benchr
Most "best AI for customer service" lists rank a dozen logos by feature checklist and never tell you the one thing that decides your bill: how you buy. There are three ways to put an AI agent in front of your customers, and they price on completely different units. An off-the-shelf resolution bot charges per resolution. A helpdesk platform agent charges per seat plus per resolution. Building your own charges per token. The right answer depends on your monthly conversation volume and whether you have engineers to spare, not on which demo looked slickest. This page walks the three tiers in order, cheapest path of least resistance first.
Tier 1: off-the-shelf resolution bots
This is the fastest path live, and for most teams under 50K conversations a month it's also the cheapest in total cost of ownership. You point a bot at your help center, it answers, and you pay only when it closes a ticket. No engineers required.
Intercom Fin sets the bar. It's $0.99 per resolution flat, with no integration fees, no setup fees, and no separate platform charge, and Intercom reports it leading independent benchmarks at a 67% average resolution rate. That combination, a documented win rate behind a sub-dollar price, is why it's the default recommendation here. The catch is the definition of "resolution": you pay for tickets the bot closes, so your real cost depends on how many it deflects, which varies by how clean your help docs are.
Lorikeet is the runner-up on price-performance, the one to shortlist when you want strong resolution quality without anchoring to the Intercom ecosystem. Below the brand names, Quickchat AI and similar tools go as low as $0.50 to $0.60 per resolution, but they don't publish the case studies that back up Fin's quality claims, so the cheaper number carries more risk. Treat sub-dollar pricing as a starting point to validate, not a settled win.
Tier 2: helpdesk platform agents
If your support team already lives in a helpdesk, the AI agent bolts onto the suite you pay for. You get tight integration with tickets, macros, and routing you already run, but you pay twice: a per-seat license for your human agents plus a per-resolution fee for the bot.
Zendesk AI is the common case. The AI agent resolves at $1.50 to $2.00 per automated resolution, and the Advanced AI add-on runs $50 per agent per month on top of a Professional Suite seat. For a 20-agent team resolving 3K tickets a month, the realistic all-in lands around $6K to $8K a month once seats, add-on, and resolutions stack. The trap to watch: Zendesk introduced automatic overage billing in January 2026 with no prior notification, so usage above your committed volume can bill itself without a heads-up. Budget for the ceiling, not the average.
Higher up, the enterprise platforms quote custom and don't publish list prices. Ada starts around a $30K-a-year platform fee plus $1 to $3.50 per resolved interaction, with enterprise contracts often landing in the $150K to $300K-plus range. Sierra doesn't publish pricing; third-party estimates put annual contracts near $150K to start with $50K to $200K setup fees. Decagon also stays quiet on price, with median contracts reported around $400K a year on custom per-conversation or per-resolution terms. These earn their entry cost only at real enterprise scale, above roughly 500 seats and 100K conversations a month, where conversational depth, voice, and outcome-based contracts pay off. For an SMB, they're the wrong tier.
Tier 3: build-your-own on a model API
The third path skips the bot vendor and wires a model straight into your stack. You build a retrieval-augmented agent on a raw API, own the prompts and the knowledge base, and pay per token instead of per resolution. It's the cheapest unit cost at high volume and the most work to stand up.
Claude Haiku 4.5 is the natural engine for this tier: $1 per million input tokens and $5 per million output, the cost-control rung in Anthropic's lineup. Prompt caching cuts as much as 90% off repeated context, and batch processing saves another 50% on non-urgent work, which is exactly the shape of support traffic where the same help-center context gets read on every ticket. The Claude Haiku 4.5 review covers where the cheapest Claude is enough for production, and the per-use-case pricing guide breaks the same models down by workload, including RAG and agents, so you can sanity-check the math before committing an engineer to it.
The economics are simple but unforgiving. Building a RAG support chatbot runs $8K to $25K upfront on a 2-to-6-week timeline for a simple single-LLM agent, and a production agent serving 10K conversations a month costs roughly $500 to $1K in API spend. That's why DIY only wins at scale: at 10K conversations it barely beats Fin, but the token line grows slower than per-resolution fees, so it pulls ahead above roughly 30K a month. Two costs hide in that estimate. Embedding and vectorization for RAG add 10 to 30 percent on top of raw token spend, and the $8K-to-$25K figure is for a simple agent: Gartner puts enterprise-scale knowledge retrieval at $750K to $1M-plus, so don't price a single-LLM project as if it were an enterprise platform.
The cost angle, side by side
The three tiers price on three different units, so the only honest comparison is total dollars at a fixed workload. Here's the same volume, 10K monthly conversations at a 75% resolution rate, run through each model.
| Tier and tool | Pricing unit | Rate | ~Monthly cost |
|---|---|---|---|
| Build-your-own (Claude Haiku 4.5) | Per token | $1 / $5 per 1M | $500–$1K |
| Off-the-shelf (Quickchat AI) | Per resolution | $0.50 | $3.75K |
| Off-the-shelf (Intercom Fin) | Per resolution | $0.99 | $7.42K |
| Platform (Salesforce Agentforce) | Per resolution | $2.00 | $15K |
Read the build-your-own row with a grain of salt. The $500-to-$1K figure is API spend only and doesn't carry the upfront build, the embedding overhead, or the engineer who maintains it. Fold those in and the per-resolution bots win clean at this volume, which is the whole reason Tier 1 is the default. The token line only takes the lead once conversations climb well past this point, and the curve crosses earlier the more your traffic repeats the same questions.
One more variable sits underneath every row: which model you put behind the agent. Support is a high-volume, latency-sensitive job where a cheap fast model usually beats a flagship, so the token tier leans on Haiku rather than a top-end model. The same reasoning shows up in the guide to the best AI for research, where the bottleneck is honest citations rather than raw cost, and in the built-in-versus-standalone breakdown for email AI, where high inbound volume tips the math toward automation the same way support tickets do.
Frequently asked
Which off-shelf AI customer service platform offers the lowest per-resolution cost?
Intercom Fin at $0.99 per resolution, with no platform fees or integration charges. Quickchat AI and similar alternatives cost $0.50 to $0.60 per resolution but lack published case studies on resolution quality, so Fin is the safer pick when you want a documented benchmark behind the price.
When should a team build their own support agent instead of buying off-shelf?
Build your own if three things are true: you handle more than 100K conversations a month, so token cost becomes cheaper than per-resolution fees; you have internal ML or engineering capacity, since a DIY agent needs ongoing tuning; and proprietary integrations are critical. For volumes under 50K a month, Zendesk, Fin, or Ada are the better buy.
What is the true total cost of Zendesk AI per agent per month?
Add three lines: the Professional Suite seat at about $55, the Advanced AI add-on at $50 per seat, and per-resolution fees of $1.50 to $2.00 multiplied by the average resolutions per agent that month. For 3K tickets a month at 50% automation, expect $350 to $450 per agent per month all-in. Zendesk also added automatic overage billing in January 2026, so usage above your committed volume can push that higher without notice.
Does Claude Haiku 4.5 make sense for production customer support at scale?
Yes, if you handle more than 50K conversations a month. With a RAG setup cost of $5K to $15K and $500 to $1K a month in API spend at 10K conversations, Haiku becomes cheaper than Fin at volumes above roughly 30K a month. It needs an engineering team for RAG maintenance, and embedding and vectorization costs can add 10 to 30 percent on top of the raw token spend.
Are Ada, Sierra, and Decagon worth the high entry cost?
Only for enterprises above roughly 500 seats with complex integrations and more than 100K conversations a month. For SMBs, Fin or building on Claude Haiku run 3 to 5 times cheaper. Ada and Sierra excel at conversational depth and voice; Decagon leans on outcome-based contracts. None of the three publish pricing, so every quoted figure is a third-party estimate that varies by contract.
Changelog
- May 30, 2026 — Originally published. Intercom Fin, Zendesk AI, Ada, Sierra, Decagon, and Claude Haiku 4.5 pricing verified against vendor documentation and third-party 2026 pricing analyses; estimates for vendors that don't publish list prices are flagged as such.
References
- Fin by Intercom, "AI Customer Service Agent Pricing Comparison," fin.ai, accessed May 2026.
- Lorikeet, "Best AI Customer Support Platforms 2026: Ranked by Resolution Rate," lorikeetcx.ai, accessed May 2026.
- CorePiper, "Zendesk AI Agent Pricing Per Resolution in 2026: Complete Guide," corepiper.com, accessed May 2026.
- Anthropic, "Claude API Pricing — Official Documentation," platform.claude.com, accessed May 2026.
- DestiLabs, "How Much Does It Cost to Build an AI Agent in 2026?," destilabs.com, accessed May 2026.
- Quickchat AI, "AI Agent Pricing Models 2026: Per-Resolution vs Per-Seat Compared," quickchat.ai, accessed May 2026.