A customer service manager opens their dashboard. Overnight, an AI agent has categorized requests, located orders and prepared refunds. Ten employees still use the software, but some of the work no longer depends on their clicks. So why keep paying for ten identical subscriptions? This scenario illustrates the economic dilemma of AI agents: they promise greater productivity while undermining the traditional unit of sale for business software. By September 2026, the seat could become a less clear-cut measure of value. This projection draws on offerings and trends already observed, without assuming they will become widespread.
The seat, a quiet pillar of the SaaS model
Per-user billing accompanied the rise of online software, or SaaS. It offered a practical compromise: predictable spending for the customer, recurring revenue for the vendor and growth tied to the number of employees using the software. A company would hire salespeople, buy additional licenses and automatically increase its contract value. Premium features then provided a way to raise average spending.
This model does not disappear the moment a machine drafts an email. An assistant that helps every employee can even justify an additional per-user charge. Copilot offerings have largely followed this approach. But agents could change the equation: rather than merely suggesting a response, they can carry out a sequence of actions across several applications, within defined rules and controls.
The risk for vendors is therefore not just a decline in jobs. It is also a decline in the number of people who need full access. If an agent consults the CRM, updates records and prepares follow-ups, some employees may need only a report or a simplified interface. The volume of work increases; the number of billable seats does not necessarily follow.
Three ways to replace the user
Actual offerings have begun exploring alternative units. Intercom has offered pricing for its Fin agent tied to resolved requests. When Salesforce launched Agentforce in 2024, it highlighted a per-conversation price. These approaches illustrate the shift underway: selling less exclusively access to a tool, and more of its activity or impact.
Usage-based: measurable, but hard to predict
Model calls, credits, documents processed: usage brings revenue closer to technical resource consumption. This limits the vendor’s risk of serving an extremely active customer at a fixed price. For the buyer, however, budgeting can become opaque. A simple request and a complex search do not require the same resources; an agent that makes repeated attempts can drive up the bill without delivering more value.
Task-based: tangible, provided the terms are defined
Qualifying a lead, processing an invoice or opening a case are units more closely aligned with business operations. But their boundaries still need to be agreed upon. Does an invoice containing an anomaly count as an additional task? Is another attempt billed? This model is compelling when the process is repetitive and clearly defined. It is less so when every case turns into an investigation.
Outcome-based: appealing, but open to dispute
Paying for a request that is actually resolved appears to align interests. Yet the outcome requires a contractual definition. Is a customer who stops responding satisfied? Does a resolution still count if the case is reopened? And who gets the credit when the agent prepares the work but a human makes the final decision? An outcome is not an inherent metric: it is a negotiated agreement.
Margins also depend on what happens behind the screen
Traditional SaaS already incurs hosting, support and development costs. Agents add variable consumption tied to models, the tools they call and checks. A request can trigger several searches, document reads and verification steps. The actual cost of a reliable action can therefore exceed that of the generated response alone.
Falling costs for some models do not solve everything. More ambitious systems may consume more resources because they perform longer tasks. Vendors must balance economics, speed and reliability: using a less expensive model to categorize a document, reserving a more capable system for exceptions and involving a human when the risk becomes too high.
This reality is encouraging hybrid contracts: a base subscription for the platform, an included volume and variable charges beyond that. Seats remain necessary to supervise agents, manage permissions and handle exceptions. Subscriptions thus retain a role, but no longer account for the entire value of the contract on their own.
Customers want to buy savings, not credits
For a finance department, the relevant comparison is not simply between an existing subscription and an AI bill. It concerns the full cost of the process: integration, data preparation, supervision, error handling and time actually saved. Automating a quick operation does not offer the same value as resolving a case that had tied up several teams.
Nor do time savings automatically translate into budget savings. They can improve turnaround times, accommodate more requests or free up time for sales. That is valuable, but different from an immediate reduction in spending. A vendor that equates every minute saved with guaranteed savings risks losing credibility at renewal time.
Negotiations should therefore focus on a few concrete safeguards:
- a spending cap and alerts before it is exceeded;
- a verifiable definition of successful tasks and exclusions;
- rules covering errors, rework and human intervention;
- access to logs that allow billed actions to be audited.
A battle for the customer relationship
The shift also threatens vendors’ position within companies. If an agent becomes the primary interface, the underlying applications could find themselves relegated to databases or specialized engines. Whoever controls the agent could capture a growing share of the budget, while others would seek to protect their revenue through APIs, transactions or data access.
Established players nevertheless have strengths: integrations, access permissions, business history and approval workflows. An effective agent must be able to act without bypassing these safeguards. New entrants, meanwhile, can design an offering directly around completed work, without having to protect an existing license business. The advantage will depend as much on operational trust as on model performance.
What next? The most plausible scenario for September 2026 is not the death of the seat, but the coexistence of several billing metrics. The user will remain relevant for collaborative tools; tasks and outcomes will gain ground in clearly defined processes. Vendors will have to prove that their pricing tracks value rather than technical complexity. Buyers, meanwhile, would do well to experiment within a measurable scope before signing up to a promise of large-scale automation.