Over the past few months, we’ve heard Artificial Intelligence (AI) agent pricing discussed repeatedly across the IT-BPS outsourcing ecosystem. Everyone seems to have an opinion, and the consensus seems to be that AI agent pricing is a definite and highly prevalent outcome of how today’s models are evolving. By AI agent pricing, we mean a model in which an AI agent is treated as a countable commercial unit. While we’re seeing this model in deals today, whether AI agent pricing becomes mainstream depends on several hard questions that still must be answered.
Reach out to discuss this topic in depth.
To start with, a human agent is a measurable unit that delivers about 1,800 productive hours per year from an offshore location such as India. This unit has three fundamental properties: it is countable, because actual human agents need seats and the same can be audited; it is bounded; and the relationship between workload, number of FTEs, and provider revenue stays linear. Historically, the FTE has been the unit of measure across BPO, delivering finite hours and handling a finite and predictable number of work units, such as invoices processed and calls answered.
In an AI-first world, these properties break down. The number of AI agents in a process becomes a function of architecture and design choices, not the quantum of work. For example, one agent could deliver an end-to-end process, or multiple agents could deliver discrete parts of the process. A collections agent, for instance, could prioritize accounts and draft communications, or it could autonomously contact customers, interpret responses, negotiate payments, update the underlying ERP system, and escalate disputes. These architectural and design decisions can vary from one deal to the next and one provider to the next. The underlying quantum of work, or the number of work units to be processed, remains unchanged.
AI agent pricing is also decoupled from the value and outcomes that matter to clients. Buyers don’t care whether one AI agent or multiple AI agents are working autonomously; they care about STP rates, cost takeout, improvements in DSO, and working capital. AI agent pricing does not capture or reflect the value the underlying system is designed to deliver.
Another limitation of AI agent-based pricing is that it is largely disconnected from both provider cost and customer value. A provider’s underlying costs are driven by AI inference, compute consumption, platform infrastructure, orchestration, and the one-time effort to build, train, and deploy the solution. The number of agents configured is not an accurate indicator of underlying costs. Equally, customer value is determined by business outcomes, not the number of agents deployed. As a result, AI agent count represents an implementation choice rather than a meaningful commercial metric. AI agent count and its definition also become a long-term versioning issue. Buyers and providers should consider the following questions:
- When does an enhanced agent become a new billable agent?
- Does an agent remain the same unit after its underlying model changes?
- Does the provider or the buyer fund upgrades required to maintain or enhance performance?
- How are changes in consumption economics treated?
What we have seen in deals under our advisement are models in which outcomes are priced, such as gainsharing from an improvement in DSO or cost takeout, or output is priced, such as per invoice, per conversation, or per minute. Run costs associated with the AI-powered tools are built into the outcome-based or output-based price, and development and deployment charges are positioned as a one-time fee, increasingly amortized over the deal term or treated as an investment by the provider.
None of this means AI agents don’t reshape pricing. They do, profoundly. But they aren’t creating a new unit of price so much as accelerating a shift that was already underway, from input-based pricing around FTEs toward output- and outcome-based pricing. The agent is the delivery mechanism, not the meter. If it is to become the meter, buyers and providers should answer the following questions before adopting it:
- What is an agent? When one solution deploys one agent for a specific quantum of work and another deploys three to do the same work, what determines the price?
- How does agent count stay tethered to the outcomes buyers actually buy, such as STP rates, cost takeout, and DSO, rather than floating free of them?
- How are consumption costs, which are linked to the underlying work units being processed rather than to agent count, passed through, capped, or bundled?
- How can AI agents across different provider solutions be benchmarked, especially in a competitive Request for Proposal (RFP) situation?
If you found this blog interesting, check out, Everest Group launches Innovation Watch on Agentic AI in Wealth Management Technology – Everest Group Research Portal, which delves deeper into another topic relating to AI.
To take the conversation forward, contact Rahul Gehani ([email protected]) and Sahil Shah ([email protected]).