The Price of Intelligence - AI Agent Pricing in 2025

Hey, I'm Kshitij. I want to talk to you about how you should think about pricing AI agents. Now, there's a ton of different examples in this talk, so let's dive right in.

I'm the co-founder and CTO at Orb, and Orb is a usage-based billing infrastructure company. So we work with a ton of different companies across AI, infrastructure, developer tooling, and more, and we really help them think about not only the monetization of their business and how they should evolve it over time, but also, of course, specific billing implementations

of different pricing models that we're seeing across the industry. So today, I want to talk through a bunch of different parts of the AI agent market, but making generalizations about this whole market isn't easy.

So we'll try to pick examples from everything you see on this list, from customer support and sales, all the way to specific enterprise search use cases or enterprise agents, and we'll do our best to talk about pricing strategies that you should be thinking about as you price your AI agent.

Now, let's flip through a, a couple examples really quickly. Here you see Intercom and their AI agent, Fin. Now, obviously, Intercom is a super successful company, has been around for a while, but they're really leaning now into outcome-based pricing.

So you're seeing a essential, advanced, expert sort of tiered model, which is very traditional. But then on the top, you're seeing a ninety-nine cent cost per resolution. So here, they're really making on Fin being a, a great AI agent, really helping you with these resolutions and customer support tickets, and you're only paying for the success of Fin

as a feature and as a service. And so that's, that's an indication of how confident they are in their product, and they're aligning their pricing to match that. Here's another example.

Unify is a go-to-market tool, and again, you have a good, better, best growth pro enterprise model, but there's a couple other things you'll notice on the page. You have your pricing calculator because their pricing is fairly complicated.

You have credits, which are incredibly common now in the AI agent space, and you have a bunch of different axes of pricing. So you'll see this a lot where there's specific usage limits or caps.

There's often a seat-based price built into each tier. Um, and then each of the features that your platform offers will have some sort of usage-based or hybrid pricing.

Here's Cursor. Here's an agent that everyone knows, I'm sure. And on the surface, their pricing looks pretty simple, right? They have a free price, a twenty dollar a month, forty dollar a month price, but there's actually a bunch of complexity hidden under the hood.

So they have completions versus requests, different usage limits. They have fast versus slow, and even on the specific models, there are certain models like GPT-4, 4o, and Claude 3.5 Sonnet that they consider premium models and that either have different usage caps or limits or cost a little bit more.

So even a tool like Cursor, kind of aimed at the everyday developer, has a ton of complexity when it comes to its pricing. Here's another interesting example. This is an example of Chargeflow, which is a chargeback recovery tool, and they're charging you a percent per recovered chargeback.

Now, this is interesting because, again, it's leaning into that outcome-based pricing, where really you're not paying anything until chargeback recovery is successful, and they're giving you what is effectively an ROI guarantee on their tool.

So let, let's talk about some key principles with those examples in mind when you're thinking about pricing your own AI agent. Um, I think what's really important is that these principles don't actually differ specifically for AI agents.

This is just how we should think about pricing as a practice and how great pricing has always been done. But we'll pu- pick some examples from AI specifically just to highlight how this applies to your industry.

Now, let's talk about the first one, which is considering your target audience. When you're thinking about building an AI agent, or of course, any product, you're building it for someone, and that someone is gonna be a buyer and a person at a company, and it's, it's not just who they are, but the whole buying process they're involved

in. So if you're selling to an SMB, you might have maybe an individual developer going in, entering their credit card, and checking out. If you're selling to a Fortune 100 or generally an enterprise company, you probably are dealing with a procurement team that's budgeting your software against other, perhaps more traditional solutions.

So you have to keep in mind where your audience is and how they're thinking through the purchase of your tool. Now, two things that almost never should be compromised are simplicity and predictability.

Especially with these usage-based models, uh, the predictability of spend over time is really important, and people just have an easier time making a purchasing decision when pricing is simple.

And finally, with these AI agents and AI at large, you're often not just thinking about pricing in the context of the initial land. You're thinking about it as the usage of your product expands over time.

And so when you're setting your pricing, you'll want to think what use cases you're actively encouraging or discouraging and what sorts of workloads your product is a great fit for, because those are sorts of things you want to encourage.

But let's talk through a couple examples to make this concrete. So let's talk about audience. Here's an example from, uh, Clay, which is a go-to-market tool, and you'll see a couple things off the bat from their pricing page.

They're really highlighting in bold colors this Explorer tier and this Pro tier. This tells you about what their potentially typical price point is. Um, even when you're seeing this people company searches, that's indicating kind of how people use Clay to prospect and how many search results they can expect.

So, you know, ten thousand, twenty-five thousand is a lot, so it's, it, it's leaning into a specific sort of persona or use case. And finally, you, you'll see these logo gardens actually quite a bit in pricing pages.

And Clay's is telling you this story of companies like OpenAI and Airbnb and Anthropic and Canva, these really fast-growing companies that use Clay, not just traditional, again, Fortune 100 or enterprise companies.

So it's telling you a story about who Clay is for. Here's another example which is quite different. This is Replit, and of course, Replit is a programming tool and, and Replit's mission is to make programming accessible to, you know, everyone across the world.

So it's a much lower price point. They emphasize that they also have a free tier. Um, and what's really important is, uh, it's a quite technical pricing page. So you're seeing agent checkpoints.

As you get to these add-ons, whether it's auto scale or deployments, you're seeing very granular pricing. And so this is telling you something about who Replit is built for and what story they're trying to tell with the accessibility of their product.

Now, one thing I'll often hear when it comes to pricing is people being annoyed by pricing pages that kinda look like this, right? There's no real price point here.

There's not even a good, better, best. It's just book a demo. But one thing I wanna be clear about is these can be incredibly, incredibly effective. It just comes down to your target audience and who you're actually selling to.

So in this case, HeavyAI is really oriented around the enterprise, and so it makes a lot of sense for them to push you towards a demo because h- that's how they want your purchasing decision to look.

Similarly, ServiceNow, obviously a much, much larger company than many companies in this space, has an AI agent's landing page, but again, they push you towards a demo. There's no price point here.

You'll see in the logos that they're fairly enterprise logos, and so ServiceNow is almost definitely cross-selling their AI agent product to existing products and existing companies that they already have a relationship with.

So they do want a salesperson in the loop. This is not intended to be an individual developer seeing this page and signing up for ServiceNow, because that's not their target audience.

Let's move to a different example, which is AI SDRs. Now, this one's interesting because it's a flat rate per month, and there's a lot of different indications on this site about, you know, how you should think about an AI SDR, their output, and their effectiveness.

You'll see there's an email sent number, a meetings per month number. And the thing that stands out to me here is it's actually pretty easy to compare an AI SDR to what you might already be employing in your SDR team.

So you're, you're thinking about the adoption journey of one of your customers who's, who's really making a staffing decision of, "Should I be staffing more SDRs and hiring more people, or should I be considering these AI SDRs?"

And in that adoption journey, it's really important that they can figure out and benchmark the productivity of your AI SDR service.

Here's a final example, which is Devin, of course, from Cognition. And Devin is supposed to be this engineer that can work on your team that can really integrate into your existing workflows.

Now, typically, of course, you're paying engineers a salary per year, but with Devin, you're paying a cost per month as well as having to think about ACUs. So ACUs are compute resources effectively, but, you know, they, they scale with virtual machine time, inference, networking bandwidth.

And this is interesting because when I'm thinking about staffing an engineering team, I'm not typically thinking about these things. I'm just thinking about heads and maybe milestones or costing, but it's pretty hard to translate that into ACUs.

Now, all that to say is I think this might work for Devin, but adding these new paradigms does have some cost associated with it because now buyers need to think about their workload and translate that into ACUs.

Let's move on to the next point, which is costs and margin structure. Now, there's this famous Jeff Bezos quote about, you know, your margin being, uh, his opportunity. And what this really points at is in typical SaaS, you often have 80, 90% margins, and oftentimes that doesn't last forever.

Uh, there are competing products that can often erode that margin. So what's really important is, especially in this evolving AI agent landscape, not necessarily to over-index on cost, because oftentimes cost will actually change pretty rapidly, but to understand the axes of cost or your cost structure.

Now, some proxies to get at this are thinking about different use cases and workloads. Oftentimes, AI agents, even if they're targeted at a specific vertical, will actually be used in very different ways.

Some, some companies very sparsely, some companies will use them a ton. And it's, it's not necessary that you have to defend your margin in every case, but you should think about what are the degenerate cases, how are you gonna defend against those, at least roughly speaking, and which are the workloads that you, you really wanna promote and

incentivize on your platform. Ultimately, it's your responsibility to defend your margin, and one of the best ways to do that is to take the R&D, the ultimate technical innovation of your company, and pass that down in your cost structure.

So let's look at an example of that. And before we do that, of course, just talking about the different costs, there's models, training, and ops. Oftentimes ops i- is a little bit ignored, and for AI agents, usually the cost is gonna come in inference, in, in running the actual agent model.

So here's an example from Character.ai. So Character.ai, of course, is a massive consumer product. Um, they have a ton of traffic on their platform. Uh, I think at-- in this blog post, uh, back in June of 2024, they were serving something like 20% of Google Search queries with 100 million daily active users.

Now, they invested actually very early on, all the way back in 2022 to 2023, in, in optimizing their inference infrastructure, and that made them sustain a B2C product where people are spending hours or, or almost an hour per day on their service.

So this is an example where in order to make their business viable, they really had to invest in their inference technology. Here's another example. Jasper is a marketing tool used to generate marketing copy, and what they found is that as marketers are, are, you know, using their brand voice product and trying to, uh, generate output with Jasper,

it's very hard for them to think about things like word count. And so Jasper actually was able to move to an unlimited credits model on some of their paid tiers, and this is really important because it supports their core value prop of being able to iterate really fast on marketing copy.

And the underlying technical work here was what they call some sort of model decision-making engine that can pull from a variety of different models, whether it's OpenAI, Anthropic, or Cohere, to, to pick the right one for the job, and that allowed them to reduce their costs and offer this unlimited credits model.

The final point I wanna talk about is pricing flexibility, and this one is perhaps the most important because you really don't want to paint yourself in a corner.

Pricing really just can't be a set it and forget it exercise. Uh, we often talk at Orb about pricing product market fit. So obviously, you're iterating on your product and trying to build the best thing for your users and your prospects, but you also really need to be measuring their willingness to pay it every time, right?

Es-especially as these industries mature, the price points become more and more important. And as your inputs change, you need to set yourself up for flexibility because the industry's gonna evolve, your customers, the market, how they buy, the use cases are gonna evolve, and you don't wanna be stuck on a different pricing model that perhaps that's just not

the way that people expect to buy from you. And finally, of course, your product will change, right? So as you increase your R&D and as you increase the value that you're delivering to your customers, you wanna be able to capture that value.

And here's actually a good graphic that I think really gets at that. Most companies think about pricing like this once a year or maybe twice a year change, and that's pretty sudden.

It can be harmful for your user relationships, and ultimately, you're just not capturing the value that you're building. Instead, you wanna think about pricing like this continuous exercise in evolution, where as your product value increases, you're able to capture that and bring your customers along for the ride.

Here's a good illustration of why flexibility is so important. So of course, what this graphic is showing is that OpenAI's, uh, cost per token has decreased quite drastically, in fact, over the last year, year and a half.

Um, and that's really important because, of course, these inference costs or, you know, OpenAI costs are an input to a lot of these agents. And so you might need to change your pricing as people expect more competitive pricing for inference or for the, the kind of underlying model.

But perhaps even more fundamentally, there might be use cases that seem out of reach today that really just don't seem possible that will make a lot of sense in a year or two years.

We, we're seeing this across healthcare AI and legal AI, where there's a ton of data to digest, which, you know, maybe the token counts don't make sense in 2024 or 2025, but it's very possible that these use cases can be unlocked within the next one to two years.

And so it, it affects not just your pricing, but also the product landscape available.

And finally, flexibility is not just about shifting a dollar price point. So this is an example from Luma Labs, and there's a ton going on on this page. But what you can see is there's a lot of levers, right?

There's the platform that you're running Dream Machine on, whether that's web or iOS. There's whether you're in relaxed mode or not. There's a credit system, um, and there's like rate limits available.

And so when you think about flexibility, you wanna maximize, again, within the bounds of simplicity for your audience, the, the number of things you can change, because ultimately, that allows you to meet your customers where they are.

One specific pricing model I wanted to emphasize is prepaid credits, just because of how popular and how common it is. Prepaid credits is, is good for a couple of things.

It lets you bring in immediate cash flows, so oftentimes there's that upfront payment like we saw with Clay, and that's very important for a COGS-heavy business where perhaps you can't afford to take even that one month sort of pay-as-you-go credit risk.

Now that... There's a kind of interesting corollary to that where oftentimes prepaid credits are used a-as a way to counteract things like fraud because you're asking the user to put down money up front for their use case.

Prepaid credits, interestingly enough, are also a pretty easy discounting mechanism. You'll see a lot of AI companies have four, five, ten different SKUs, each with a different unit rate, and now when you're discounting, you can just discount one conversion rate from credits to dollars instead of having to discount every line item.

There's also a lot of fluctuating demand in AI products. There, there's industries that are very seasonal, and so if you have a prepaid credit model where you grant all your credits up front, people can choose how they burn those down over the course of a year, which can be very attractive for the consumer.

And finally, prepaid credits generalize across very different use cases. So at Orb, we see prepaid credits as a trial motion for very small companies, but of course, as a commitment motion for the largest multi-million dollar deals, and so prepaid credits can work in both use cases.

Here's what I think is gonna happen to AI agent pricing in 2025. So of course, as competition heats up, the price wars are gonna continue, no doubt. And what I think this will lead to, of course, is that, uh, we're gonna have a race to the bottom on some pricing in some verticals.

On the other hand, I think we'll continue to have COGS pressure or margins pressure from, uh, the market and venture capital and investors. And so you'll, you'll kind of see both of these tensions at play, and one particular outcome I expect is, uh, you'll see more companies try to get closer and closer to offering effectively unlimited plans

where there's not very much of a usage limit. And again, I think that'll be a function of the inputs getting more commoditized, as well as just competition heating up to a point where that needs to be the case.

The second thing I think will happen is we will lean into this outcome or success-based pricing, but we'll need to get real about what exactly are the guarantees and the SLAs that each of these providers are providing for you, right?

And, and so we'll see more discussions around clear definitions of success.

And finally, this one I think is underappreciated, uh, there's gonna be a lot more R&D investment in monetization and pricing because you'll need to provide your customers a lot of control over how they use your product, uh, throttling use cases, being able to set spend caps, being able to see exactly how they're burning down their credit allotment

over time. And I, I think this is gonna be really important because ultimately at, at the end of the day, your users are gonna wanna be able to audit their usage very, very carefully so they know how much they're spending on their AI agents.

There's a lot of technical challenges in that even 2025 roadmap, a lot of complex business logic. Of course, a- as you get these more complicated pricing models, you still need to layer on things like enterprise agreements and discounting and ramps, and so you'll start to see companies leaning into that complexity, at least in some markets.

Again, with this customer experience and visibility point, that can be quite a bit of a challenge to maintain that whole product experience. And with flexibility, you'll see companies make a ton more pricing changes, and that, of course, can be its own technical challenge as you have customers on legacy price points.

And so whether it's from the input, high volume data infrastructure, the kind of core billing business logic, or the output financial accounting, I think you'll see technical challenges at each stage of the billing stack.

And of course, that's, that's what we do. We-- we're a billing engine that's really specialized at these use cases and, and that really thinks about each step of that journey we just outlined.

One particular feature I wanna highlight is we think a lot about price changes. We, we go through a lot of them, oftentimes now many times a month with, uh, single customers, and this is a, a really fast evolving landscape, and so having a billing system that can have versioning and migrations as a first-class feature set is very,

very important. So yeah, that's my talk, and, uh, hopefully that gave you some insight into AI agents. You can reach me at my first name @[REDACTED:email_address]. Thanks so much.