AI Agents Are Starting to Break the Old SaaS Pricing Model

The next fight in enterprise software is not only over who has the smartest artificial intelligence agent. It is over who gets paid when that agent actually does the work.

That shift moved from investor slideware into the daily business news cycle Tuesday after Sierra co-founder Clay Bavor told CNBC that AI agents are moving from demos into customer service, sales and support workflows — and that outcome-based pricing could challenge the way software companies have traditionally charged customers. CNBC framed the interview around a simple business question: if agents are built to complete tasks rather than merely answer questions, should software vendors still charge mainly by user seat?

For the technology industry, that is not a narrow pricing debate. It cuts into the heart of the software-as-a-service model that made cloud companies some of the most durable businesses of the past 20 years. The classic SaaS bargain was easy to understand: a company paid a subscription fee, usually per employee or per seat, and the vendor promised ongoing access, updates and uptime. AI agents complicate that bargain because the unit of value may no longer be a human user logging into a dashboard. It may be a support case resolved, a customer authenticated, a reservation changed, a sales lead qualified, a refund processed, or a workflow completed without a person touching every step.

Sierra is leaning into that change explicitly. The company’s own product page tells prospective customers to “pay for a job well done” and says its model uses outcome-based pricing so customers “only pay for the value Sierra delivers.” Its platform pitch is aimed at customer-facing agents that can work across chat, SMS, WhatsApp, email, voice and ChatGPT, with tools for building, testing, monitoring and improving agents before they go live.

That is the story: AI agents are not just a new feature layer on top of enterprise software. They are becoming a pressure test for how enterprise software gets bought, measured and renewed.

Why this matters now

The timing matters because the market is moving from “Can this model answer a question?” to “Can this system run a business process reliably enough to price against?”

In the CNBC interview, Bavor described agents moving into real workflows, especially in customer service, sales and support. He also pointed to a problem every enterprise buyer recognizes: return on investment is still hard to measure when AI tools are sold as broad productivity boosters. A chatbot that drafts text may be useful, but the CFO wants a cleaner line between cost and value. Did it reduce support time? Did it lift conversion? Did it resolve more issues without escalation? Did it create enough measurable savings or revenue to justify the bill?

Outcome-based pricing tries to answer that by changing the invoice. Instead of charging only for access to software, the vendor charges around the action, conversation or completed job. That can make the sales pitch sharper: pay when the agent does something valuable. It can also make the business risk sharper: if the agent fails to perform, the vendor’s revenue is exposed too.

This is why the agent pricing debate is bigger than Sierra. Salesforce, Microsoft and other enterprise software companies are already experimenting with pricing structures that move beyond simple per-seat subscriptions.

Salesforce’s Agentforce pricing page says customers can choose among consumption-based pricing, Flex Credits, Conversations and per-user licensing. Salesforce describes Flex Credits as a unit of payment intended to align cost with the value agents create, and says agent actions are metered individually. The same page gives examples such as an agent checking an order status, assisting a Salesforce user with a customer case, helping a field service technician schedule appointments, and answering onboarding questions for a new employee. Salesforce also lists conversation pricing for external-facing customer agents and says Flex Credits and Conversations are not supported in the same org.

Microsoft is also mixing old and new models. Its Copilot Studio page still lists Microsoft 365 Copilot at $30 per user per month, paid yearly, but it also offers Copilot Studio plans built around prepaid credit commitments and pay-as-you-go usage through Azure. Microsoft describes Copilot Studio as a platform for building and managing agents that connect to business data, complete tasks and publish across channels used by employees and customers.

Put together, the pattern is clear: the industry is not abandoning seats overnight, but the unit economics are being renegotiated. Seats still matter for internal users. Consumption matters for model usage. Outcomes matter when agents are supposed to replace or compress work.

The old SaaS model was built around human users

Traditional SaaS priced around a person because a person was the scarce interface. Sales reps used CRM. Support agents used ticketing systems. Finance teams used planning software. Engineers used development tools. The software was valuable because it organized human work and made that work more measurable.

AI agents change the interface. If a customer asks about an order and an agent authenticates the customer, checks the order system, retrieves shipping data and gives a status update, the customer may never become a “seat” in the old enterprise software sense. The worker may not either, at least not for that interaction. The value is in the completed task.

That has obvious appeal for buyers. If a business pays per seat, it may be paying for unused capacity or vague productivity. If it pays per resolved case, it can connect spending to an operational metric it already tracks. For startups trying to sell into large enterprises, that can shorten the trust gap: the buyer does not have to believe every claim about AI transformation; it can watch the meter.

But there is a catch. The more pricing is tied to outcomes, the harder the definition of “outcome” becomes.

A support conversation that ends without escalation might look like a success until the customer returns angry the next day. A sales agent that books a meeting might look productive until the meetings are low quality. A health care or financial services agent may need to optimize for accuracy, compliance and escalation discipline rather than speed alone. A retail agent may drive conversion but also create return risk if it oversells. In regulated industries, “done” is not always the same as “done safely.”

That is why the last mile of deployment matters. CNBC’s summary of the Bavor interview said the conversation covered testing customer-facing agents before they go live, rising AI token costs and the difficulty of deploying enterprise AI in real workflows. Sierra’s own site emphasizes observability, experiments, monitors, flagged issues and visibility into agent actions. Those are not decorative features. They are the operating controls that make outcome pricing less reckless.

The startup angle: selling accountability, not magic

For startups, outcome-based pricing is both a wedge and a trap.

It is a wedge because incumbents often have large installed bases and per-seat revenue streams they do not want to disrupt too quickly. A startup can walk into a sales cycle with a cleaner promise: do not pay for a license shelf; pay for completed work. That message lands especially well in customer support, where businesses already track resolution rates, handle times, escalation rates, satisfaction scores and cost per contact.

Sierra’s customer page shows why that market is attractive. The company lists customer stories across financial services, health care, media, retail, technology, telecommunications, travel and hospitality. It highlights examples including Rocket Mortgage, Singtel Group, Airtable, SoFi, Kraken, GoFundMe, Redfin, Chime, Vivid Seats, SiriusXM, Ramp and Sonos. The page also displays specific performance claims for some customers, including Airtable at an 80% resolution rate, Chime at a 70%+ resolution rate, Capital on Tap with a 33-point NPS improvement and Sun & Ski with a 3x conversion rate increase.

Those claims should be read as company-published customer marketing, not independent audits. Still, they show the commercial logic. AI-agent startups are trying to sell into business processes where value can be counted.

The trap is that accountability cuts both ways. If an agent is priced against outcomes, the vendor must carry more responsibility for the quality, safety and cost of the work. That means better testing, stronger integrations, clearer logs, faster rollback and more careful customer success work. It also means the vendor has to manage its own AI infrastructure costs. If token prices, model-routing choices or support volume move faster than expected, the vendor’s gross margin can get squeezed.

That margin question is the quiet business story under the hype. Classic SaaS companies became attractive to investors because incremental software usage could be highly profitable after the platform was built. AI agents are more complicated. Every interaction may carry compute cost. More complex workflows may require more model calls, tool calls, retrieval, verification and monitoring. If pricing is too low, growth can be expensive. If pricing is too high, buyers may retreat to pilots.

So the winners will not simply be the companies with the flashiest demos. They will be the companies that can connect agent performance, compute cost and customer value tightly enough to make the invoice feel fair.

Big Tech has a different problem

For Microsoft, Salesforce, ServiceNow, Adobe, Oracle, Google Cloud and other enterprise platforms, the agent era creates a balancing act. They want AI agents to expand revenue, protect core products and keep customers from drifting to startups. But they also have to protect existing subscription lines.

That is why hybrid pricing is emerging. Microsoft can keep a per-user Copilot plan for knowledge workers while using credit and pay-as-you-go models for custom agents. Salesforce can offer per-user licensing, conversation pricing and Flex Credits. The point is not that one model has won. The point is that enterprise AI is forcing vendors to maintain several models at once.

That creates complexity for buyers. A CIO comparing agent products now has to ask not just “Does it work?” but also:

• What counts as an action, conversation or completed outcome?
• What happens when an agent needs several model calls to finish one task?
• Who pays for failed or escalated interactions?
• Can the buyer cap usage or receive threshold alerts?
• How are human handoffs counted?
• What audit trail exists if the agent makes a wrong recommendation?
• Does the pricing reward speed at the expense of quality?

These questions are not procurement trivia. They determine whether AI agents become trusted operating infrastructure or another line item that finance teams cut after pilot season.

The customer-service beachhead

Customer service is the logical first battleground because the pain is obvious and the metrics already exist. Businesses spend heavily on contact centers, outsourced support, help desks and internal operations. Customers hate waiting. Employees hate repetitive tickets. Executives want lower costs without torching customer trust.

AI agents promise a tidy answer: resolve routine issues instantly, escalate messy cases to humans, personalize responses using customer data and learn from every interaction. In practice, the work is messier. Enterprise support sits on top of old systems, inconsistent policies, edge cases, privacy obligations and brand-sensitive moments. An agent that works in a demo may break when it meets a customer with a half-canceled order, a changed address, a loyalty credit, a prior complaint and a legal restriction on what can be said.

That is why the companies building in this category talk so much about deployment, monitoring and guardrails. The agent has to connect to real systems, not just generate fluent text. It has to know when to stop. It has to leave a record. It has to avoid inventing policy. It has to escalate without making the customer start over.

If vendors can prove that reliability, outcome pricing becomes powerful. It lets buyers pay for measurable value, and it lets vendors share in that value. If vendors cannot prove it, the model becomes a fancy way to argue over bills.

What to watch next

The next phase of the AI-agent market will be less about launch announcements and more about contracts. Watch how companies define “resolution.” Watch whether pricing pages become more transparent or more complicated. Watch whether customers demand caps, refunds or service-level guarantees. Watch whether agents are measured by cost savings alone or by a broader scorecard that includes accuracy, escalation quality, customer satisfaction and compliance.

Also watch gross margins. If agent companies promise outcomes but absorb rising compute and support costs, investors will start asking whether the economics look like software, services or something in between. That distinction matters for valuations, IPO readiness and acquisition appetite.

For buyers, the practical takeaway is simple: do not buy an AI agent only as a productivity story. Buy it as an operating system for a specific workflow, with a pricing model that matches the work and a measurement plan that survives contact with real customers.

For software companies, the message is more uncomfortable. The seat is no longer sacred. In the agent era, the invoice is going to follow the work.

Sources and verification

This article is based on CNBC’s July 14 report and interview summary on AI agents and SaaS pricing; Sierra’s public product and customer pages; Salesforce’s Agentforce pricing page; and Microsoft’s Copilot Studio product and pricing page. Company-published customer results are attributed as company claims, not independently audited performance figures.

• CNBC: “AI agents are moving into real work. That could change how software companies get paid” — https://www.cnbc.com/video/2026/07/14/ai-agents-saas-pricing-sierra-clay-bavor.html
• Sierra product page — https://sierra.ai/
• Sierra customer stories — https://sierra.ai/customers
• Salesforce Agentforce pricing — https://www.salesforce.com/agentforce/pricing/
• Microsoft Copilot Studio — https://www.microsoft.com/en-us/microsoft-copilot/microsoft-copilot-studio

---

Shadowfetch is a technology publication. Explore Shadowfetch Linux — our own Linux build — and the Shadowfetch apps on the App Store.