Revenue Engineering: How to Price (and Reprice) Your AI Product

00:00:00.000 | Thanks for coming to my talk. I'm Shitej. I'm one of the co-founders at ORB, and I'm going to be

00:00:20.540 | talking about how to think about pricing. Maybe top-level takeaway from this talk is that pricing

00:00:26.460 | is a deep, complicated topic. We're going to cover some examples. We're going to cover some

00:00:30.260 | tactical advice. But in general, the way you should think about pricing is pricing is a form of friction

00:00:35.020 | for your product. And sometimes that friction can be applied for a very good reason. Sometimes that

00:00:40.480 | friction can just prevent people from using your product. And so you really have to think about the

00:00:44.680 | value your product is delivering, as well as the audience you're delivering it to, and the way that

00:00:49.040 | you're building and innovating in your product to start really thinking about pricing. Just a quick

00:00:53.780 | intro to ORB. ORB is a usage-based billing infrastructure company, so we help companies

00:00:58.780 | at lots of layers of the stack think about pricing and billing. Nowadays, a lot of our conversations

00:01:05.220 | first center around monetization and pricing as a practice, and then we get into the weeds of how do

00:01:09.860 | we actually bill for this stuff. So maybe an intro tweet to start. This is Amjad. He's the CEO and

00:01:19.420 | founder of Replit. Replit is one of ORB's customers. And this is a tweet actually pretty recent, starting

00:01:24.520 | to think about how Replit agent should price itself. And so a pretty short tweet, but a lot of complexity

00:01:30.060 | actually packed in here. Really starting to think about, as you have Replit, which is a very popular

00:01:35.920 | agent that does programming, is starting to build full-stack web apps. How should a company like Replit

00:01:42.580 | price their agent? Should it be like a super simple, every time the agent checkpoints, it charges you a

00:01:47.740 | fixed amount? Or should it start aligning with the complexity of the change it made? And as you can

00:01:53.780 | start thinking through this question, you have to think through considerations like who is Replit built

00:01:57.840 | for, right? What are the costs that Replit incurs? But also just the user experience of if I get charged

00:02:03.800 | some amount at the end of something that this agent has done without me really having control over what

00:02:08.720 | the agent did other than the initial prompt, is that going to be surprising? Is that not going to be

00:02:12.840 | surprising, right? So these considerations apply particularly to agents because the work that the

00:02:18.760 | agent is doing is fairly opaque or can be opaque, but they actually apply at every layer of the stack,

00:02:23.800 | including the kind of infrastructure layer of the AI stack. So monetization as a whole is changing.

00:02:29.560 | Maybe let's start with some kind of traditional principles in pricing and how pricing has always been

00:02:35.320 | thought about, right? So we're maybe coming from way back in the day, you know, you're selling a fixed

00:02:40.080 | license, you go to like Fry's Electronics and buy your TurboTax license. And then of course, now you have

00:02:46.600 | seat-based pricing subscription, maybe you're paying $30 a seat, all the way to you have this usage-based

00:02:52.380 | pricing, which is very dynamic. And then maybe you can consider outcome-based pricing as a further

00:02:56.600 | evolution of that. So in traditional pricing, you have a few key principles. You want to make sure that the

00:03:02.460 | pricing is simple so that users can actually understand it. They can grok your pricing. They

00:03:05.600 | know how they're going to pay for your tool. You want to produce some friction because it determines

00:03:10.680 | the value of your product. Like willingness to pay is a very real signal. And then you want to protect

00:03:14.520 | your margins, right? Traditionally, software margins have been very healthy. You know, a good software

00:03:20.540 | margin looks like 80%. Obviously, as you start getting closer to the infrastructure layer and as you start

00:03:26.160 | moving into the world of AI, these margins are much, much more variable, not necessarily unhealthy, but

00:03:31.360 | there's a lot more degenerate workloads that you can incur with an agent. So now starting to think

00:03:37.080 | about AI native pricing, predictability matters a lot, right? I think especially when you're selling

00:03:42.420 | to more mature companies that need to budget and think about their, you know, cost profile using your

00:03:47.540 | tool, they need to be able to predict how that cost is going to scale, not just on an individual

00:03:53.680 | developer, how that cost might scale, but just generally throughout the company's buying process.

00:03:58.380 | AI is very early. So speed and just showing the value of your product and people being able to play

00:04:06.740 | around and experiment matters a lot. Oftentimes, at least in this space, it's not like people are coming

00:04:12.000 | to you buying a very established product. It's kind of your burden of proof to say, you know, here's how

00:04:17.100 | our agent works. Just go try it and then start thinking about pricing. So maybe the friction point and

00:04:21.940 | where it is has changed. And then I was just saying, um, cogs are variable, but, but not only are they

00:04:27.780 | higher, they're also changing very, very quickly. I'm sure everyone in this room seen, okay, great. Um,

00:04:32.440 | you know, open AI, uh, cut, cut model costs by a third overnight. Uh, that has a pretty material impact

00:04:38.960 | on like potentially the underlying cogs of many of these agentic tools. So, um, the three key principles

00:04:45.760 | that, that I think matter here are like really thinking about the audience, who your product is

00:04:50.140 | optimized for, what the value delivery mechanism looks like your margin structure. Um, and, and I say

00:04:56.320 | margin structure and not actually margins, because again, the underlying costs are changing very, very

00:05:01.060 | rapidly. So you want to think about what are the axes of scaling rather than what is the literal margin

00:05:05.980 | on day one, especially if you're an earlier product. And then you want to give yourself the

00:05:10.140 | flexibility to experiment over time. This comes with a lot of complexity, but I think it's very important

00:05:14.520 | because you're just not going to get it right the first time. Um, so, so these are kind of the principles

00:05:18.980 | to keep in mind, maybe going through one by one. Um, the, the way to kind of think about pricing for your

00:05:25.080 | audiences, um, you know, what is their buying journey and what are they coming to you to, to buy and what value

00:05:32.540 | are they looking to get out of it? Right. Uh, if you're like, I, I, we, we talked to lots of people

00:05:37.000 | who are like, Oh, you know, service now is pricing is so silly. You go to their pricing page and it says

00:05:41.680 | like contact sales. It's like, well, that's, that's a very, very different audience that they're selling

00:05:47.240 | to where they're probably cross selling some, some existing contracts. That is a big enterprise commit.

00:05:52.160 | They want you to talk to sales because they really want to tailor the value to their sales motion.

00:05:57.320 | Right. And then on the other hand, companies that are selling to individual developers,

00:06:00.800 | there's not a procurement team. There's no purchasing process. They want you to click,

00:06:04.740 | try it and start using the product immediately. So, so, um, don't just think about the kind of

00:06:11.100 | point of entry, but think about who's behind that at the company. Uh, is it going to be a cross

00:06:15.280 | functional decision? Is it something that someone can just insert into, you know, as, as a vendor on day

00:06:20.500 | one, or is there going to be more of a process around it? And then also, uh, start to kind of think

00:06:25.760 | about, uh, what users are seeking as quickly as possible. If they're seeking a

00:06:30.620 | proof point, then perhaps you want to be able to give them that proof point before you insert

00:06:34.240 | pricing into the equation. Here's, here's an example, again, using replet, um, couple of things

00:06:39.280 | to note here. Uh, you'll see that there's one, a lot of tiers and pricing is actually quite

00:06:45.360 | transparent. Obviously it's a free tier, right? Indicating that they want you to see the value

00:06:49.560 | without having to do a ton of work. Think about the monetization. Um, and, uh, one thing that's kind

00:06:56.000 | subtle is oftentimes pricing and packaging really starts to, uh, dictate what sort of use case they

00:07:03.680 | are imagining you to use, right? Like, um, you know, what models you have access to, how many agent

00:07:09.260 | checkpoints you have access to included, how many seats you might have included. Uh, it kind of starts

00:07:13.900 | positioning your product. Is it a multiplayer product? Is it something that people are going to

00:07:18.660 | be using daily or is it something that you're going to be using, you know, maybe like five times a month,

00:07:22.620 | uh, the way you package it really determines the incentives that you're pushing onto your users

00:07:27.020 | and obviously controlling for the costs that you might pay on the backend. Here's another example

00:07:31.660 | from Unify, which is, I think, uh, uh, agent agent to company, but it's like targeted at revenue and go

00:07:36.780 | to market teams. One, you'll see the, like the price point is a lot higher. Um, I want you to like

00:07:41.480 | forget whether it's monthly or annually or any of that, just like you come on this page, the price point

00:07:45.320 | is higher that communicates something psychologically. And then they're like second duty years are custom,

00:07:50.180 | right. Um, again, I, and, and maybe the third thing to point out here is that 50,000 credits,

00:07:56.180 | 200,000 credits, 600,000 credits. I actually don't know what that means, but that seems like a large

00:08:00.840 | number that communicates something about the workflows they expect to participate in and like

00:08:04.720 | what that might mean for how you use Unify. Right. Um, of course, like it maybe goes instead,

00:08:10.100 | but the set of logos you put on your pricing page is also very important. Um, are you tailoring it

00:08:14.820 | towards, okay, we, we service the best developer brands or we service the like fortune 100 enterprises.

00:08:19.620 | Um, and so pricing is, is not just the price point, but, but of course also the, the whole experience

00:08:25.060 | of what someone sees when they land on this page. Um, let's, let's talk a little bit about margins. Um,

00:08:31.040 | you know, as, as, uh, the, the underlying architecture of your product changes, of course,

00:08:37.320 | your, your margins are going to change. These inputs are changing very quickly. Um, one,

00:08:40.820 | one, one interesting example here is, uh, you, you, as much as possible, you want to think about,

00:08:46.520 | uh, the differentiation of your product also in R and D innovation and try to pass that on to your users as

00:08:52.020 | basically pricing leverage. Uh, a good example of this is like Cloudflare has built their

00:08:57.420 | infrastructure on and their workers infrastructure on these like isolates, right? So, so they're kind

00:09:01.880 | of lightweight compute machines. And so they can charge for CPU milliseconds. And so if someone is

00:09:07.640 | building like an agentic app on Cloudflare and it makes like a, you know, call out to open AI or

00:09:12.700 | Anthropic, you're not getting charged for that because Cloudflare is just charging you for CPU

00:09:16.340 | milliseconds instead of wall time. I think that's a really good example of them, you know, taking an

00:09:20.820 | architectural bet, um, which, which might be good for many, many reasons, but it's a really good fit

00:09:25.500 | for building AI agents or generally AI software on top of Cloudflare. So that's, I see that as an

00:09:30.860 | example of like passing down technical innovation to the end user. Um, and then the other thing to

00:09:35.840 | consider about your margins is you don't have to protect them at all costs. You just have to think

00:09:40.920 | about what are the extreme edge cases and what are you doing to like prevent those outcomes, right? So

00:09:45.940 | if you have these potentially really degenerate workloads, uh, how do you incentivize reasonable

00:09:50.880 | usage? Do you put in rate limits or guardrails? Um, you don't have to kind of linearly necessarily

00:09:56.100 | scale costs with usage. You, you can protect your margins potentially in, in different ways. Um,

00:10:02.220 | this is an example from Jasper. I think this, this blog post and pricing has changed since,

00:10:06.600 | since I pulled this example, but you know, Jasper realized like, Hey, we're, we're targeted toward

00:10:11.180 | marketing teams. Marketing teams don't want to count credits. Like when they're, you know,

00:10:14.560 | basically writing marketing copy. And so, uh, they, they talk a little bit in this blog post about

00:10:20.560 | how they've architected around being able to like switch between models under the hood seamlessly.

00:10:25.580 | You don't have to choose the right model for the job. And as a result of that, they went unlimited

00:10:29.780 | on, uh, each of the three tiers. So you're, you're no longer counting credits as you're using Jasper.

00:10:34.960 | Now, uh, this kind of highlights the core value prop of the product, which is they want to be

00:10:40.060 | a really critical part of that workflow. And they don't want you to have to like pick and choose when

00:10:44.680 | you use Jasper because you might have credits or might run out of credits. So again, kind of trying to

00:10:50.220 | take, is there like an architectural, uh, advantage you have and how can you pass that on to your

00:10:55.480 | pricing model? Finally, I want to talk about flexibility. Um, you know, you, you all, as,

00:11:01.500 | as you're working in your teams, uh, you're constantly improving your product. You're putting

00:11:05.280 | a lot of R and D into it. Um, one thing that is used to be true, especially in the seat based world

00:11:10.440 | is like, you'd have R and D over here shipping a ton, um, increasing the product value. And then you'd

00:11:16.080 | have like a once a year, okay, we're changing our seat price from $20 to $25. Um, one advantage

00:11:20.980 | you have in the AI world is like, uh, the value is much closer to the end users. It's not necessarily

00:11:26.500 | like a $20 a seat thing. It's like someone is literally seeing, okay, I'm, I'm taking this

00:11:30.780 | action. This agent is giving me this output. Okay. It's getting better. It's getting, it's

00:11:34.760 | better, it's getting better. So I think not only do you as a business have a responsibility

00:11:38.740 | to incrementally evolve your price as, uh, you know, you invest R and D, but I think it's,

00:11:44.780 | it's honestly a lot more understandable for customers why that's happening because the,

00:11:49.000 | the monetization is much closer to the value of the product. Um, so, so, uh, you know,

00:11:54.420 | obviously in orbs customer base, but in general, we are seeing people, uh, make many, many more

00:12:00.020 | price changes, not just once a year or twice a year, but in fact, some customers making price

00:12:04.100 | changes, uh, two, three times a month. Um, and again, there's, there's caveats that come

00:12:09.760 | with that. Uh, you have to think about, you know, what, what is that complexity that introduces in

00:12:14.500 | the, in the mental model of your end users, how are you going to manage change? Um, and a kind of

00:12:20.120 | good internal dynamic example of this is, uh, if your go to market team is, you know, commissioned

00:12:25.900 | and, and they're selling these big enterprise contracts and your pricing is changing, or you're

00:12:29.480 | going to like a pay as you go model, you need to think about like, what does that mean for your

00:12:33.040 | sales team? What does that mean for the incentives that they have in selling your product?

00:12:37.020 | classic example of this is a lot of usage based companies now commission their sales reps, not

00:12:41.620 | on the initial contract land, but on the expansion of that account over time. And then, okay, second

00:12:47.480 | order effect. How do you think about customer success versus sales as an organization? And how do you

00:12:51.320 | organize that internally? Right? So, so repricing can have lots of impacts, not just on your end user,

00:12:56.300 | but also in the internal dynamics of your org. Um, finally, the thing I'll say is, uh, as you're

00:13:01.800 | iterating, you, you really want to, as much as possible, be able to simulate the impact that this

00:13:07.740 | pricing change is going to have on your users. In particular, I think, uh, you want to be able to

00:13:12.460 | simulate, uh, different usage patterns. So if you think about your users and cohorts, um, some might

00:13:17.840 | be using it for very different use cases. What is that revenue mix going to look like? What is your top,

00:13:22.840 | uh, revenue generating customer going to, how are they going to change with a new pricing model? You want to

00:13:27.920 | be able to do that in a, in a data informed way, especially if you're running at a, at a reasonable

00:13:32.420 | scale. Here's some predictions I have as, as we're wrapping up for AI agent pricing. I think, um, you

00:13:39.380 | know, the, the price wars will continue. We're seeing agents get cheaper and cheaper. Um, and, and, you

00:13:44.780 | know, I think you're, you, we should expect people to try to give effectively unlimited plans. I think

00:13:51.040 | there'll still be caps. I think there'll still be guardrails, but I think, um, we, we will move to a

00:13:55.700 | world where, uh, you know, people just want these agents embedded in every workflow. So we'll try to,

00:14:01.380 | try to converge to effectively unlimited. Um, I think outcome-based pricing is going to get more

00:14:07.260 | real. We are seeing some companies, uh, do that already, but the, the, the hard part about outcome

00:14:12.700 | based pricing is defining what the actual SLA is like, what, what is the outcome and how do you measure

00:14:18.300 | it? And does it make sense to the end user? Um, I think like, we're going to see a lot of clear

00:14:23.780 | definition of success language and contracts and like people really trying to get, uh, precise about

00:14:30.260 | what the outcome is that they're charging for. And then, um, I think there's going to be a lot more

00:14:36.040 | real-time visibility, spend management, uh, balance alerts. I think the kind of workflow around pricing

00:14:41.860 | and monetization is going to get, uh, more and more sophisticated. You think about like, you know,

00:14:46.860 | obviously in, in database queries, you have like query planners. I think it's quite possible that,

00:14:50.700 | you know, you, you prompt an agent, it tells you like, here's my estimate for how much, uh, how many

00:14:55.780 | credits I'm going to use, or, Hey, here's three options for, um, how I might execute, how I might plan

00:15:02.800 | this workload. One's going to take 10,000 credits. The other is going to take, you know, 4,000 credits.

00:15:08.280 | I think that customer experience of like being able to control the workflow around the spend

00:15:13.300 | is likely to get much, much more sophisticated, um, this year. So, uh, just, just final note,

00:15:19.200 | um, we have a whole ebook on how to price your AI agent that goes into much more tactical advice,

00:15:23.440 | talks about specific, you know, threshold billing, fraud, prepaid pricing models. Um,

00:15:28.300 | if you want to check that out, that is at this link. Thank you.

00:15:31.540 | Thank you.

00:15:32.540 | Thank you.

00:15:33.040 | Thank you.

00:15:34.040 | Thank you.

00:15:34.040 | I'll see you next time.

Revenue Engineering: How to Price (and Reprice) Your AI Product — Kshitij Grover, Orb

Chapters