AI customer support pricing compared: Fin, Zendesk, Agentforce, Ada, Decagon vs BYO LLM

TL;DR

At 5,000 AI-handled conversations a month, Intercom's Fin bills about $59,400 a year, Zendesk's AI agents land between $90,000 and $120,000, and Salesforce Agentforce lists at $120,000. The same volume on OpenAgent with your own Gemini API key runs roughly $1,200 to $1,800 a year, platform fee included. That is not a clever discount; it is a different architecture. Per-resolution vendors resell LLM inference with a margin measured in thousands of percent. BYO LLM removes the reseller.

The four billing models in this market

• Per resolution / per outcome. Intercom Fin ($0.99), Zendesk AI (~$1.50 to $2.00), Ada (reported $1 to $3.50). You pay every time the AI succeeds. The vendor defines “success”.
• Per conversation or per action. Salesforce Agentforce: $2 per conversation, or Flex Credits at $0.10 per agent action (a multi-step answer consumes several actions).
• Quote-only enterprise contracts. Ada and Decagon publish no prices. Procurement data puts Decagon's median contract near $386,000 a year with a $50,000 platform fee before usage.
• Flat platform fee + your own LLM tokens. The OpenAgent model. The platform charges for the software; inference is billed by Google, OpenAI, or Anthropic directly to you, at list price, with no markup.

Published pricing, July 2026

Annualised at 5,000 AI-handled conversations a month. Seat fees for human agents are excluded on all rows (every vendor charges them separately except OpenAgent, where seats are unlimited and free).

The 30× math, shown honestly

The headline example: 5,000 monthly resolutions on Fin is 5,000 × $0.99 = $4,950 a month, or $59,400 a year. The same volume on OpenAgent is the flat platform fee ($36/site/year billed annually) plus roughly $100 a month of Gemini API: about $103 a month, $1,236 a year, a 48× difference. Round hard against ourselves and budget $150 a month all-in, and it is still $1,800 a year, 33×. Against Zendesk's pay-as-you-go rate or Agentforce the multiple runs 50× to 90×.

The $100 API figure is deliberately generous. Raw list-price math for 5,000 single-pass support answers on Gemini 2.5 Flash ($0.30 per million input tokens, $2.50 per million output) is closer to $6. We budget $100 to cover what real conversations actually cost: multiple turns, RAG context stuffed into each prompt, embedding calls for retrieval, and retries. If your bill comes in under that, the multiple only gets bigger.

Three caveats, because a comparison without them is marketing:

• Not every conversation is a billable outcome. Fin bills when the visitor confirms resolution, stops replying, or gets handed off through a configured workflow. If only 70% of your 5,000 conversations become billable outcomes, the Fin figure drops to about $41,600 a year. The order of magnitude does not change.
• Enterprise buyers negotiate. Committed volume gets discounts everywhere. The table shows list rates because that is what teams without a procurement department pay.
• Seats are excluded on both sides. Intercom and Zendesk also charge per human seat ($74+ and $55+ per agent per month respectively), which makes their real totals higher than the table shows, not lower. OpenAgent seats are unlimited.

Why per-resolution billing gets worse, not better

Per-resolution pricing has a structural problem: you are charged more as the product works better and as your business grows. It is a tax on your own success, and the meter is defined by the vendor. Zendesk's January 2026 change is the pattern to watch: every resolution above your committed volume now auto-bills at the full per-resolution rate, with no discounted overage band. Budgeting a usage-metered AI line item means forecasting your own ticket volume a year out and paying for the error either way.

Consolidation is pushing the same direction. Salesforce signed an agreement to acquire Fin (the company formerly known as Intercom) in June 2026 for about $3.6 billion. Salesforce already prices Agentforce at $2 per conversation, twice Fin's rate. When the two biggest per-resolution vendors share an owner, pricing power sits with them, not with you. A flat platform fee plus tokens at provider list price is the one structure where a vendor cannot quietly reprice your success.

When a per-resolution vendor is still the right call

• You want one throat to choke. Fin, Zendesk, and Agentforce bundle the model, the guardrails, and the liability into one contract. If your board wants a single accountable vendor for the whole AI surface, that bundle has real value.
• Your volume is tiny and stable. A few hundred conversations a month is under $300 on Fin. The gap is real but may not justify a migration.
• You are already deep in the ecosystem. If your org runs on Salesforce or Zendesk end to end, the integration gravity can outweigh the AI line item.

The regulated-industry and EU angle

For banks, insurers, healthcare, and the EU Mittelstand, the cost story is usually the second reason to avoid per-resolution vendors. The first is data control. With hosted AI agents, every customer conversation transits the vendor's model pipeline under the vendor's DPA. With BYO LLM, the inference contract is between you and the model provider you have already cleared with legal (including EU-region endpoints), the API keys are yours, and the RAG pipeline that grounds each answer is auditable. Teams with residency requirements run the Business tier: dedicated infrastructure in the region you need, audit log export, and the model endpoint region chosen by you, not by the vendor. Predictable spend and predictable data flow turn out to be the same architectural decision.

Check the math on your own volume

Every number above is a default, not a conclusion. The cost comparison calculator takes your monthly conversation volume, lets you edit every vendor's rate (if you have negotiated numbers, use them), and shows the annualised bill side by side. The related AI vs human ROI calculator answers the prior question: how much of your queue AI should be handling at all.

Sources

Rates checked July 15, 2026. Vendors change pricing; if a number here has drifted, the source links are the ground truth and we re-verify this page monthly.

• Intercom pricing page ($0.99 per Fin outcome, 50-outcome monthly minimum)
• Fin help center: what counts as a billable outcome
• Zendesk pricing page (Suite seats; AI resolution rates are quoted on contact)
• Richpanel teardown of Zendesk's AI meter (~$1.50 committed / ~$2 pay-as-you-go per resolution)
• Salesforce Agentforce pricing ($2 per conversation; Flex Credits at $0.10 per action)
• Featurebase teardown of Ada's quote-based pricing
• eesel analysis of Decagon contracts (Vendr procurement data)
• Gemini API list pricing (2.5 Flash: $0.30/M input, $2.50/M output tokens)
• Salesforce press release: definitive agreement to acquire Fin (June 15, 2026)