A customer service department handles more requests but hires fewer people. A sales team automates its follow-ups without adding licenses. This is the paradox AI agents are bringing to the heart of business software: a product can become more useful while having fewer human users. For vendors, the challenge is no longer simply to sell a better interface. They need to find a new unit of value—and convince customers that it will not turn their budgets into a meter running out of control.
Looking ahead to September 2026, this is a defining tension in the market. The following analysis draws on developments documented through 2024; the potential paths beyond that point are forward-looking scenarios, not established findings.
The seat: a unit that no longer seems so obvious
The success of subscription software rests on an easy-to-understand promise: a set number of euros per user, per month. Businesses know their headcount, negotiate a volume and forecast their spending. Vendors benefit from recurring revenue that grows when customers hire or equip additional teams. This model has accompanied the rise of sales management, collaboration and human resources tools.
It has never been perfect. Licenses sit unused, some employees use several competing solutions, and others barely log in. But the seat provides a convenient proxy for value: the more people work with the software, the more useful it is assumed to be.
AI agents challenge that shortcut. Unlike an assistant that suggests text, an agent aims to carry out a sequence of actions: look up information, open a case, update data, trigger a procedure. Its autonomy remains bounded and its reliability varies. But if an agent performs some of the work of several people, the number of human accounts becomes a less accurate reflection of the workload the software actually handles.
Three ways to sell automated work
The first response is to retain the subscription and add an AI option. Microsoft 365 Copilot and GitHub Copilot, both commercially available before 2025, illustrate this per-user add-on approach. It is straightforward and compatible with existing purchasing arrangements. It is particularly well suited to individual assistance: everyone keeps their workspace, simply enhanced.
The second approach is usage-based pricing. Cloud platforms and generative model APIs have already familiarized businesses with requests, data volumes and text tokens. For agents, the unit might become a conversation, an action or a workflow run. This brings vendors’ revenue more closely into line with activity and, to some extent, their own technical costs.
The third, more ambitious approach charges for an outcome. As early as 2023, Intercom offered resolution-based pricing for its Fin agent. This is a significant shift: customers no longer buy only access to a tool, but a measurable service. This approach could extend to other processes, provided success is precisely defined.
Outcomes seem simple—until the first dispute
Consider an online store. A customer asks where their parcel is. The agent checks with the carrier, replies and closes the conversation. Should a resolution be billed if the customer returns the next day? What if the answer was accurate but incomprehensible? What if a human adviser had prepared the necessary information beforehand?
A business outcome is not always a technical event. A closed ticket can conceal a discouraged customer. A booked appointment may be of no value. A processed invoice may need correcting later. The more pricing depends on the outcome, the more its definition becomes a matter for contracts, audits and sometimes disputes.
Incentives can even diverge. A supplier paid by the number of actions does not have the same economic incentive as a customer seeking to eliminate unnecessary steps. A supplier paid per resolution might favor easy requests. This does not invalidate these models, but it does require quality metrics, reopening windows and exclusion rules.
Variable billing spreads across business functions
For finance departments, the risk resembles one already encountered in the cloud: consumption spread across many teams, then aggregated into a bill that is difficult to predict. With agents, the execution path adds another layer of uncertainty. A seemingly simple request can trigger multiple searches, model calls and correction attempts.
A successful sales campaign can therefore increase sales, conversations and automation spending simultaneously. Conversely, a poor configuration can generate activity without creating value: cases processed again, agents passing a task back and forth, answers requiring human rework. Usage rises; the benefit does not necessarily follow.
Comparing advertised prices is therefore no longer enough. The relevant calculation is the total cost of a correctly handled case: subscription, consumption, integration, oversight and exceptions. An agent that costs less per interaction may prove more expensive if it requires more exchanges or transfers requests to an employee too often.
What buyers need to negotiate
The issue therefore goes beyond negotiating a discount. Before rolling out an agent widely, a business should test it within a limited scope representative of its operations, measuring performance before and after. The aim is to distinguish tasks that have genuinely been eliminated from those merely shifted into oversight.
- An explicit unit: specify what triggers billing, including failures, duplicates and rework.
- Budget safeguards: caps, alerts, approval thresholds and defined behavior when the quota is exhausted.
- Usable traceability: link spending to a process, a team and a verifiable outcome.
- Contractual quality standards: define acceptable errors, remedies and conditions under which no charge applies.
A cap alone is not enough, however. Automatically stopping an agent that handles urgent requests could disrupt service. The contract and operating arrangements must provide for a fallback mode: queuing, human handoff or continued processing of priority operations only.
The seat is not disappearing—it is losing its monopoly
For vendors, abruptly abandoning licenses remains risky. Revenue becomes more variable, while each execution can incur computing costs. For customers, paying for outcomes may seem reassuring, but it makes comparisons between suppliers complex if each defines its own unit.
The most plausible future scenario is therefore coexistence: a subscription for access, security and governance; a usage allowance for automation; and potentially outcome-based payments for clearly defined tasks. The seat remains relevant for collaboration and decision-making tools. It is less relevant when software works largely without human intervention.
What comes next? Looking ahead to September 2026 and beyond, the central issue may be less the disappearance of per-user pricing than the ability to make automated work verifiable. Vendors will have to prove what they accomplish, not merely count what they execute. Businesses, meanwhile, would do well to buy demonstrated savings rather than a promise of autonomy.
