Your AI agent charges you even when it fails: the hidden tax of per-conversation pricing

Salesforce Agentforce charges $2 per AI conversation even when the case escalates to a human unresolved. At a 60% resolution rate, you end up paying $3.33 for every real resolution because you're also billed for the 40% of failed attempts. Zendesk auto-bills overages with no warning since January 2026. Per-conversation pricing turns every vendor failure into a line on your invoice. The way out: custom AI with fixed, predictable cost.

The model that charges for trying, not for succeeding

Salesforce launched Agentforce at $2 per AI conversation. The detail almost nobody read in the fine print: it charges for every interaction, resolved or not. If the agent fails to close the case and escalates it to a person, you still pay your $2 (Fin AI, 2026). The math is brutal: at a 60% resolution rate —optimistic for many support cases— the effective cost isn't $2, it's $3.33 per real resolution, because you're funding the 40% of conversations the AI couldn't close (SaaStr, 2026). It's a model where the vendor's failure becomes the vendor's revenue. Marc Benioff himself admitted the per-conversation scheme caused confusion and that customers 'pushed for more flexibility', nudging Salesforce to introduce Flex Credits.

It's not an isolated case: it's the new billing standard

The pattern is spreading across the whole category. Zendesk charges between $1.50 and $2.00 per automated resolution, plus a mandatory 'Advanced AI' add-on of $50 per agent per month, on top of your existing plan; and since January 2026 it auto-bills overages with no prior warning (Fin AI; CorePiper, 2026). HubSpot moved its Customer Agent to $0.50 per resolved conversation in April 2026. Each vendor paints its variant as 'you only pay for value', but the aggregate effect is the same: your bill no longer depends on how many users you have, but on how much the AI works —and that's exactly what you can't predict or cap. On June 4, 2026 Deloitte published a full technical guide on how to account for these outcome-based prices, a sign that billing chaos has already reached finance departments.

Why a cost you can't predict is a cost you can't govern

The underlying problem isn't the unit price, it's unpredictability. A single AI conversation can trigger hundreds of micro-actions, and with consumption billing your cost scales with usage, not with value. Finance leaders are trying to govern a dynamic model with tools built for fixed licenses, and the result is no longer theoretical: one company burned its entire 2026 AI budget by April after usage nearly doubled in two months (Citi Ventures, 2026). When your bill rises exactly as your business grows —more customers, more tickets, more conversations— the vendor has put a tax on your success. And unlike a license, you can't negotiate it at year-end: it's already spent.

The defense of outcome pricing (and where it breaks)

The case in favor is real: paying for outcomes aligns price with value, avoids buying seats nobody uses and, at low or irregular volumes, can be cheaper than a fixed license. For a company with few conversations a month, the usage model is perfectly rational. Where it breaks is on three points. First, 'outcome' is defined by the vendor, not you: Agentforce charges per conversation even when it resolves nothing. Second, the marginal cost never drops: every new customer you serve charges again, forever, with no economies of scale. Third, you lose control of the ceiling: with auto-billed overages, the limit is set by your volume, not your budget. The rule of thumb: if your volume is low and stable, pay-per-use works; if your business is growing or your support is intensive, fixed cost almost always wins.

The alternative: custom AI you pay once and scale for free

There's a third path between 'don't use AI' and 'pay the vendor for every conversation'. You build your own custom agent: you call the OpenAI, Anthropic or Google APIs directly and pay the real token cost —cents per conversation, not $2— integrated into your CRM, your helpdesk and your processes. The software is paid once; compute costs what it costs, with no middleman margin or failure surcharge. As the comparison shows, for 10,000 conversations a month the saving is around 80% in year 1, but the decisive part is the cost profile: your bill stops rising every time your business grows. If it unsettled you to learn you're paying for the conversations your AI doesn't resolve, request a custom AI software quote and take back a cost you can predict, cap and govern.