It’s an AI-first future, it’s an agentic future—but it’s also a pricey future. Talk about some of the big announcements in the world of AI, and we’re easily talking thousands of dollars.
OpenAI’s latest full-blown ChatGPT Pro with o1 Pro mode costs $200 a month. Last week, the much-feared yet highly anticipated Devin was made available to the public for $500 a month. Recently, OpenAI CFO Sarah Friar suggested that the company may charge $2000 a month to replace humans.
But is it fair to charge a flat fee per month?
Elliot Greenwald, head of GTM operations at Sierra, wrote a blog post explaining the company’s use of outcome-based pricing. This is vastly different from traditional and consumption-based pricing models offered by tools.
An outcome-based pricing model only charges you when there is a tangible impact for your business—such as a business, upsell, or cross-sell, among other examples that Greenwald quoted. “If the conversation is unresolved, in most cases, there’s no charge,” he said.
A few months ago, Zendesk, a customer support and sales SaaS company, switched to an outcome-based pricing model for its AI agents. Meanwhile, Intercom’s Fin, an AI agent for customer service, charges $0.99 per resolution.
‘Charging for Behind-The-Scenes Processing Feels Unfair’
Recently, Replit devised a new pricing model. When it launched the new assistant inside Replit, the company announced ‘checkpoint-based pricing’. This allows users unlimited access to the agent, and once the monthly credits are exhausted, users transition to pay-as-you-go agent checkpoints.
Replit mentioned that it considered multiple pricing methods for the agent but ultimately landed on checkpoints. “Each checkpoint is a tangible piece of work the agent has completed—whether that’s generating a new component or implementing a feature,” read the announcement.
The company also said that charging users with tokens may be more precise, but it ‘doesn’t always reflect the actual value you get’.
Another approach they discarded was the price per message. The company said, “This can lead to users cramming everything into fewer, longer messages to save money.” This may not be ideal for either the AI or the users for tasks that require multiple back-and-forth communication.
Even a per-seat pricing methodology may not be ideal. Take customer support, for instance; when AI can handle most parts of it, companies will not need as many human agents. “Companies currently pay per support agent [per seat], but when AI can handle ticket resolution, the natural pricing metric becomes successful outcomes,” said a16z, a venture capital firm, in a blog post.
Moreover, when AI startups build on top of foundational models, which have significant variable costs that increase with usage, each API call and token processed adds to expenses, said the company. Concerns like this are leading more and more AI companies to use outcome-based pricing.
Several companies and business leaders are resonating with this shift. However, as outcome-based pricing moves towards more diverse use cases beyond customer support, it may not be as easy as it looks.
Aaron Levie, CEO of Box, the cloud service platform, said exploring these pricing models is one of ‘the most fun questions’ in AI right now. In a post on X, he outlined four approaches to the same.
One is a traditional labour model that prices agents like discounted human work, while another is an outcome-based approach that charges based on specific results. “The moment your service offers N types of value props or outcomes, you need N pricing models to go along with it,” added Levie.
Third, according to him, is a cost-plus model that closely tracks underlying AI expenses—‘potentially good for customers, but maybe not for shareholder returns’.
Lastly, one that offers unlimited agent use per seat. It could work where many seats are used, but otherwise, it may not provide enough value.
Hadi Partovi, CEO of Code.org and an early investor in Facebook, SpaceX, and Uber, said on X, “Another option is to charge a percentage of underlying transactions if the agent is making reservations, buying tickets, or shopping.”
That said, an outcome-based approach to AI agents has plenty of caveats.
Outcome is Ambiguous
AIM spoke to Karan Peri, a product advisor who has worked for companies such as Coinbase, Amazon, and Microsoft. He outlined several complexities involved in outcome-based pricing.
“How clearly can you understand the outcome, and can you figure out whether it has been achieved or not?” asked Peri. He further questioned that companies can provide excellent metrics about the outcomes of their AI agents, but does that help their customers make money?
Moreover, one has to differentiate between outcomes and needs. Companies may need to use AI agents to accomplish certain goals, but the outcome may not be clear.
Peri illustrates this with a simple example. “There is a need for Microsoft Excel—but is the outcome clear”?
There are more such factors, such as the regularity with which AI agents are used or the predictability of the need itself. Then, of course, the cost structures, the cost of the models and GPUs for the developers, and the upfront cost for the customers.
Peri said that only when all these parameters are perceived with clarity will outcome-based pricing prove beneficial.
Moreover, if AI agents are said to mimic the jobs of humans in the real world today, it is also appropriate to make sense of outcome-based pricing in our lives.
“If your ticket is booked, you pay the travel agent. If not, you don’t pay them. As opposed to a lawyer, or an accountant – it doesn’t matter what happens, you will need to pay for their time,” said Peri.
“So now, if you think about the real world, you will gravitate towards which models you fall under,” he added.
But, problems don’t end here. What do you attribute the outcome to? “You won’t know if you delivered an outcome for the customer; you won’t know if the customer themselves had to apply effort by other means to get the outcome.”
“Just because your tool was part of the workflow, it doesn’t mean that your tool was the one that delivered the outcome,” he added. There is also a substantial value added by the techniques and methodologies added by the customer while using the AI agent.
Moreover, one shouldn’t be confused with output and outcome. An AI agent may provide multiple outputs, and charge you per output. But what if only one output is crucial to the outcome?
“Which is why you see ChatGPT charging $20 a month, right? They cannot say X cents per message because they don’t get a value per message. It is an ongoing activity, and the outcome may not be clear all the time”, said Peri.
Concerns like these are why companies often rely on a SaaS-based model: a flat fee, with usage limits and usage-based pricing.
Peri suggests that paid AI add-ons to existing software or a bring-your-own API key approach could be more efficient for now. He also said that developers may prefer using API keys directly inside frameworks because it gives them a sense of ‘psychological predictability’, as they are used to utilising API keys from OpenAI, Grok, etc.
For now, it is fair to conclude that outcome-based pricing may only work well for customer service applications, which goes back to Sierra CX.
“But for nearly everything else, I think there is essentially zero chance this catches on or makes sense,” said Elaine Zelby, CEO of Tofu, a startup providing omnichannel marketing campaigns for B2B go-to-market teams.
“The only other example I can think of is lead gen companies where you pay per meeting or qualified lead. I guess that is also outcome-based pricing, and this scenario also makes sense,” she added.
The post Outcome-Based Pricing Won’t Work for AI appeared first on Analytics India Magazine.
