bonus. And that is like depending on the outcome. And so certainly you would not collect that outcome bonus if the outcome was not delivered. It's really important to align on the outcome definition and measurability upfront. If you look at like a— how does Zendesk count an AI resolution? It's actually fairly complicated because if it's using AI to resolve a customer support question, well, just because the customer closed out of the chat doesn't necessarily mean that support question was resolved. They could come back an hour later. With the same issue. They could come back the next day. They could just get mad and churn off the product, right? And so there's— there actually ends up being a good amount of tension in these outcome-based models over making sure that you have the right measurability and that you're aligned up front on what an outcome really means. I highly encourage folks to be very clear up front around that, that measurability, and look for ways that the customer is recognizing the outcome as generated by you. And you're sort of celebrating the success together. I think the biggest challenges come from instances when it feels like you're taking credit for something the customer did, or you're trying to charge them for something that they feel like there was not a business impact. And, you know, as much work as you can do upfront to get ahead of that, the better. But that certainly becomes a challenge with, you know, outcome attribution.

Kyle, thank you so much. Thank you. Happy 15-year anniversary.