How to price AI

Steven Forth is a Managing Partner at Ibbaka. See his Skill Profile on Ibbaka Talent.

“There is grandeur in this view of life, with its several powers, having been originally breathed into a few forms or into one; and that, whilst this planet has gone cycling on according to the fixed law of gravity, from so simple a beginning endless forms most beautiful and most wonderful have been, and are being, evolved.”

Charles Darwin, The Origin of Species, 1859


AIs will transform all parts of the software industry over the next few years. If ‘software is eating the world’ (Marc Andreessen) then AI is eating software. We need to develop a framework for pricing AI and understand how the value metrics and pricing metrics for AI will work. This is an exploration of this topic.

To get started, I thought I would start by asking Open.ai’s ChatGPT how to price AI.

ChatGPT’s thoughts on pricing AI and pricing itself

“How should AI applications be priced?”

“The pricing of AI applications can vary depending on a number of factors, including the type and complexity of the application, the value it provides to users, and the market demand for the product. In general, it is important for companies to carefully consider the potential costs and benefits of their AI products and to price them in a way that is fair and reasonable for both the company and its customers. It may also be helpful to consider the pricing strategies of other similar products on the market and to conduct market research to understand the demand for the product and the price point that potential customers are willing to pay.”

I then asked it how it would price itself.

“How should ChatGPT be priced?”

“As with any AI product, the pricing of a ChatAI application will depend on a number of factors, including the type and complexity of the application, the value it provides to users, and the market demand for the product. In general, companies should consider the potential costs of developing and maintaining a ChatAI application, as well as the value that the application provides to users. For example, a ChatAI application that provides valuable customer support or personalized recommendations to users may be able to command a higher price than a more basic chatbot. It may also be helpful to conduct market research to understand the demand for ChatAI applications and the price point that potential customers are willing to pay.”

ChatGPT brought up ‘willingness to pay’ so I asked

“How does one figure out customer willingness to pay?”

“One way to figure out a customer's willingness to pay for a product or service is to conduct market research. This can involve surveying potential customers to ask them about their interest in the product, their willingness to pay for it, and their budget for similar products or services. Market research can also involve studying the pricing of similar products or services on the market to understand the competitive landscape and how customers may react to different price points. Additionally, companies can use data analytics to analyze customer behavior and purchase patterns to gain insight into their willingness to pay for a product or service.

Not happy with an answer? ChatAI invites you to try again.

“There are several ways that a company can determine a customer's willingness to pay for a product or service. One common method is to conduct market research, such as surveys or focus groups, to gather data on customers' perceptions of the product and their willingness to pay a certain price for it. Another approach is to offer the product at different price points and observe the resulting sales to see which price point is the most successful. Additionally, companies can look at the pricing of similar products on the market and use that information to inform their own pricing decisions.”

And I had to ask

“CanGPT be used to design pricing?”

“Yes, AI can be used in the design thinking process to help designers generate and evaluate design ideas more efficiently. For example, AI algorithms can be used to analyze large datasets of user behavior and preferences, providing designers with valuable insights into the needs and wants of their users. AI can also be used to create simulations or prototypes of design ideas, allowing designers to test and refine their designs more quickly and efficiently. Additionally, AI can be used to automate certain repetitive or time-consuming tasks, such as data collection and analysis, freeing up designers to focus on more creative and strategic aspects of the design process. Overall, while AI is not a replacement for human creativity and empathy, it can be a valuable tool for designers as they seek to create innovative and user-friendly products and services.”

ChatGPT is a lot of fun to explore. I am using it in the design thinking community to explore the current understanding of design thinking and where it is going (at Ibbaka we apply design thinking to the design of pricing models). I am also using it with my granddaughter as we co-author our novel The Enchanted Library to generate the dialog for a couple of characters. But let’s go on and look at some frameworks for pricing artificial intelligence.

How can we organize AI applications?

Before we can talk about pricing AI applications we need a way to organize the different types of applications. This is not an easy task as new applications appear everyday and AI is being applied much more widely, and more quickly, than most of us expected.

Earlier this year I proposed the parallel hypothesis,

If one can use AI to describe or categorize a piece of content, you can use AI to generate similar content.

This hypothesis is proving to be true, with new AI applications able to generate more and more types of content, from diagnosis and recommendations, to code and designs, to reports, dialogs, voices, music and many forms of art.

There are 5 main types of AI applications.

  1. Classification

  2. Pattern Recognition

  3. Recommendation

  4. Prediction

  5. Generation

The value metrics (and therefore, the pricing metrics) will be different for each of these application areas.

The AI Ecology

AI applications are created using a complex set of technologies and resources. One way to organize these is shown below. Different parts of this ecology deliver value, and are therefore priced, in different ways.

This way of organizing the AI ecosystem is designed to help understand how to price. There are many other ways to organize the ecosystem, as a search for ‘AI ecology’ or ‘AI marketmap’ will reveal.

Infrastructure: AI development, management and application runs on a complex of custom chips (from foundries like Nvidia), servers, databases, data pipes and all the software required for management of a complex IT infrastructure. Some large companies develop and manage their own infrastructure, but most rely on cloud vendors for this. Note that in the above sketch, ‘training’ means training a model and developing a model from a set of data.

Data: AIs, especially machine learning AIs, cannot be built without data, lots of data. Data that can be used to build AIs is becoming more and more valuable. Organizations that hold such data are looking for ways to monetize it, whether directly or by providing it to AI companies.

Tools: Programming languages like Python, Lisp and Julia; frameworks like Tensorflow or Pytorch; and APIs like Keras. This is just the tip of the iceberg. GitHub has thousands of tools, and the big AI shops have their own custom versions.

Workflows: With all that data being run through so many different tools, there is a need for orchestration systems that control how data flows through the many different tools. Workflow systems are also used to manage the models generated.

Models: Models are developed from data using tools. An AI is basically a model, static or dynamic, that takes some form of input and generates an output.

Outputs: At the end of the day, we are not looking for models, but the outputs of the model. For many AI applications, it is the outputs that will be priced.

Outcomes: AI outputs are meant to have an impact on the world. The type of impact will depend on the application. An AI that is used to identify and diagnose tumors on an image will impact cancer survival. An image generated by an AI might be packaged as an NFT (Non Fungible Token) and the outcome could be the price at which it is sold. Pricing outcomes is the future of pricing and will be enabled by ongoing advances.

Pricing Components of the AI Ecology

How we price AIs depends on the role played in the ecology.

There is nothing magic here. Basic pricing methods apply with the goal being to understand the value metric and to find pricing metrics that track the value metric.

Value Driver: An equation quantifying one aspect of how a solution delivers value to a specific customer or narrowly defined customer segment.

Value Metric: The unit of consumption by which a user gets value.

Pricing Metric: The unit of consumption for which a buyer pays.
Here is a brief guide to the different approaches to pricing seen in different parts of the AI ecology.

For the next couple of years, the most common use of AIs will be to enhance existing applications. This is an example of sustaining innovation. There will also be cases of disruptive innovation and even cases of category creation. Pricing of innovation has been endlessly discussed. Ibbaka’s approach was shared in Pricing innovation and value drivers.

How can we price the different parts of the AI ecology?

Infrastructure: This is the most developed part of the ecology. Companies like Amazon AWS, Google, IBM and Microsoft Azure are pricing their AI infrastructure services the way they are pricing their other infrastructure. One pays for servers, databases, volume of data stored, throughput and latency. Looking into the details of the pricing plans can give a lot of insight into the strategy of each company. This is something we will be analyzing in 2023.

Data: Generally data will be priced based on the number of records, the quality of the records and the relevance of the records. Quality and relevance will differ by application, so it is probable that the same data will be sold at different prices for different applications. This will be easier to do if approaches like Open Algorithms (OPAL) from the MIT Connection Lab are applied. In this model the data stays put and the algorithm comes to the data. If this becomes a common approach, the complexity and value of the algorithm may be used to set the value of the data the algorithm is being applied to.

Tools: Today most of the best tools are open source. That is likely to change though and some companies will want to find ways to monetize their tools directly. The most likely pricing metric here will be ‘per user.’ More sophisticated companies will price on the outputs of the tools. There are likely to be free versions of each tool, for you to try, and then Good Better Best packaging with a combination of the number of outputs and functionality available used as pricing metrics and fences between packages.

Workflows: Applications for managing AI workflows are not all that different from those used to manage any data workflow. This is an application of a mature technology to a new space. Pricing metrics will likely be the same as those for other data management workflows: process steps, the number of times the process is invoked, throughput, the number of applications connected and in advanced applications network measures based on nodes and edges.

Models: This is where it starts to get interesting. All of the above, from infrastructure to data and tools to workflows, are meant to generate models. Many AI companies charge for use of their model. This is what Open.ai does to price the GPT model.

In this case, Open.ai prices inputs into the model (tokens) and then fences the models using speed and power. This is an example of usage-based pricing.

Outputs: Models have outputs. Open.ai takes this approach to its image generation model Dall-E.

Here one pays a higher amount per image based on image resolution.

For many applications, paying for the output of a model will make the most sense. One could pay per item classified, for pattern found, recommendation or prediction made.

Outcomes: The future of pricing is to price outcomes. There have been whole books written about this, see The Ends Game: How Smart Companies Stop Selling Products and Start Delivering Value by Marco Bertini and Oded Koenigsberg. In the long run, this is how the outputs of AIs will be priced, and the AIs will help to address the current objections to outcome-based pricing.

  • There are too many causes contributing to any significant outcome

  • Outcomes are too unpredictable to price

AIs can be trained to apply approaches like Judea Pearl’s causal reasoning to estimate the causal contribution of different factors (this is already routinely done in Health Economics and Outcomes Research). And prediction is one of the core use cases of modern AIs, and can be used to predict the outcomes of any intervention or solution (in this way AIs will eventually replace value models and value based pricing).

Research into how to price AI

At Ibbaka, we plan to develop this framework for pricing AI in 3 ways:

  1. Work with clients to apply it to their pricing challenges.
    Ibbaka is focussed on developing and implementing pricing models for B2B SaaS. As much of the innovation over the next few years will come from applications of artificial intelligence, we expect to do a lot of work in this area. Key SaaS verticals like Cyber Security, Marketing Automation and Financial Management are all actively incorporating AI into their solutions.

  2. Research and compare AI pricing as it appears in the market.
    For example, a great deal can be learned by comparing how Amazon Web Services, Google Vertex AI, MS Azure or IBM Watson are priced. We will also be watching to see how Open.ai, Stability AI and other innovators approach pricing.

  3. Develop general value models for AI-based applications and services.
    At this point in time, the best practice is to base pricing on the variables used in value models. By developing some general value models for AI we will be able to explore how value is being created in the market.

To follow this work, please subscribe to the Ibbaka Value Pricing Blog.

 
Previous
Previous

Pricing and Package Design - Two packaging patterns

Next
Next

Core Concepts: Emotional Value Driver