Pricing models for B2B AI products
how to think about usage-, value-, seat-based pricing
Hi! Welcome to the last of a three-part series on building AI products.
Scroll to the bottom for more resources and related readings.
Generative AI introduces a dirty word in the traditional SaaS model: marginal cost. This makes it tempting for some to consider a dangerous pricing model: cost-based pricing. Well, to be fair, many are not seriously thinking about pricing model at all because they’re racing to have a “AI strategy”.
With the maturation of product led growth, it’s prudent to think about pricing models and monetization early in the product development lifecycle. There hasn’t been much exploration into this topic, so I researched how some of the popular AI products are adopting usage-based, seat-based, and value-based pricing and compiled some patterns below.
In this post:
the value pyramid of AI products
three pricing principles
breaking down usage-, seat-, and value-based pricing
Value pyramid of AI products
This is a generalization but there seems to be 3 basic categories of AI products and a likely pricing model for each.
These large sophisticated models and AI infrastructures are the foundation of many SaaS products. Soon enough, their costs will become commoditized like AWS. These companies often charge usage-based as they’re basically a utilities cost.
That said, companies that use these utilities will need to layer on new differentiation to create unique value for customers.
Examples: OpenAI, Anthropic, Databricks
Layering on top of the infrastructure are horizontal solutions that carry out a workflow to support the user’s productivity, such as customer support, note workspace, spend management. These products tend to add LLMs in a chat interface or to automate parts of the workflow to further improve the user’s productivity.
They’ll use foundational models as the base but the innovation or new value created is the new productivity unlocked within the context of that workflow. They often charge per seat for access to this new value.
Examples: Notion AI, Gong, Brex
E2E (end to end) Work
Further layering on an augmented experience are products that complete an end to end workflow/task in a specific vertical use case. They’re able to codify expertise in a specific field and train AI to complete the full task. Sarah Tavel best puts it as “work products”.
They improve productivity by 95% (rather than 5 or 10% in the above augmented experiences) and essentially replace a human service. These opportunities can be found in services where there already exists outsourced groups (eg. bookkeeping, call centre, customer support, medical billing, data scrapping, etc.)
These products can leverage value-based pricing and charge per work product, in direct comparison to the human capital it saves.
Examples: Even up, Intercom, Cresta
Before we explore some of the common pricing models, it’s important to keep in mind three criteria that guide a successful pricing strategy.
Simple: easy for customers to understand.
Whether you have a product-led self-serve motion or sales-led motion, pricing that’s easy to understand is an important factor in driving conversion on the pricing page and win rate for sales. The uproar over Unity’s confusing pricing update is a great counter example of this.
Predictable: customers can predict and budget the cost.
If you’re lucky, at some point in a sales conversation, the buyer will ask about pricing. And they will then immediately try to estimate the total cost against their budget. The friction to a “buy” decision is much greater if they can’t predict their total cost.
Scalable: aligns with customer’s growth, so you can grow together.
It’s always hard to increase prices, so having a pricing model that can naturally grow with customers is ideal. It completely aligns pricing to the product value received by the customer, as they depend on the product to grow their own business.
Common pricing options
Pricing AI products should consider how much differentiated value is delivered by the AI feature and how the customer perceives that value.
Per seat pricing: value is in accessing the AI feature regardless of usage
Per seat pricing with usage cap: value is in generating ideas to augment a creative process
Value-based pricing, per work unit: value is in delivering an end work product, replacing an outsourced human service
Usage-based pricing: value is the underlying infrastructure that powers the product
Let’s dig into each further.
Usage based pricing (token-based) is the most common for infrastructure/language models providers, as it’s based on the cost to run the models on GPUs. Obvious examples include OpenAI, Anthropic, Cohere, Databricks.
The argument against this model is that it’s hard to translate usage to value of work completed. It’s unavoidable as a model provider, and the GPU costs will go down overtime. But for companies that are using these models to deliver their own products, it’d be a bad idea to do cost-plus pricing. It’s difficult for customers to understand the discrete value provided on top of just using the models directly themselves, and it’s difficult to predict and budget cost.
As it gets cheaper to build LLMs, larger companies who have the data and resources could invest in building their own models instead of using a model provider. They’d then be able to enjoy minimal marginal cost and leverage value-based pricing.
OpenAI API: GPT-4 turbo input $0.01 / 1K tokens, output $0.03 / 1K tokens
Hume (analyze changes in people’s emotions based on voice intonation and facial expressions): charges per minute, annotation and word.
Databricks: $0.07/DBU for jobs. They have a handy pricing calculator to help customers estimate cost.
The classic SaaS per seat subscription model is simple to understand and easy to budget for. It’s the most straightforward model to charge for new AI features added to products that are already priced this way.
As we’re still in an experimental phase with AI, many companies are racing to add AI in their products to optimize for growth. They might not be as concerned about whether their token costs are covered. For products that are particularly expensive to run (such as images and videos), adding usage caps/credit bundles to a per seat model is also a good idea.
Flat rate per seat
ChatGPT: $20/month per user for ChatGPT
Copy.ai (AI OS to automate workflows):
Pro plan is $36/month for 5 seats, unlimited words, 500 workflow credits
Runway (Generates images and videos):
Standard plan is $12/user/month, 625 credits/month, can buy extra credits as needed
*Note how Copy.ai and Runway have usage caps.
Add-on per seat (AI features added to core product)
Notion AI: $8/user/month add-on to paid plan
Microsoft Copilot Pro (consumer): $20/user/month add-on to Microsoft 365
Microsoft Copilot (enterprise): $30/user/month add-on to Microsoft 365
Free AI add-on (bonus value to drive growth)
When AI gets sophisticated enough to replace a full end to end human task, there comes a unique opportunity to price in relation to the cost savings of having human complete that task. (Again,’s Sell work, not software is the best exploration of this concept).
The alternative would be outsourced human capital (BPO services). So it’s easy to calculate the cost of these outsourced tasks, which makes it easy to justify the ROI for the AI replacement.
Ian Clark’s How to price generative AI products also makes the argument for why the right way to price AI products is to charge for the value. You shouldn’t consider your cost at all when pricing your software product. It should be all about customer’s willingness to pay.
Intercom: AI chatbot Fin charges $0.99 for every resolved customer request. (Its core customer service product is priced per seat.)
Even up: charges per demand package (the work product produced by lawyers in injury law).
Any new pricing model innovations not covered here? How are you thinking about pricing models for B2B generative AI products? How is that decision impacting how you build the product?