AI Pricing: Has Microsoft Copilot set a reference price for other Generative AI Applications?
Steven Forth is a Managing Partner at Ibbaka. See his Skill Profile on Ibbaka Talio.
The PeakSpan/Ibbaka survey results on Net Revenue Retention | State of the Industry - High Performers VS The Rest are now available to download.
Over the past few weeks, Ibbaka has had a number of conversations with companies bringing generative AI solutions to market. In a surprising number of cases, the company is mentioning US$30 per user per month as a reference price for generative AI functionality. They say this comes from customer comments and in some cases market research. This number most likely comes from the price that Microsoft set for MS365 Copilot in July.
For some research on Microsoft’s pricing of Copilot see Pricing AI assistants for productivity suites: survey results.
Some definitions of common terms in generative AI.
Generative AI: Any type of artificial intelligence (AI) that can be used to create new text, images, video, audio, code, or synthetic data. In general, these AIs leverage large language models (LLMs) built using the transformer architecture introduced by Google in 2017.
Large Language Model: Models that given a prompt as an input generate text, images, music, video, code, or synthetic data. These models generally have hundreds of billions of parameters (weighted connections between tokens) to a trillion or more parameters. GPT-3 and GPT-4 from OpenAI, LLaMA from Meta, Cohere, and PaLM2 from Google are common examples.
Transformer Architecture: A neural network architecture composed of encoders and/or decoders. The design aims to solve sequence-to-sequence tasks while handling long-range dependencies. It was first proposed in the paper “Attention Is All You Need” and is now a state-of-the-art technique in the field of Natural Language Processing.
What does it mean to say that MS 365 Copilot pricing is acting as an anchor for generative AI pricing?
When people and businesses make buying decisions they expect the price to be in an acceptable range. This can be a wide range, generally 3X to 10X orders of magnitude. Where a solution falls on the range suggests its positioning and expected value. In pricing, this is often presented in Value Maps, where the price is on the Y-axis and some measure of value on the X-Axis.
At present, the two most common reference prices for generative AI are Open.ai’s price of US$20 per user per month and Microsoft’s price of US$30 per user per month.
Given this anchoring, it will be difficult for anyone to get more than $100 per user per month for generative AI functionality. To do so, one will have to convince people that one is providing a great deal more value than MS 365 Copilot.
The other anchoring effect is that generative AI pricing should be ‘per user’. This was a choice Microsoft made for Copilot to align it with MS 365 pricing. They needed to make it easy to build bundles, so defaulted to a pricing metric they were already using. There are other pricing metrics that could be used.
In value-based pricing, one looks for a pricing metric that tracks the value of the solution. Value drivers for generative AI are not yet well-defined or validated. They tend to come in the following buckets.
Time saving (for an individual)
Labor saving (for an organization)
Option generating (generate and explore more ideas)
Risk reduction (identify and plan for more risks)
The ‘time-saving’ and ‘option generating’ value drivers map to ‘per user’ pricing, though even here they are not compelling. With more time and thought Microsoft could have come up with other pricing metrics.
ChatGPT’s suggestions for pricing metrics for ChatGPT
Microsoft could have asked ChatGPT for suggestions on pricing metrics (perhaps they did). Here is our conversation with ChatGPT.
Prompt: “A pricing metric is the unit of consumption for which a buyer pays. Per user, per transaction, per search, per query, and number of tokens are all examples of pricing metrics. What pricing metrics could be used for ChatGPT?
ChatGPT
“For a service like ChatGPT, which provides AI-powered conversational capabilities, various pricing metrics can be employed to charge buyers based on their usage. Here are some potential pricing metrics that could be used for ChatGPT:
Per User: Charging is based on the number of unique users or accounts that interact with the ChatGPT service. This could be a common pricing metric for platforms that integrate ChatGPT into their applications.
Per Message/Interaction: Charging is based on the number of messages or interactions between the user and the ChatGPT. Each back-and-forth exchange in a conversation could be counted as an interaction.
Per Conversation Session: Charging based on the initiation of a conversation session with ChatGPT. A session could encompass multiple interactions within a single conversation context.
Per Minute of Usage: Charging by the duration of time a user spends interacting with ChatGPT. This metric could be suitable for scenarios where users have longer conversations.
Per Query/Question: Charging is based on the number of queries or questions submitted to ChatGPT, regardless of the length of the conversation.
Per Token: Charging based on the number of tokens processed by ChatGPT. Tokens are the units of text, and longer inputs or outputs require more tokens.
Tiered Plans: Offering different pricing tiers with varying levels of usage or features. For example, a basic plan might include a certain number of interactions or tokens per month, while a premium plan might offer higher limits or additional features.
Subscription Plans: Providing users with the option to subscribe on a monthly or annual basis, giving them access to a certain level of usage without worrying about individual transaction costs.
Pay-as-You-Go: Charging users based on the actual usage, calculated at the end of a billing cycle. This could be a flexible option for users with unpredictable usage patterns.
Custom Pricing: Offering customized pricing models for large enterprises or businesses with unique requirements. This could involve negotiations based on factors like expected usage volume, support, and integration needs.
It's important for providers of the ChatGPT service to carefully consider which pricing metric(s) align with their business goals and the expectations of their users. Different pricing metrics might be suitable for different types of users or use cases, and finding the right balance can help ensure a fair and sustainable pricing structure.”
Nothing revelatory here, but it reinforces the point that many different pricing metrics could have been used.
Pricing metrics used in AI applications
PeakSpan and Ibbaka will be publishing the results of a survey on Net Revenue Retention in early September. One of the questions we asked was about pricing metrics. We were looking for connections between the number and type of pricing metrics and Net Revenue Retention.
As a result, we have insight into the pricing metrics being used by different SaaS verticals, including AI. The results for AI are summarized in the figure below, excerpted from the survey. There were 341 survey responses, 25 of which were from AI companies.
Per token pricing is how Open.ai charges for API access to its models and has become quite common in generative AI.
AI vendors are using more different pricing metrics than any other SaaS vertical. In addition to those mentioned above, AI vendors also price by model, by server, by consumption of server resources and they often customize the pricing model for each client’s unique needs.
This in part reflects the early stage of the AI market where solutions and business models are evolving quickly and value has not yet been modeled and validated.
Contact info@ibbaka.com if you would like to be notified when the report on the survey is released.
How should generative content applications position themselves relative to Microsoft Copilot and its $30 per user per month pricing?
Companies developing generative AI solutions have need to make some key choices.
Adopt user-based pricing or find a different pricing metric
If user-based pricing, reference MS Copilot or try to distance one’s solution and price without this as an anchor price
If anchoring on MS Copilot, position at a discount, on par, or at a premium
In most cases, one should not use per-user pricing. Only use per-user pricing if (i) each user requires their own unique configuration and history (ii) data privacy needs to be managed at the individual user level or (iii) the solution is bundled as part of a platform that is priced per user (this was the main reason Microsoft went with per-user pricing for Copilot).
Rather than per-user pricing, create a value model for the application and see if one more of the value drivers can be used as a pricing metric.
Definition of a Value Model: A system of equations used to estimate the economic value a solution provides to a user or buyer over time. Each equation formalizes a value driver. A value driver is the impact a solution has on a customer’s business by …
Increasing revenue
Reducing cost
Reducing operating capital
Reducing or deferring capital investment
Reducing risk
Increasing optionality
If you do go with per-user pricing, you need to decide if it is in your interest to have Microsoft Copilot as an anchor.
The answer will be yes if your customers are already MS365 users, your solution integrates with MS365, and your solution will be thought of as an office productivity application or an extension of such an application.
If your application is used for design, software development, business planning and forecasting (beyond what is possible in spreadsheets), scientific research or any of the many other domains that go beyond office productivity you probably do not want to anchor yourself to Copilot. Instead, anchor yourself to the pricing metrics and pricing levels already used in your category. If you are instead creating a new category it is even more important not to have pricing anchored in an existing category.
See Pricing and value for category creation.
Can you price higher than an anchor price? Absolutely yes. The way to do this is to show that you create orders of magnitude more value than the anchor. This is at the heart of value-based pricing. An application that has done some combination of the following is probably creating a lot more value than Copilot.
Enhanced the Large Language Model
Tuned the hyperparameters of the model
Constructed special prompts
Built a prompt orchestration engine for these prompts
Set up a GAN (Generative Adversarial Network) so that two models are interacting to compete with and improve each other
Designed in learning loops so that the application gets more intelligent with use
Integrated inputs from or outputs to other applications
Invested in a coherent user experience that supports understanding and effective use
On the other hand, if the solution is seen as an add-on to Copilot or to have less functionality and less ability to create value, then it will need to be priced at a discount of $30 per user per month.
That is not necessarily a bad thing. Many SaaS companies would love to get even an extra dollar per user per month. For an application priced at $10 per user per month that would be a 10% price increase.
Price increases are not the only way to get value from new functionality. Before fixating on pricing the new functionality, ask how much it would be worth to
Reduce churn
Increase use (and if you have some form of usage-based pricing this will lead to a price increase)
Drive up-sell (if you have a tiered packaging architecture like Good Better Best)
In some cases having a specific price is not the best way to monetize innovation.
Google has not yet clarified if and how it will price its Copilot alternative DuetAI.
Google has an opportunity to re-frame this market and create a new anchor. It will be interesting to see what they make of this.
