How AI apps make money - an interesting study of 40 companies

Every software vendor or organization creating intellectual property (IP) using AI should pay attention to the solution's monetization. In my previous articles, I emphasized that the pricing of a solution with Generative AI embedded needs to be considered differently. It

I wrote about the harsh reality that software vendors are facing a new reality in software pricing when using AI. This is because they require a new and different approach to pricing due to changed cost dynamics and solution characteristics . I also wrote about AI pricing strategies for SaaS companies that offer Copilots (like Microsoft) and what kind of approach they take in their pricing.

For a while, the trend has been to apply a consumption-based pricing model, and I gave some reasons why this could be the case in my blog . I also referred to the TSIA Fish model and its impact on software vendors, value-added resellers (VARs), and their business models.

Kyle Polar and Palle Broe conducted an exciting research initiative (published May 1st, 2024) on 40 companies and how they have priced their AI-based solution. The research focused on AI-native apps around pricing models, value metrics, packaging, free versions, and pricing transparency.

The primary criteria for the companies selected for the research were that they had received external funding and traction from public sources like the Forbes AI 50 List and Sequoia's Generative AI market map . According to Crunchbase , AI products are attracting a considerable share of the VC funding in 2024.

The research study asked the question: how are AI apps making money? Some products make people more productive (like ChatGPT or Microsoft 365 Copilot), and others create new categories and entirely new work products. The latter category allows building brand new pricing models, while the previous could be challenging as the customer is already paying for a solution based on specific premises, and adding AI pricing to the existing model might create resistance.

For example, what if Microsoft charged double the price of its Microsoft 365 Copilot ($60/user/month and not $30/user/month)? Would the customer still pay and receive the value of the solution? In my case, I would, but for large organizations where lots of the licenses might be unused, I'm not sure the $60/user/month would fly for users with hundreds or thousands of users with Office 365 E3 or E5 subscription that is already around that price point. Ibbka refers to an interesting price sensitivity meter that shows a sensitivity analysis of Microsoft 365 Copilot pricing. The picture of the sensitivity meter is as follows:

When discussing AI solutions, we need to understand what part of the "technology stack" we are discussing. A typical division of layers can be portrayed as follows :

The infrastructure layer: The Large Language Model (LLM) is typically hosted by one of the cloud providers such as Amazon AWS, Microsoft Azure, Google Cloud, OpenAI, Facebook, etc. The pricing model for these is typically usage-based, and many of these organizations charge on a per-token basis, which is closely aligned with the cost of computing. However, I am not a fan of the token method, especially with new reasoning LLMs such as OpenAI o1 LLM.

Reasoning LLMs use reasoning tokens that perform internal thinking processes. These tokens are not visible in the output but are used by the model to solve the problem. For a software vendor, the cost of the "internal thinking" could become a challenge from a costing perspective as the vendor does not have visibility on how those tokens are used.

The model layer represents the vendor producing LLMs, and there are many others besides OpenAPI, Anthropic, Mistral, etc.

The application layer is the layer that typically shows the end user where the application consumes the LLMs and portrays the results to the end user. An excellent example is Humantelligence , which has infused AI and LLM technology with its award-winning assessments (based on psychometric tests). I wrote about this in my blog and why Humantelligence is an interesting company to watch. It is an example of a changing industry in learning & development and how tools can help manage team dynamics.

The research study was based on five research questions, and the results were as follows :

1. Limited pricing innovation—seven in ten have a subscription model, and very few offer pure usage-based or pay-as-you-go pricing.
2. Most companies charge based on the number of users, which is?consistent with the notion of AI apps as “Copilots” (assist people) rather than digital “workers.”
3. Free versions are extremely popular for initial adoption—one-in-two have a free plan, another one-in-five offer a free trial.
4. There’s a “Good-Better-Best” paradigm in terms of packages/tiers .
5. Varying degrees of pricing transparency—two in three have public pricing.

The following pictures portray the different pricing models, value metrics, packaging, and whether the software vendor offers a free version and if the pricing is publicly displayed:

The research concludes that software companies have historically leaned towards a subscription and per-user model, and this is still the case with the first wave of AI apps. However, according to the study, the second wave of AI companies is applying pricing models that could unlock faster customer adoption while capturing more overall revenue. The study refers to Microsoft's testing of the new AI Copilot for Security and its innovative pay-as-you-go pricing model, where the customer pays $4 per hour of usage as part of a consumption model.

Software vendors will be challenged when accustomed to per-seat pricing, especially if moving towards value-based or outcome-based pricing models. Organizations tend to oversubscribe to seats, and according to the research, it is not uncommon to see monthly active user rates between 20% and 40 %. According to the article by Poyar et al., the assumption is that the traditional per-seat subscription model will increasingly come under pressure as AI products deliver work rather than augment personal productivity.

The research showed interesting results in respect to the five research questions:

?Finding #1: Limited pricing innovation

The vast majority (71%) have adopted a traditional SaaS subscription pricing model. 10 companies (26%) have adopted a hybrid pricing model, and only 3% have adopted a purely usage-based model. The reason, according to the study, is as follows:

1. Organizations want to keep the pricing simple and provide an easy way to adopt the solution.
2. Organizations feel it is hard to develop and measure usage-based pricing.
3. Organizations have a hard time quantifying the value of the solution.
4. Organizations do not want to limit adoption.
5. Organizations are not yet focused on profitability.

?Finding #2: Most companies are charging based on the number of users

The primary value metric driver is still charge-based and focused on users. This is the most typical metric, well-known in the SaaS world, and easy for the customer to understand. Microsoft decided to price Microsoft 365 Copilot based on seats and not on consumption. In the study, a dozen companies are using either per-user and usage-based components or a pure usage-based model with value metrics such as credits, characters, minutes of video, subtitles, etc.

It is clear that as AI use cases advance, where humans will play a lesser role, the per-user pricing model will also become less relevant as a value metric.

?Finding #3: Free versions are popular for initial adoption

According to the study, 47% of the organizations offered a version that is free forever, but which often has limited capability but allows the user to try the feature/product. 3% of the companies had a free version but limited in usage. 16% of the companies had a free time-bound trial, which had all the capabilities but was tied into 7 or 14 days.

?Finding #4: There is a "Good-Better-Best: paradigm in terms of packages/tiers

The Good-Better-Best pricing model has been typical for many SaaS companies, especially startups when figuring out what features/functionality are valuable for the user. The model allows the company to differentiate offerings depending on the customer and creates a clear upsell path. The number of tiers varies from two to five, including freemium and an enterprise version.

?Finding #5: Varying degrees of pricing transparency

According to the study, around two-thirds (65%) currently show pricing on their websites, while 35% do not. According to the study, transparent pricing tends to be the norm for individual—or prosumer-focused apps. Enterprise-focused apps are more likely to hide the pricing and require the potential buyer to call to inquire about it.

In enterprise cases, it is also more common to include implementation delivery, which makes the pricing fluctuate from company to company. Enterprise software vendors are also protecting themselves from competition as AI is still new, and every company is trying to figure out the best approach to pricing. Enterprise applications typically need to be customized, and software vendors want to keep their pricing flexible and not yet lock in the pricing.

The article also shows an interesting flowchart that portrays how a software vendor might react based on competition:

The flowchart clearly shows a software vendor's different routes in its pricing journey. Based on our experience after having facilitated tens of different software pricing and packaging workshops , pricing is an evolving thing. It must be considered at all times when advancing the product. Sometimes, new features will be added to the existing product (or tier), and sometimes, new functionality will be added as an add-on, as Microsoft did for its Microsoft 365 Copilot pricing.

It is clear that software vendors have much to learn about pricing these new-generation AI solutions, especially if the new reasoning LLMs are used. The uncertainty of the costing model and usage patterns will eventually educate software architects about what to expect from an application's consumption. I have always said that a bad architectural model will lead to an unsustainable costing model, so if I were the leader of a software company, I would keep software architects as part of the journey at all times to ensure that the expectations are met concerning what customers are willing to pay for the solution and what the underlying costing model can carry.

I would love to hear your views on this topic if you are a software vendor or an organization commercializing AI-based IP. How are you planning to price your solutions?

Yours,

Dr. Petri I. Salonen

PS. If you would like to get my business model in the AI Era newsletters to your inbox on a weekly or bi-weekly basis, you can subscribe to them here on LinkedIn https://www.dhirubhai.net/newsletters/business-models-in-the-ai-era-7165724425013673985/