With the rise of AI systems, especially autonomous agents — or agentic systems as they are called in the latest parlance — pricing becomes an interesting issue.
Vendors have chosen several paths that buyers need to assess and understand carefully, as there is often a mixture of them with limited transparency. These pricing models are often used in addition to more “traditional” SaaS pricing models like per-user pricing. It is easy to understand that a per-user pricing model is incompatible with autonomous AI agents. If there is no user anymore, just the AI at work, how do you charge for “bums on seats”? This is also the main driver behind Salesforce’s shift to per conversation, i.e., usage-based billing for its Agentforce offering, besides investors being concerned about revenue loss .
AI Pricing Models
On a high level, we see the following models in the SaaS world:
- All-in pricing. The subscription price already covers the AI features that the software delivers. This method treats added AI capabilities simply as continuous improvement of the delivered software. Vendors that offer this type of pricing include Oracle and Zoho.
- Subscription pricing. Customers pay a monthly or annual fee for access to (some of) the vendor’s AI capabilities. There are also often tier levels, where higher tiers offer more or more advanced AI capabilities. Companies that apply this type of pricing include Salesforce for some of its Einstein features, HubSpot for some of its features, and Microsoft for various AI services.
- If all capabilities are part of a single subscription at a fixed price regardless of usage, then this can also be called flat-rate pricing. This model is also used for perpetual licenses. Companies that apply this model include Atlassian, Dropbox, and Basecamp.
- Feature pricing. This is a variation of the subscription pricing model that requires different subscriptions for different AI capabilities. It is similar to pricing tiers, except that additional capabilities can be freely selected. Companies that include this type of pricing are SAP, IBM for Watson capabilities, SugarCRM for its insight capabilities, and Zendesk for some of its capabilities.
- Usage-based or pay-as-you-go pricing. AI capabilities are paid for according to usage, measured by several KPIs. These KPIs can include API calls, CPU or memory consumption, or the volume of processed data. This is widely used by companies like Amazon Web Services, Google, and some Microsoft Azure services
- Token-based pricing as, for example, OpenAI offers. Each operation consumes tokens based on the complexity of input and output. In text-based AI scenarios, a token is often roughly equivalent to a syllable. In other scenarios, this is harder to identify. For example, images are processed and charged for by visual units, which are essentially smaller parts of the image, even down to the number of pixels. Token-based pricing is essentially a form of usage-based pricing.
- Outcome-based pricing is a revenue model that ties price to results, outcomes, or value generated by the AI system. Customers pay based on predefined performance metrics. These can be cost savings, increased efficiency, or tickets resolved. Vendors that use this model include C3.AI or Zendesk.
- Last, but not least, is hybrid pricing. Hybrid pricing combines different models for different AI services. What is often seen is subscription pricing in combination with feature pricing, or token-based pricing, depending on which additional AI services are added to a contract. For example, Microsoft offers subscription-based pricing for Azure and allows customers to add usage-based components like machine learning. Salesforce includes basic Einstein features in its subscription price and offers usage-based for some of its more advanced analytical features. SAP offers subscriptions for its software and has feature pricing for advanced AI capabilities.
- Freemium models are a variety of hybrid pricing models. They offer a basic version of the AI for free, while advanced or “premium” capabilities are available in a paid version. This model is regularly combined with a tiered pricing model. The tiers are based on users or usage, or a combination thereof. This is a common pricing model used by vendors like Zoom, Canva, and many more.
Finding a Balance Between Customer and Vendor Value
In brief, most of these approaches focus on seats, activities, or outputs as approximations for what customers want, which are outcomes or (business) results. Incentives for vendors are not aligned with those for the users.
Let’s take per-seat pricing as an example. Heavy and occasional users pay the same price. Assuming heavy users are heavy users because they gain a lot of value from their use, the value they receive is high. Conversely, occasional users may not use the AI capability often. In effect, they are subsidizing the heavy users. Interestingly, this leads to a problem for the vendor, especially, if more and more automations (i.e., autonomous agents) come into play. A 2023 Techradar article suggests that Microsoft loses significant money on its GitHub copilot. The same is probably true for Canva, which announced quite a steep price increase for some of its plans this September.
In essence, per-seat models offer predictability for customers and vendors. The vendor at the outset shares some of the risk, as wrong usage calculations may lead to losses. This comes from the different incentive structures that vendors and customers have: Vendors, offer low prices to drive adoption; customers, use heavily, to drive benefits. The rift opens when the vendor increases the pricing.
Usage-based metrics use API calls, or conversations, as a proxy for value. The argument is that the higher the use, the higher the value that the customer gains, as they wouldn’t use the capability else. Sincerely, this is a poor approximation, as it, for example, heavily relies on how efficiently the vendor (or an implementation partner) has implemented a capability or how close system outputs come to desired outcomes. Usage-based metrics have a low barrier to entry, as they initially come at a low price, driving adoption. They usually are quite transparent due to modern systems’ ability to collect real-time metrics. However, with adoption, some cost surprises may arise; and, as they are often combined with other metrics (like users), a complete cost overview can get lost due to the complexity of the overall pricing.
Although, in theory, they should incentivize the vendor to efficiently deliver to customer requirements, this cannot be taken for granted. Instead, one often sees a decoupling of vendors and customers, especially where the vendor may become “sticky.” Stickiness is often the easier route than higher efficiency, quality, or delivery to expectations. For all usage-based and output-based pricing models, customers should tie price to value via desired outcomes. However, this is more difficult than it sounds.
Customer-Centric Pricing
From a customer point of view, outcome-based pricing is ideal. With this type of pricing, a desired outcome is tied to both, a value and a price. They are centered around the customer and the customers’ strategic or tactical desired outcomes. A simple example is case resolutions, like Zendesk recently introduced for its AI agents. A ticket is either fully resolved by the AI agent or not. If it is resolved by the agent, Zendesk charges, otherwise not. Vendor and customer objectives are aligned. The more cases get solved automatically, the more the customer service team can concentrate on more difficult ones, so they get resolved faster. With increasing automation, the cost of service goes down while quality goes up. These are tangible values for the customer. Zendesk wins with increasing automation, too. This effectively aligns vendor and customer objectives while sharing the risk.
I wish more vendors would opt for outcome-based pricing. What do you think?