Capturing part of the value generated for customers has always been a north star in pricing. If customers don’t benefit from using your product, they shouldn’t pay you for it. On the flip side, if they’re getting tons of value from it, you should be getting paid accordingly. It’s about aligning incentives, and different pricing models can help tie these incentives together. We’ve seen different attempts to solve this in the past, from seat-based to usage-based pricing, creating an endless dance between software providers and buyers as they each try to take a larger share of the pie. But now that AI companies are taking off, do we need to rethink software pricing models? Can AI companies more directly tie pricing to actual customer outcomes? The pricing dance historically has been driven by misaligned incentives, with software sellers trying to get more money for the same product, and buyers trying to get more use for the same money. Outcomes-based pricing could simplify that, much more clearly tying spend to customer value. Sure, there may be some bumps along the way as both sellers and buyers adjust to new pricing models, but the clear money-to-value relationship will be worth the effort.
How it was done in the past
Traditionally, seat-based subscription pricing was the most common “value-based” pricing strategy in SaaS – as more users started using the software (and supposedly getting value from it), you would get paid more. This indirect tying of incentives could lead to some less-than-ideal dynamics. Customers would try to pay as little as they could while getting as much value as they could – e.g., paying for just one subscription and sharing a login across employees (like consumers do with Netflix logins) – while software companies would try to justify never-ending price increases with incremental feature releases (“With this new update we added better filtering! Oh also prices went up 50%. Enjoy.“) Both software providers and users try to ameliorate some of these market inefficiencies through long-term fixed-price contracts, TOS prohibiting account sharing, etc., but fundamentally there will always be a conflict where each party tries to pull a fast one on the other by capturing additional value for themselves without giving any up to the other side. This teetering equilibrium between software companies and their customers becomes even more precarious when AI enters the equation, where software suppliers’ costs are more directly tied to product usage and API calls to model providers.
There’s also been significant adoption of usage-based pricing in SaaS recently, where software pricing is more closely tied to usage metrics like API calls, # of reports generated, the volume of data processed, etc. Just across our portfolio, the percentage of companies with usage-based pricing has more than doubled over the past couple of years, and we’ve seen a whole wave of companies that have been founded to address the complexities that come with usage-based pricing. My colleague Max did an in-depth analysis of usage-based pricing and billing infrastructure a couple years ago, and that work is even more salient today given the increasingly-complex AI products coming out.
Aside from an ever-increasing shift towards usage-based pricing I mentioned above, a new software pricing strategy that we’ve seen emerging in the last couple of years is outcome-based pricing. Simon-Kucher laid out their thesis on outcome-based pricing in 2021, when outcome-based SaaS pricing was still early. Typically, more traditional services like managed IT, value-based healthcare, or consulting, “fees-at-risk”, services were priced based on the business outcomes and agreed-upon milestones. In the past couple of years we’ve also seen some SaaS companies price some products based on outcomes. Within our portfolio, Vantage has a cloud cost-optimization “autopilot” that charges 5% of the savings it identifies. As AI apps encroach on the more traditional services, we’re seeing AI companies adopt some of the same outcome-based pricing models, especially when the use case is replacing human work.
How things are changing with the rise of AI
In today’s world of AI services, software providers’ costs are directly tied to product usage – as customers use the product more, there are more calls to backend AI models, and the costs of serving the software increases. Because of this, it’s easy to end up in a situation where heavy users drive up costs to the point where they’re a negative gross-margin customer, and freemium costs are also a lot higher (see OpenAI’s low gross margins). With traditional SaaS this was never that big of an issue, as software costs were relatively limited to cheap cloud calls and data storage costs, but with the increasing expense of foundation models and compute, AI products face higher dynamic costs. Seat-based pricing might not make sense anymore.
Think of consumer / prosumer AI chat apps like ChatGPT or Perplexity. For a flat monthly fee, you get near-unlimited usage of the chat app. This again leads to some of the tensions we saw with seat-based pricing in the traditional SaaS world. Low-usage customers will question why, if they use the product once a week, they’re paying $20 per month, while high-usage customers will be more than happy to pay $20 per month while using the product 100 times a week (and costing the provider much more than that $20 per month. Even at $200 per month, OpenAI is still losing money.). On the flip side, the chat app companies can also pull out a couple tools to mitigate this by throttling users, directing them to lower-cost models, or tiering pricing, but it again leads to a seesaw of value between customers and software providers.
The best pricing solutions in today’s AI-dominated software world tie customer and provider value together much more closely than purely seat-based models, and align incentives so that as customers get more value out of the product, some of that value still passes on to the provider. The more work the AI does for you, the more it should get paid. If it’s a bad product and can’t do the work, it shouldn’t get paid, but if it’s able to book you X customer calls, generate Y hours of video content, or drive Z ad impressions, it should get paid commensurately.
AI agents clearly demonstrate this shift in pricing models
We see this very clearly with AI agents – notably Intercom and Zendesk, which were two of the first companies to align their customer service AI agents pricing directly to the number of tickets resolved autonomously. Intercom, for example, charges $0.99 for every resolution that their AI agent, Fin, successfully resolves without escalating to a human. This clearly ties the price of the AI agent to the value delivered to the company – human time and cost saved. As to how exactly they got to the $0.99 number, in this case we can see that the average cost for a human to resolve a customer ticket is $5, so Intercom is saving their customers $4 per ticket and capturing 20% of the value. The more tickets Fin can resolve, the more money Intercom’s customer saves, and the more money Intercom gets paid. The pros are very clear, with incentives perfectly aligned and value being split fairly among all parties.
Another way of pricing AI agents, especially those that are taking on larger roles within a company like AI SDRs, lawyers, coders, etc., is to think of how much you’d normally pay a human in that role. An SDR in SF, for example, has a median total comp of ~$100k. AI SDR companies like 11x and AiSDR are charging anywhere from $10k to $50k per agent, showing how wide the range of value capture can be. In this case, you can “hire” an AI SDR for only 10-50% the cost of an actual human. OpenAI is planning on going even more upmarket, charging up to $240k a year for high-end agents supposedly replacing PhD-level researchers and scientists. Some companies, like Regie.ai, offer a mix of standalone SDR agents as well as a platform for human SDRs to enhance their workflows – letting the customer choose which pricing model and where on the spectrum of copilot to full agent they want to be.
Now, these outcome-based pricing models also open up AI agent startups to some new risks as well. How do we even attribute outcomes to AI agents? If an employee isn’t performing, companies usually put them on a PIP or fire them. The same will be true of AI agents. If your tech isn’t performing up to expectations, expect customers to “fire” your agents and churn. Much like in the workforce, it will be a true meritocracy for AI agents – no more hiding behind annual contracts or flat-fee pricing structures. Customers, too, need to make sure they’re ready for the efficiencies AI can bring. What’s the point of your AI SDR booking 100 intro calls if your AEs can only handle 50? (that is, until we get AI AEs) As with traditional outcome-based pricing models, modeling revenue and costs will be a bit harder, but it’s worth the simplicity in pricing and packaging.
All that said, it is still early in the outcome-based pricing AI agent world and we’ll have to see how both AI companies and customers respond to the rise of new pricing models. Are companies like Intercom and Zendesk pushing the envelope too far? Will anyone pay $200k for an AI agent? Or will buyers realize the benefits of pay-for-work agents and get comfortable with these new models? Hopefully it’s the latter, but we’ll have to wait a couple of years for things to play out.
Acknowledgements:
Thank you Pete Giordano, GTM advisor at Scale. He is a pricing guru and was a great sounding board for our ideas.
