Authentication sequences that seem straightforward can burn through compute unpredictably when you're running them at scale. A single login might require handling two-factor flows, CAPTCHA challenges, session token management, and cookie persistence. Each has its own failure modes. When TinyFish's enterprise web agent infrastructure is orchestrating browser fleets across thousands of sites simultaneously, some authentication sequences complete in seconds while others require multiple retry attempts. The resource consumption varies by 10x or more, yet most pricing models treat all "actions" or "conversations" as equivalent.

Salesforce offers three different ways to pay for the same agent platform: per conversation, per action, or per user. ServiceNow charges "per assist". Microsoft sticks with flat $30 per user. Each pricing model represents a different hypothesis about where resources actually concentrate when agents run continuously.

Vendor	Pricing Model	What It Assumes
Salesforce	$2 per conversation → $0.10 per action	Shifted from conversation-based to action-based after learning conversations don't map to resource consumption
ServiceNow	Per assist	Discrete, bounded operations with predictable costs
Microsoft	$30 per user	Flat usage patterns across users
Amazon	No additional cost	Scale advantages make marginal costs negligible

This fragmentation reflects rapid learning playing out in public pricing decisions.

We see this pattern clearly in our infrastructure. Session management that feels trivial compounds in ways that don't scale linearly. Error recovery that looks like an edge case becomes the dominant cost driver at millions of browser sessions. The operational reality: obvious cost proxies like conversations, actions, or users don't map cleanly to actual resource consumption.

Salesforce's pricing shift suggests they learned conversations don't correlate with what's expensive to run. But even "actions" might miss the mark. Maintaining authentication across different booking systems, handling rate-limiting across regions, recovering from site structure changes—we learn quickly that some operations genuinely consume resources while others just look complex on capability lists.

Each pricing structure reveals different assumptions about what scales predictably. Right now, 68% of vendors charge separately for AI features, and 90% of CIOs report that managing AI costs limits their ability to drive value. That tension suggests vendors will need to consolidate around sustainable models faster than typical enterprise software cycles.

The trajectory over the next few months points toward consolidation around models that reflect actual resource consumption patterns. Vendors who've operated these systems at scale are pricing more accurately because they understand which operations actually compound. Running web agents that need reliable authentication, regional rate-limiting, and site structure adaptation reveals quickly that margin problems appear after deployment. Either you're overpaying for operations that don't actually cost much, or you're underpriced on operations that compound unexpectedly.

For enterprise teams evaluating agent platforms, the pricing model reveals what the vendor thinks is expensive to operate. When those assumptions align with how agents actually consume resources in your environment, costs stay predictable as you scale.

Worth considering: Does the vendor's pricing structure account for authentication complexity? Session persistence costs? Error recovery patterns? Or does it assume uniform resource consumption that won't hold at scale?

The market seems to be learning this faster than usual enterprise software pricing cycles. Pricing models are adjusting within months, not years. Deploy at scale, discover what's expensive, adjust pricing, repeat. The vendors who understand their true cost drivers will price sustainably. Those who don't will discover margin problems when usage scales beyond pilot deployments.

Pricing fragmentation is the visible surface of vendors learning what actually costs money at enterprise scale. And that learning is happening fast.