Generative Pricing
Steven Forth is a Managing Partner at Ibbaka. See his Skill Profile on Ibbaka Talio.
TL:DR Generative AI applications require a new approach to pricing due to changed cost dynamics and emergent properties. Key aspects include:
Responsive Pricing: Prices adapt quickly to changing configurations.
Explainable Pricing: Pricing is transparent and justified.
Value-Based Pricing: Prices reflect the value created for customers.
Composable Applications: Configurations change in real time to meet user needs.
Conversational Interfaces: Users interact with applications through structured conversations.
Integrative Solutions: Applications draw on multiple sources of information.
Synthetic Data: Applications generate data to support insights and design spaces.
This new approach will need to balance value creation for customers with cost minimization for sellers.
(Generated by Perplexity using Sonar Large - this gave a better result than using GPT4.o)
“Pricing an AI product will be a defining question in software for the next few years. AI products offer productivity gains. But greater productivity may reduce the demand for seats over time, ultimately decreasing the size of software markets.” Tomasz Tunguz, July 11, 2024
There is an emerging view that generative AI and the applications it spawns will require a new approach to pricing.
Ibbaka has published four key blog posts to help you get oriented and begin to execute with generative pricing.
Generative Pricing (where we introduce the approach)
Pricing approaches for generative AI applications (where the alternatives are considered)
The path to generative pricing (where we introduced the five step approach to generative pricing)
Steps on the path to generative pricing (where we expand on the five steps)
Current approaches - whether implemented in pricing software like PROS, Zilliant, Vendavo, and Pricefx, or in CPQ (Configure Price Quote like Salesforce CPQ), or in price books in Excel with multiple lookup tables - all fail to address the emerging properties of generative AI applications.
What are these emergent properties? They come from two directions.
Changed cost dynamics
Emergent properties of the applications themselves
Generative AI cost dynamics
Thinking about the changed cost dynamics, there has been a lot of angst here as the SaaS business and investment model, especially the investment model, has become dependent on high operating margins and rapid growth. This is seen in common benchmarks like the ‘rule of forty’ (the sum of growth and profit margin must be larger than 40).
Generative AI cost dynamics mean that it is easier and faster to develop new applications and then to configure or compose these applications for a specific buyer or use case. That will lead to a huge change in development and reinforce the tendency to agile development and fail forward (“Fail means to not achieve an objective. Fail forward means to keep on trying, despite making mistakes, until you reach the objective.”)
There will be a lot of exploration over the next few years as we discover the many ways that generative AI applications can create value. The cost of that exploration will go way down. The cost of launching and supporting applications will also go down. Generative AI will improve discovery (reduce search costs for the buyer) and will be used to evaluate new applications and proposals (one emerging trend we will have to look at in the future is the use of generative AI to evaluate proposals).
Many costs will go down, but the fundamental cost of operating a service is likely to go up. Processing a prompt requires a lot of compute, even at the best of times, and to make matters worse that compute power, which depends on specialized chips like Nvidia’s Blackwell architecture, is going to be in short supply for the next few years. SaaS grew up in a time when computing was cheap and always getting cheaper. That has changed. Some new rules are needed.
Only incur operating costs if
those costs create value for customers
the value created can be monetized and captured in price
Generative AI pricing models will need to factor in operating costs. This will call into question some accepted practices. Freemium models (where there is a large population of free users and a smaller group of paying users who need to cover the costs) and free trials will need to be rethought and carefully calibrated. Freemium and free trial are not likely to go away completely, they are deeply engrained in the SaaS culture, but they will be subject to more scrutiny and much better fenced.
Per-user pricing will become less common as it does not generally correlate with value or cost. One get a glimpse of how this is evolving, see: What does Box pricing tell us about AI pricing trends?
Emergent properties of generative AI applications
The first generation of generative AI apps were largely examples of sustaining innovation. They provided incremental improvements to things we already do, and they were priced accordingly. In How AI Apps Make Money, Kyle Poyar and Palle Broe noted “Lack of pricing innovation presents an opportunity for the second wave of AI apps.”
For that second wave of apps to really make a difference they will need to take advantage of the emergent properties of generative AI applications. What are these?
In an interview with Ibbaka, Michael Mansard from Zuora identified three properties that he looks for. Generative AI apps are actionable, adaptable, and support process efficiency.
By actionable, he means that when “you go into a GenAI application and it provides you with direct actionable insights and outputs.”
Adaptability is where GenAI shines. “It customizes and adapts to what we just said in a user-friendly way, in a way that you do not expect.”
Process efficiency is driven by the “extreme speed, actionable and adaptable outputs with limited and sometimes no human effort.”
Each of these points to how generative AI apps will create value.
At Theory Ventures, the focus is on composable applications. See Composable software platforms in practice. Composable applications are built on platforms, enabled by modern infrastructure and AI-based automation, that allow companies to configure applications, workflows, and data to their specific business needs. This happens in near real-time. An example we discussed in Pricing AI: What role could AI play in pricing? is Totogi, a BSS platform (used to enable virtual mobile networks). Totogi promises a fully customized BSS that can be figured in minutes using AI.
At Ibbaka, we see generative AI platforms as being Composable, Conversational, Integrative, and Synthetic.
Composable
This could be the most important of these properties. If applications become composable in real time we will be able to adapt them to each users current needs and evolve the configurations quickly over time. This will completely change B2B software.
Applications are configured real-time
Buyers and users have direct control over the configuration
Different users at the same company can have different configurations
Configurations can change depending on the conversation
Conversational
Conversational interfaces will help make composable software possible. The configuration will be created through a structured conversation with the user. And as Michael Mansard pointed out above, the conversational experience makes use of the application more actionable.
Conversations with the application will be one of the main modes of interaction
Conversations will be an input into configuration and data generation
Conversations will be multithreaded and may include multiple applications
Integrative
Adaptive solutions need to be able to draw on many different sources of information. Depending on data locked in one silo cripples applications today and the situation will only get worse with the current data explosion. At the same time, most advanced applications will use more than one language model and will integrate other approaches to artificial intelligence. Hugging Face currently has more than 775,000 models.
LangChain and similar approaches will build applications from multiple models
Data will come from multiple sources and be integrated real-time using RAG
Integration models will be open and responsive to real-time configuration
LangChain is an open-source framework designed to facilitate the development of applications powered by large language models (LLMs).
RAG or Retrieval-Augmented Generation is an AI technique designed to enhance the accuracy and reliability of generative AI models by integrating them with external knowledge sources.
Synthetic
Synthetic data is data generated by applications, especially AI applications, rather than collected from the real world. The use of synthetic data (and synthetic users) is one of the open questions about AI. Increasingly models will be trained on data created by models. Some think this will lead to model collapse. Others have shown that synthetic data can be used to meaningfully enrich data sets and support better reasoning. For one inspirational use of synthetic data see Scaling neural machine translation to 200 languages. Many languages do not have enough text available to build a large language model. This research showed ways to leverage synthetic data to build better models. Sometimes one can hoist oneself up by one’s own bootstraps.
Suffering from sparse data? You need to explore synthetic data as a solution while being aware of how this could lead you astray.
Applications will generate data that they use to bootstrap insights
Synthetic data will be used to generate design spaces that the applications can explore
These emergent properties are what drive the need for a new approach to pricing.
Generative pricing for generative AI applications
Generative AI applications need generative pricing.
What is generative pricing, and how does it differ from conventional approaches to pricing like cost plus, willingness to pay (WTP), dynamic pricing of value based pricing?
Generative pricing will need to have three key attributes, it will need to be responsive, explainable and value based.
Responsive Pricing
If one can compose (or configure) an application in seconds your pricing needs to be just as quick. And if the configurations are adaptive, changing as needs change, the pricing needs to respond to that as well.
This is fundamentally different from dynamic pricing. Dynamic pricing is based on various external factors, including current market demand, the season, supply changes, and price bounding. With dynamic pricing, product prices continuously adjust – sometimes in minutes – in response to real-time supply and demand. It is generally calculated by pricing engines such as those provided by PROS, Vendavo, Zillant and Pricefx.
Responsive pricing is responsive to the configuration, which is driven by the customer’s constantly evolving needs. In responsive pricing, the product itself changes and the price along with it.
Systems supporting generative pricing will need to be able to generate quotes at the same cadence that configurations change.
Explainable Pricing
One of the challenges of pricing optimization engines is that they are seen as black boxes. They take in all sorts of data, run some fancy algorithms, and provide sales with an estimate of Willingness to Pay (WTP). When challenged on price a salesperson cannot very well say ‘My AI told me you should be willing to pay this much.’
AI also has this problem. There is a perception that AIs are black boxes. This is not really true, there has been a lot of progress in explainable AI over the past few years (Google for example has some good approaches). Generative AI apps are getting better and better at explaining and justifying the content they create.
Pricing, if it going to be responsive, will also need to be explainable.
We expect the chatbots that are used so much now in pre-sales and customer success to be able to explain pricing. They will be asked questions like …
“How did you come up with that price?”
“How can I get a lower price?”
“Does this price reflect the value I am getting?”
“Who is getting a lower price?”
Your chatbots had better be primed to answer questions on your pricing.
Value Based Pricing
Value based pricing has been recognized as the most compelling approach to pricing for some thirty years (the first edition The Strategy and Tactics of Pricing by Tom Nagle and Reed Holden was published back in 1994). Despite this, most companies have struggled to implement this approach and there is still a lot of confusion about what value based pricing actually is (some people are so misguided as to confuse value and willingness to pay). Generative pricing is going to require value based pricing.
The driver for this will be the composable nature of generative AI applications. If it is easy to compose applications and support dynamic configuration how should you configure? The best configuration at any point of time will solve for value optimization for the buyer and cost minimization for the seller. Price will be a result of these two optimizations.
Value will continue to be represented through a value model, but that model will become more dynamic. During the buying process there will be a conversation with an AI that uncovers the value, quantifies the value model, and then explores how different configurations deliver value.
Value Model: A value model is a system of equations built from value drivers that estimate how much value a solution will provide to a customer or set of customers (market segment).
Once the solution is in use the value model can be kept current by integrating data from other systems. This will help to guide adaptive configuration. The configuration that optimizes value will change over time. At the same time, synthetic data can be used to enrich the model (today most value models struggle with sparse data) and to explore other parts of the design space and test counterfactuals (working with counterfactuals being central to causal reasoning systems).
The path to generative pricing
Begin with a value model. If you do not have one build one (this is often the job of value engineers).
Develop a way to assess the value model and improve it, generative AI can be used for this purpose.
Map the value drivers in the value model to the possible configurations, generative AI can also support this task.
Build a chatbot that can be used to
Have a conversation about value that can be used to build and quantify the value model
Propose pricing
Answer questions about pricing (support explainable pricing)
Use generative AI with techniques like feature flags to support more dynamic configuration (this is a step on a longer path, but as many applications already use feature flags it is a step towards composability).
Have lots of conversations with buyers, users, potential buyers about value, how you create value, how you can better create value, how you can help them to create value, and about how their own businesses are changing.
For some simple steps you can take to get started with generative pricing see The path to generative pricing.
LeadLander