What if one day OpenAI tells businesses: "If the AI didn't deliver, you don't have to pay"?
This has already begun. According to The Information, citing a person with direct knowledge, over the past few months, OpenAI has quietly offered select enterprise customers a new payment model: businesses only pay when the AI successfully completes tasks such as customer service. However, an OpenAI spokesperson declined to comment.
It may seem like just a change in billing method, but it actually impacts the entire business model that has shaped the software industry over the past two decades.
In the past, calculating software costs was straightforward: one employee, one account, a fixed monthly fee—companies simply subscribed based on the number of users. But in the era of generative AI, this model is becoming increasingly outdated. An agent might answer one question in a minute, or work continuously for hours to complete an entire workflow for multiple people. Even more complicated is the fact that increasing model usage doesn’t necessarily translate to higher value created for the business.
So, more and more companies are asking a more direct question: if AI claims it can get the work done, why not pay only after it actually completes the task?
AI is selling more and more,
The software company doesn't know how to charge anymore.
The software industry has been exploring this issue over the past two years.
The most straightforward approach is to charge based on usage, such as tokens, number of API calls, or tasks completed. This method is at least closer to AI’s cost structure than traditional subscription models based on accounts, but another issue quickly arises: higher usage does not necessarily mean higher value.
For example, the number of calls an agent makes—say, 1,000—is clearly not the same as the number of sales it ultimately generates. Similarly, an engineering team calling hundreds of millions of tokens per day does not indicate how much revenue those tokens ultimately generated for the company.
Therefore, usage-based billing makes it easy for businesses to feel like their money is being spent without clear justification. What they truly care about is return on investment.
This is why "pay-for-results" is rapidly entering the pricing models of AI companies. Customer management company Sierra uses a similar model: customers only pay when the AI successfully completes a task without human intervention. Fin, which Salesforce is acquiring for $3.6 billion, also charges based on the number of problems the AI successfully resolves.
The programming agent company Cognition has gone further, promising enterprise clients that if the final engineering deliverables do not match the value of what they paid, the company will offer up to $10 million in credits. Established software companies like Adobe, HubSpot, and Zendesk are also beginning to shift toward a model where they only charge after AI has truly completed the work.
OpenAI's entry into this space means that this pricing model is no longer just an experiment by AI startups.
Similar approaches have also emerged domestically. For example, Lingxi Technologies employs a model closely aligned with “outcome delivery” in sales scenarios such as insurance and automotive. Its AI does not merely serve as a copilot for sales staff but actively participates in customer communication, needs assessment, follow-up, and closing deals. The key metric for evaluating the system’s value is straightforward: whether it generates real orders, premiums, and sales growth.
This aligns with the same shift that companies like Sierra in Silicon Valley are attempting: AI companies are beginning to proactively take on some of the risks that were traditionally borne by customers.
Previously, software vendors sold you a tool, and whether anyone used it or whether it increased revenue was up to the customer. Now, some companies are willing to directly tie their own revenue to their customers’ business outcomes. This sounds more appealing, but it’s much harder than selling software.
Salesforce has also started changing its pricing model.
In large corporations, Salesforce is one of the most typical examples of this change.
The company is offering increasingly flexible payment options for Agentforce. In addition to purchasing based on AI usage, enterprises can now negotiate custom contracts with Salesforce, tying costs directly to business outcomes—such as the number of additional orders closed by sales teams thanks to Agentforce, or the cost savings achieved through automated customer service tasks.
CEO Marc Benioff said during a recent earnings call that he has recently felt strongly that "customers want to buy differently and be priced differently." His thinking has moved beyond simply "charging once per task."
Benioff gave a straightforward example. The original outcome-based pricing might simply have been: “We made this many calls, so you pay us $2.” But what Salesforce truly aims for is: “We helped you increase your revenue by this much, so you pay us $2 because we helped you earn $20—or even $40.”
If this can be achieved, software companies can naturally charge higher prices. Benioff bluntly stated that, under this model, software vendors could command "very high prices" for their products.
Salesforce stated that its flexible AI pricing model has helped the company close several "very large deals." Last week, it also disclosed that sales of Agentforce integrated with a data management service more than tripled year-over-year.
This segment has not yet significantly boosted Salesforce’s overall revenue, but investors have responded—the company’s stock price has risen approximately 23% since the announcement.
How can you still make money after being "bypassed"?
Salesforce is eager to change its pricing model for another, deeper reason: AI agents are reshaping the place of enterprise software.
In the past, after a company purchased Salesforce, employees had to actively open it and manually enter customer information, track sales, and handle service requests within the interface. The software itself served as the entry point for employees' work, so enterprise software companies naturally charged based on the number of users.
After the emergence of agents like Claude, this matter has become more nuanced.
When AI can directly operate software like Salesforce, employees may open these applications less frequently. Users simply need to tell Claude: “Organize this client’s information,” “Modify the order,” or “Check the inventory,” and the agent will handle the remaining tasks in the background.
If this way of working becomes mainstream, one of traditional enterprise software’s greatest assets—user access—could gradually be taken over by AI agents.
Salesforce chose not to block it. Last week, the company launched Claudeforce, enabling customers to directly access Salesforce data and perform numerous tasks involving Salesforce applications through Claude, without needing to manually navigate those applications.
According to individuals familiar with its sales strategy, Salesforce is likely to offer multiple pricing options for Claudeforce, though enterprises will first need to purchase a higher-tier Salesforce subscription plan.
Behind this lies a larger business strategy: even if users no longer directly access Salesforce in the future, Salesforce aims to generate revenue whenever other companies’ AI systems need to access its data and capabilities.
In other words, as the software interface is gradually pushed behind the scenes by agents, Salesforce is also seeking new "revenue channels."
Is the SaaS industry about to change dramatically?
Charging based on the income generated or costs saved by AI is not entirely new.
Palantir has been doing this for years. It signs highly customized contracts with large enterprises to unify their data and develop applications, combining fixed fees, usage-based charges, and performance-based pricing.
Today, Salesforce's AI pricing strategy is increasingly approaching this model.
This is also an interesting role reversal. More than twenty years ago, Salesforce itself was the company that redefined how software should be priced.
Before its emergence, businesses typically had to spend a large sum to purchase commercial software outright and manage their own upgrades and maintenance. After Salesforce popularized SaaS, businesses could pay subscription fees based on the number of employees, avoiding expensive upfront software licensing costs; as staff increased, they simply added more accounts, while all software updates were handled by Salesforce in the cloud.
This model later sustained the entire SaaS industry and sparked over two decades of software prosperity.
The transition isn't always smooth. For example, when the system monitoring software company Splunk shifted from traditional software licensing to a subscription model, its revenue briefly declined.
But now, AI has once again disrupted the old rules. This time, however, Salesforce is no longer the first to introduce new ways of doing things—it has begun learning anew from newer companies like Sierra and Fin about how software should be sold.
Can your account be balanced? How do you calculate it?
Charging based on results may sound customer-friendly, but it actually brings to the forefront a previously less important question: What truly counts as a result created by AI?
There are at least two barriers here. The first is that AI must first reliably complete tasks. Previously, pricing was based on tokens or API calls—even if an agent failed midway, the computing resources had already been consumed, and customers typically still paid. If the model shifts to “charge only upon task completion,” software vendors will bear a greater portion of the cost for each failed execution.
This is no easy task for current AI agents. Independent tests show that OpenAI’s Operator still has a relatively high failure rate on real desktop tasks; another survey covering 8,128 users found that agents successfully complete only about three-quarters of assigned tasks on average. In other words, when OpenAI begins experimenting with “pay only if successful,” it directly turns product reliability into a revenue issue: the more frequently agents fail, the more computational resources are wasted. Can this equation ever add up?
But even if AI successfully accomplishes the task, a second hurdle emerges: how much of this result should actually be attributed to AI?
Whether customer service issues have been independently resolved is relatively easy to determine; but once the basis for charging shifts to “how much more money was earned for the client” or “how much cost was saved,” things become much more complex. For example, if a company’s sales increased by 20% after using AI for three months, that growth could be due to the agent’s contribution, but it could also stem from new product launches, marketing campaigns, or seasonal fluctuations.
Stripe has specifically released an outcome-based pricing guide, reminding software vendors to clearly define attribution rules in advance, as otherwise customers and suppliers may easily dispute whether a given revenue result was generated by AI or would have been achieved regardless.
Therefore, shifting from tokens and seats to "outcomes" is much more than just changing a unit of measurement. For software companies to share in the incremental value created by their customers, they must first be willing to pay for failure and establish a mutually agreed-upon method to prove that those successes are genuinely tied to their AI.
From OpenAI and Sierra to Salesforce and Palantir, and now to a growing number of similar companies emerging domestically, this experiment is becoming increasingly vibrant.
As for how things will develop in the future, let’s wait and see!
Reference link: https://www.theinformation.com/briefings/openai-starts-letting-customers-pay-ai-works?rc=jn0pp4
© THE END
This article is from the WeChat public account "Machine Heart" (ID: almosthuman2014), authored by Machine Heart.