Agentic AI, the more focused alternative to general-purpose generative AI, is gaining momentum in the enterprise, with Forrester having named it a top emerging technology for 2025 in June.
Since then, several organizations have begun using the technology, and major vendors such as Salesforce and ServiceNow have offered AI agents to customers.
Agentic AI focuses on performing specific tasks and emphasizes operational decision-making instead of the content generation often associated with gen AI tools.
The technology is in its early days, and several questions remain open — chief among them, how AI agents will be priced. So far, no agreement exists on how pricing models will ultimately shake out, but CIOs need to be aware that certain pricing models will be better suited to their specific use cases.
Salesforce, for example, offers three pricing models: one that includes 1,000 Agentforce “conversations” free with its Salesforce Foundations CRM service; another included with its standard success plan; and $2 per conversation a la carte. Salesforce defines a conversation as a customer sending at least one message or selecting at least one menu option or choice other than “end chat” within a 24-hour period.
The $2-per-conversation approach can include many back-and-forth interactions between a customer and Agentforce, says Ryan Shellack, senior director of AI product marketing at Salesforce. The company is focused on use-based pricing, with only one customer seat required to administer it, he adds.
Lots of pricing models to consider
The per-conversation model is just one of several pricing ideas. In a recent LinkedIn post, Box CEO Aaron Levie outlines four agentic AI pricing models that could emerge.
First, vendors could base the price of AI agent tasks on the traditional work they replace, with a discount on the traditional labor price. “An AI Agent performs a certain amount of work, and you pay for amount of time or units it took to do that work,” he writes. “Generally, it’s a fair trade for the customer and provider.”
Second, agents could be priced based on outcomes, with the price focused on the completion of a task. “This model allows for a simple relationship between what the customer needs and what they’re paying to get accomplished,” Levie writes. “It also has the benefit that as underlying AI costs drop over time service providers can extract more margin for this work.”
A third way that AI agents could be priced is by calculating the underlying costs and charging a small markup, he says. “This can be great for technically-savvy customers but has the risk of not being sufficiently abstracted from AI costs to hold value over time,” he says. “Potentially good for customers, but maybe not for shareholder returns.”
Finally, agentic AI vendors could offer a per-seat SaaS subscription model that gives users unlimited access to the agent, Levie says. “This model could be quite disruptive,” he writes. “In areas where there are lots of seats used by end-users, it’s possibly very strategic; in areas where there’s only a small number of seats, you’re likely giving up too much value.”
More pricing models are likely to come forward, Levie says in an interview. These are “fairly exciting times to watch new business models in software emerge after a decade plus of limited changes,” he writes.
Conversations and subscriptions
A per-conversation model seems to be an emerging approach, says Sesh Iyer, managing director, senior partner, and North America regional chair at BCG X, Boston Consulting Group’s IT building and designing group. Vendors could also charge a small price per audio input or output.
Alternatively, a token-based consumption approach would bill tokens used for assistant API tools at the chosen language model’s per-token input and output rates, he adds.
An early trend seems to be the SaaS model, with a per-conversation model emerging for infrequent users, says Ritu Jyoti, general manager and group vice president for AI, automation, data and analytics research at IDC.
Outcome-based pricing could be tricky, she says, when it’s still difficult to define a successful outcome in an AI agent intervention. Outcome-based pricing may lead to disputes between vendors and users over whether the desired effect was achieved, Jyoti says, although pricing based on resolutions can work well in customer-service situations.
“It is all dependent upon the features and usage volume,” she adds. “What is really desired from enterprises, as they kind of get into this whole adoption, is that they are looking for subscription-based pricing with tiered plans based on features and usage volume. The reason is because enterprises look for some predictability.”
However, some experts see other price models emerging. Outcome- and cost-based pricing, with variations for specific use cases, are likely to catch on, says Rogers Jeffrey Leo John, co-founder and CTO of DataChat, a no-code, gen AI platform for instant analytics. In comparison, current large language model pricing is a form of outcome-based pricing, with users paying for tokens processed or generated, he notes.
“For business users, outcome-based pricing is often the most intuitive,” Leo John says. “This model directly ties the cost to specific outcomes or successful completions, making it easier to relate to the value delivered.”
Cost-based pricing will also be appealing because it’s straightforward to calculate, he adds.
“By pricing based on the underlying costs of compute, latency, and throughput, this model provides clarity on how charges are determined and allows for more precise budgeting,” Leo John says. “While it may lack the direct ROI alignment of the outcome-based model, it simplifies the financial planning process for users who understand and manage technical resources.”
Dangers of consuming too much
While several pricing models may emerge, CIOs and IT leaders should beware of consumption pricing, says Jeremy Burton, CEO of Observe, an AI-powered observability platform.
“It all sounds good, but the challenge is that people get annual budgets and cannot tolerate variability,” he says. “Everyone assumes if they move from subscription to some form of consumption, then they’ll save money. That is until they see a spike and burn through half of their budget in a few weeks.”
Some big vendors in the IT industry can demand consumption pricing, he says, “but from what I’ve seen at the app level it’s a nightmare.”
Focus on your business needs
Pricing will evolve as the agentic AI model does, and CIOs should explore the options that best fit their needs and their use patterns, experts say.
“AI costs have been dropping significantly, with the cost per unit of AI work decreasing and capabilities improving rapidly,” Leo John says. “Vendors may move towards hybrid models that combine cost-based transparency with performance-driven incentives. The ongoing advancements in AI will drive continuous evolution in how AI services are priced to remain competitive and aligned with market demands.”
CIOs should consider specific use cases and desired outcomes with AI agents, Leo John adds. This assessment can determine whether an outcome-based pricing model or a cost-based model, based on operational expenses such as computing and throughput, is more suitable.
CIOs should also consider total cost of ownership, he says. “With new and improved models emerging almost daily, leaders must also account for the costs associated with retraining or customizing AI models in this rapidly evolving landscape,” he adds.
CIOs and IT leaders must determine how their organizations will use the AI agents to determine the best pricing for them, adds BCG X’s Iyer.
“They should assess what is available today with an in-depth understanding of pricing and volume, build forecasts for at-scale usage, and build scenarios for increase in unit costs —like cloud — with optionality to switch agents to prevent lock-in,” he says.
Transparency and predictability will be the drivers for agentic AI pricing, says Box’s Levie. The vendors that provide predictable prices — and outcomes — will win in the market, he predicts.
“You need a high degree of visibility into what you are actually going to be paying for, so you don’t have these big, opaque systems when you can’t really anticipate what’s going to happen,” he says. “You generally can’t have a situation where you see a 10x spend increase, because something happened within the system that that was a surprise.”
