Town Raised $108M at a $1 Billion Valuation by Giving Every Employee Their Own AI Coworker. The Per-Employee Model Is the New Seat License.

A Forkast analysis published October 25 tracked eight enterprise AI agent platforms and found that four of them publish no public price at all. OpenAI's enterprise agent product is fully quote-gated. Anthropic lists its Enterprise base subscription at $20 per seat per month but discloses nothing about agent execution costs. Google's Gemini for Enterprise agent capabilities require a sales call to price. Meta's enterprise AI products have no public rate card. Of the eight platforms tracked, only SAP has a clean, transparent per-action price: 0.02 AI Units per autonomous action — and SAP is the only legacy enterprise software vendor in the group, not one of the five AI labs competing for the platform layer.

A separate analysis from Foreign Policy Journal on October 25 confirmed that across OpenAI, Anthropic, Google, Microsoft, and Meta, no lab publishes its full enterprise AI agent rate card. This is not an accident of product immaturity. The five labs have been selling enterprise AI products for two or more years. They have the pricing infrastructure and the customer base to publish. They choose not to.

Understanding why they choose not to — and what it means for every enterprise buyer evaluating these platforms — is the central challenge of enterprise AI procurement in Q4 2026.

The Rate Card Nobody Will Show You

Price opacity in enterprise software is not new. Salesforce, Oracle, and SAP have operated for decades with list prices that bear little relationship to what large customers actually pay. The enterprise software industry's baseline posture is that pricing is negotiated, not posted.

What is new about enterprise AI agent pricing is that the opacity is structural, not just strategic. In traditional enterprise software, the list price exists even if it is routinely discounted. The published rate is a ceiling, and negotiation determines how far below it you land. In enterprise AI agents, there is often no list price to start from — not because vendors are withholding a known number, but because the metering model is genuinely indeterminate until the customer's specific workload is analyzed.

An enterprise AI agent executing a simple FAQ lookup consumes perhaps 500 input tokens and produces 200 output tokens. The same agent, retooled to handle a complex procurement approval workflow — evaluating a purchase request against five internal policies, querying the vendor database, checking the requester's budget authority, and generating a structured approval document — might consume 80,000 tokens across fifteen reasoning steps. These are the same "agent action" in product terms. Their actual cost differs by a factor of 160.

This is the structural problem. The pricing unit — an agent action — is not a stable unit of value. It is a variable-length sequence of compute operations whose cost scales with task complexity. Publishing a per-action price would be immediately exploited by customers who design simple agents to minimize cost and then escalate to complex workflows once committed.

ServiceNow's AI revenue crossed $1 billion in 2026, with roughly half coming from usage-based billing. ServiceNow chose to address this complexity through capacity bundles: customers buy a block of "AI requests" at a negotiated price, with overages billed at a different rate. This is honest pricing — it acknowledges that consumption varies — but it still does not answer what an AI request actually costs until you analyze your workloads.

Why Complex Value Delivery Breaks Simple Pricing

Before AI agents, enterprise software pricing had three stable models. Per-seat pricing captured the cost of access. Per-consumption pricing (per API call, per document, per row) captured usage intensity. Outcome-based pricing, which some SaaS companies experimented with in the early 2020s, attempted to align pricing with value delivered.

AI agents do not map cleanly to any of these. According to ICONIQ Growth's 2026 State of Go-to-Market report, 48% of B2B SaaS companies now run hybrid pricing as their primary model. That hybrid is typically a seat fee for access plus a consumption fee for usage above a threshold. The consumption component is almost always where enterprise AI value is generated — and where pricing becomes opaque.

The deeper problem is that per-token and per-seat pricing both measure the wrong thing. Per-token pricing measures compute cost, which is a cost metric, not a value metric. A reasoning step that identifies an error in a million-dollar procurement decision consumes 2,000 tokens. A reasoning step that drafts a routine status update consumes 2,000 tokens. They cost the same. They are not worth the same.

Per-seat pricing measures headcount, which is a proxy for usage but not for value. A user who invokes an AI agent once per day for simple lookups pays the same seat fee as a user who invokes it forty times per day for complex workflow automation. The seat pricing model systematically undercharges for high-intensity users and overcharges for low-intensity ones.

The MindStudio analysis of enterprise AI pricing models frames this as the "decoupling problem": the two metering systems that enterprise software has relied on for twenty years measure compute and headcount, neither of which maps reliably to the value that AI intelligence actually produces. Outcome-based pricing — charging per completed workflow, per decision made, per hour of human work automated — is theoretically correct but practically impossible to implement without extensive instrumentation of the customer's internal processes.

The result is that enterprise AI vendors are stuck: they cannot publish per-token pricing because it underprices high-complexity work, cannot publish per-seat pricing because it underprices high-intensity users, and cannot publish outcome-based pricing because they cannot measure outcomes they do not control. Quote-gated pricing is not the ideal solution. It is the least-bad solution given the constraint.

What Each Lab Actually Charges

The table illustrates the full spectrum. SAP is the outlier with a transparent per-action price, published as 0.02 AI Units per autonomous action. SAP can afford this transparency because it is not competing for the AI platform layer — it is pricing AI capability as a feature of its existing ERP platform, where customers are already locked in and the pricing comparison set is narrow.

The five AI labs — Anthropic, OpenAI, Google, Microsoft, and Meta — are all competing for the same enterprise AI platform contract. Transparency is a competitive vulnerability for them in a way it is not for SAP. If OpenAI publishes its enterprise agent rate card, Anthropic can immediately train its sales team to undercut it. If Anthropic publishes first, OpenAI benefits. The Nash equilibrium is for all five to keep their prices private — which is exactly what they do.

Microsoft's December 1 Problem

Five weeks before Microsoft's own December 1 licensing change turns its entry-level enterprise agent pricing into a variable structure, there is still no common unit of comparison across the eight platforms the Forkast analysis tracked.

Microsoft's December 1 change is specifically about Copilot Studio — the platform enterprise developers use to build Microsoft-native AI agents. Before December 1, Copilot Studio consumption was metered primarily on message volume: each agent interaction counted as one message, and message prices were relatively stable. After December 1, the metering model shifts to action-level billing: the price depends on how many actions an agent executes within a session, with variable rates based on action complexity.

According to SAMexpert's Agent 365 analysis, Microsoft Agent 365 is currently $15 per user per month standalone, covering agent management, governance, and security controls. The December 1 change does not affect this base fee — it affects the Copilot Credits consumption that pays for actual agent execution. Enterprise customers who have built agents in Copilot Studio and designed them assuming current message-based metering need to remodel their cost assumptions before December 1, because action-based billing can move the effective price of a complex agent workflow by 30–80% in either direction depending on how the agent is architected.

OpenAI's Decisions API, which became generally available in October 2026, was the AI lab move that most directly forced Microsoft's hand on agent pricing. The Decisions API gives enterprise customers a structured way to execute multi-step AI agent workflows with traceable action logging — which is exactly what Agent 365 and Copilot Studio provide. Microsoft's response was to make its agent platform's pricing structure more sophisticated, not simpler. A more complex pricing model is harder for OpenAI to benchmark against in sales calls.

The Competitive Dynamic Behind Price Opacity

The deeper strategic logic of hidden pricing is not just about avoiding apples-to-apples comparisons. It is about controlling which attributes the sale is made on.

When a customer can compare Vendor A at $0.03 per action and Vendor B at $0.04 per action, the sale becomes a price negotiation. When neither vendor publishes a per-action price, the sale shifts to capability demonstrations, integration quality, and relationship. This is a better sales environment for vendors with strong platforms and engineering teams — which describes all five AI labs — and a worse environment for procurement teams who need to model total cost of ownership.

PricingSaaS logged 230 verified pricing and packaging changes across 4,412 tracked pricing pages in Week 42 of 2026. Twenty increases, nine cuts. Credit pricing rose to 21% of tracked changes — up from 8% the prior year. The industry is moving toward credit-based and consumption-based models, but the headline enterprise AI labs are moving more slowly than their SaaS counterparts. The calculus is different when you are selling at eight- and nine-figure contract values: the opacity enables custom pricing that extracts more value from large accounts than a published rate card would allow.

This dynamic is well-understood in enterprise software. Workday, ServiceNow, and Salesforce all operate with list prices that function more as opening bids than real prices. Enterprise AI is following the same path, but faster and with less price memory — customers have no historical reference point for what a "fair" per-agent-action price should be, because the product category is less than two years old.

How to Navigate Quote-Gated Pricing

Enterprise AI procurement teams operating without a published rate card can still build defensible cost models. The negotiation is different, but it is not unwinnable. Here is a five-step framework:

1. Demand sample invoices from reference customers before your first sales conversation. Ask the vendor to provide actual invoice line-items — anonymized for commercial sensitivity — from a customer whose workload resembles yours in scale and complexity. If the vendor cannot or will not provide this, it means they either do not have comparable customers or they are unwilling to let you benchmark against real consumption data. Both are red flags.

2. Define the metering unit in writing before the first contract proposal. Ask: what exactly counts as one "action," one "message," or one "AI Unit"? Get the definition in the contract, not just in the sales deck. Vendors often use product-level definitions at the time of sale and contract-level definitions that are subtly different. The contract definition is what you will be billed against.

3. Request the consumption cap mechanism in the product UI, not just the contract. Every enterprise AI agent deployment should have a hard spending cap that triggers an alert or a pause before overages can accumulate. Ask the vendor to show you, in the admin interface, exactly where this cap is set and what happens when it is hit. "We can add that to the contract" is not the same as "here is where it appears in the product today." Products without consumption guardrails are pricing for your inability to monitor usage.

4. Build your own load test before accepting vendor benchmarks. Before any vendor-conducted performance or pricing demonstration, run your own benchmark: take a realistic sample of your intended agent workflows, estimate the token consumption and action count per workflow, and use the vendor's published API pricing (where it exists) or the quoted rate to compute your expected monthly spend at target scale. Compare this to the vendor's own estimate. Discrepancies of more than 25% indicate a pricing model disagreement that will appear in your first invoice.

5. Negotiate a 24-month price freeze covenant. Enterprise AI vendors are actively repricing as they understand consumption patterns at scale. A contract signed at current rates can face significant price increases at renewal, particularly as the vendor's cost modeling improves. Negotiate a price freeze for your initial term — typically 24 months — with a defined escalation cap (CPI plus 5% is common in enterprise SaaS). Vendors with confidence in their pricing will sign this. Vendors planning to reprice aggressively will resist it.

What Buyers Get Wrong at the RFP Stage

The most common enterprise AI agent procurement error is evaluating vendors primarily on model quality benchmarks — which is the equivalent of buying a commercial kitchen based on the celebrity chef endorsement rather than the utility costs, the maintenance contract, and the food cost per plate.

Model quality matters, and it is genuinely difficult to evaluate without a controlled benchmark against your specific workloads. But in enterprise AI agent deployments, the procurement-stage factors that most reliably predict long-term total cost of ownership are: the flexibility of the metering model, the quality of the consumption monitoring tooling, the contract terms around price changes, and the integration depth with your existing systems. These factors determine whether your actual costs track your projected costs over the first 18 months — which is the period in which most enterprise AI programs either demonstrate ROI or lose executive sponsorship.

Ramp's $60 billion valuation in 2026 was partly driven by enterprise demand for AI spend management — the capability to monitor, attribute, and govern AI consumption costs across a large organization. The fastest-growing category in enterprise AI tooling right now is not the agents themselves but the observability and cost management layer that sits on top of them. That is a market signal about how difficult enterprise AI cost management has become without it.

After the First Quote

When the vendor's first proposal arrives, the real negotiation begins. The proposal will quote a base subscription, a consumption allocation, and overage rates. Three things to verify before accepting any of these numbers:

The allocation is sized for your workload, not for the vendor's churn model. Vendors often quote allocations large enough that you will not hit overages in the first 90 days but structured so that realistic scale-up sends you into overage territory before year one ends. Ask the vendor to model your expected consumption at full deployment scale and confirm the allocation covers it.

The overage rate is not the true cost driver — it is the margin mechanism. Most enterprise AI agent contracts are structured so that the base allocation is reasonably priced and the overage rate is where margin is captured. Read the overage rate carefully and either negotiate it down or size the base allocation to avoid it entirely.

The metering model can be changed at renewal. Unless you negotiate otherwise, enterprise AI vendors can change the definition of their metering unit, add new billable categories, or restructure their credit model at contract renewal. The December 1 Microsoft change is the clearest example of this happening at scale. Build explicit protections against metering model changes into any multi-year agreement.

Takeaway

Takeaway: Enterprise AI agent pricing opacity is not an oversight — it is the product of three converging forces: metering models too complex for a stable published rate, competitive dynamics that reward opacity over transparency, and a product category young enough that vendors are still learning what their workloads actually cost to serve. For enterprise buyers, the practical response is to treat quote-gated pricing as a negotiation starting position, not a fixed ceiling. Demand sample invoices from reference customers, define the metering unit in the contract, negotiate a price freeze, and build your own load test before accepting any vendor's cost projection. The vendors who resist these requests most aggressively are the ones whose pricing models will surprise you most at invoice time. Microsoft's December 1 change is a preview of what the next two years of enterprise AI repricing will look like: more variable, harder to forecast, and structurally favoring buyers with strong procurement discipline over those without it.