AI agents for customer service

Intro — how this agent works

A customer service agent reads the ticket, retrieves the right answer from your help centre and past tickets, takes an action against your systems (refund, reorder, address change, subscription pause), and replies. What separates it from the chatbots of 2021 is that it acts rather than deflects — and that it is sold on outcomes, not seats.

The ticket, step by step

The word "resolution" is doing a lot of work

Every vendor bills per resolution and every vendor defines it differently. Intercom's Fin charges $0.99 per outcome — when the customer confirms it was resolved, asks no follow-up, or a workflow completes. Zendesk counts a resolution when its AI handles the issue start to finish, and charges nothing when it escalates. Salesforce bills a conversation — a 24-hour session that may contain dozens of actions — at about $2.

• Confirmation-based (Fin) is the strictest: the customer effectively has to agree it worked.
• No-escalation (Zendesk, Gorgias) is the most common: if a human touches it, you don't pay.
• Session-based (Agentforce conversations) is the loosest: one billable unit can cover a whole day of back-and-forth.

The consequence: a headline rate is not comparable across vendors without knowing the denominator. Model your own ticket mix before believing any per-resolution quote.

Why this vertical matured first

• Text is forgiving. No latency budget, no barge-in, no turn detection. A support agent can think for four seconds and nobody notices.
• The training data was already there. Every company has years of resolved tickets — a labelled corpus of question-to-correct-answer pairs sitting in the helpdesk.
• The buyer already had a budget line. Support is a measured cost centre with a known cost per ticket, so the ROI conversation is arithmetic rather than vision.
• The ceiling is your documentation, not the model. Teams that fail at this almost always fail on retrieval quality and stale help centres, not on the LLM.

Where it still breaks

• Actions, not answers. Answering "what's your refund policy" is solved. Issuing the refund against a billing system with edge cases is where deployments stall.
• Escalation quality. A bad handoff — losing context, making the customer repeat themselves — destroys more goodwill than never having automated at all.
• The long tail is the expensive tail. The 20% of tickets that don't automate are the complex ones, so your remaining human team gets a harder job, not just a smaller one.

Part 1 — the providers

Three groups: helpdesk incumbents bolting agents onto an existing seat business, AI-native platforms selling resolutions, and the CRM giants bundling agents into the suite.

Figures marked "est." come from third-party reporting; those vendors do not publish pricing. Rates change often — verify before committing.

Part 2 — how each provider handles the work

AI-native platforms — resolutions are the product

• Intercom Fin — the one to benchmark against, partly because it will run on top of Zendesk or Salesforce rather than demanding you migrate. $0.99 per outcome, and an outcome only counts when the customer confirms resolution, asks no follow-up, or a workflow completes — the strictest definition in the market, which makes the headline number more honest than most. Seats are separate ($29/$85/$132 per seat/mo) but not required for standalone Fin.
• Decagon — sells to large consumer brands with a $50,000 annual platform fee covering all channels, integrations, "Agent Operating Procedures" and its Watchtower QA monitoring, then roughly $0.99 per conversation on top. The QA layer is the differentiator: it grades its own agents continuously rather than leaving you to sample transcripts. Tripled to a $4.5B valuation on a $250M Series D in January 2026, on about $35M of 2025 revenue.
• Sierra — the same outcome-based model it uses for voice, at roughly $1.50 per resolution, with unresolved conversations typically free. Contracts are six figures with real setup fees, so it is a transformation programme rather than a signup.
• Ada — automation-first and unusually strong on multilingual, reportedly $1–$3.50 per resolution from about $30k/year, but it expects to sit alongside Zendesk or Salesforce rather than replace them.
• Forethought — blends a platform fee with outcome pricing at roughly $59,500/year, and is strongest at triage and routing on messy ticket taxonomies rather than pure deflection.

Helpdesk incumbents — agents attached to a seat business

• Zendesk AI Agents — outcome pricing grafted onto the incumbent: about $2 per automated resolution pay-as-you-go, falling to roughly $1.50 with committed volume, and nothing when the agent escalates. But it sits on top of Suite seats at $55–$115 per agent per month, and the Copilot add-on is a further $50 per agent per month. The all-in number is rarely the per-resolution number.
• Freshworks Freddy — the value option, bundled into a cheaper suite and billed per session. Fewer capabilities per resolution, but the total bill for a small team lands well below the AI-native platforms.
• Gorgias — the clearest pricing in the category and the tightest vertical fit: $0.90 per resolved conversation on annual plans, $1.00 monthly, with overage interactions at $1.50. Because it is built around Shopify, order lookups, returns and refunds work without integration work.

CRM suites — the agent is a feature of the platform

• Salesforce Agentforce — three ways to pay, which tells you the model is still settling: about $2 per conversation, Flex Credits at $500 per 100,000 (20 credits per action, so $0.10 an action, and $0.15 for voice), or per-user licensing from $125/user/mo. The arithmetic matters: conversations win above ~20 actions each, Flex Credits win below. The real argument for Agentforce is not price but that the agent already sits on your Service Cloud data.

Voice, hybrid and DIY

• PolyAI and Parloa — where support is phone-dominant, these replace the IVR rather than the chat widget. Priced as managed enterprise engagements, with tuning and support inside the contract. See the receptionist page for how their voice stacks work.
• Crescendo — sells AI and managed human agents as a single service with outcome-linked commercials, which is effectively BPO with the labour arbitrage replaced by model inference.
• Build your own — an LLM API, retrieval over your help centre and past tickets, and webhooks into your ticketing system. Marginal cost is cents per ticket rather than dollars, and for a high-volume, narrow-workflow business the savings against $1/resolution are enormous. You are buying that with evaluation harnesses, guardrails, escalation logic and QA that the platforms give you on day one.

Part 3 — pros and cons

How to pick, in 30 seconds

• Under ~2,000 tickets/month → Fin. The benchmark, no migration, honest billing unit.
• You sell on Shopify → Gorgias at $0.90/resolution beats every generalist on both price and fit.
• Already deep in Zendesk or Salesforce → use their agent and accept the premium; migration costs more.
• Large consumer brand, tone matters → Decagon for the QA layer, Sierra if it is a board-level programme.
• Many languages → Ada, then Parloa if the volume is voice.
• High volume, narrow workflows, engineers on staff → build it. The per-resolution economics invert above roughly 50k tickets a month.

The investor read

• Per-resolution pricing is the category's defining event. Support is the first software market to widely abandon seats for outcomes, and it happened in under three years. The band has already compressed to $0.90–$2.00 — which is what commoditisation looks like while it is still happening.
• The definition of "resolution" is the real pricing power. Fin's confirmation-based unit and Agentforce's 24-hour conversation are not the same product at the same price. Vendors that loosen the denominator raise prices without touching the rate card. Watch that number, not the headline.
• Distribution is beating technology. Zendesk and Salesforce ship worse agents at higher all-in cost and still win deals, because the data and the contract are already there. The AI-natives' answer — Fin running standalone on rival helpdesks — is the most important competitive move in the category.
• The DIY floor caps the ceiling. An in-house agent costs cents per ticket. Every platform is selling the gap between that and $1, filled with evals, QA and guardrails. That gap narrows every time the tooling improves — the durable vendors are the ones building QA and monitoring moats (Decagon's Watchtower), not answer quality.
• Verticals out-earn generalists. Gorgias charges the least per resolution and is the hardest to displace among Shopify merchants, because the actions — refund, reorder, address change — are pre-integrated. Expect the same pattern in healthcare, travel and fintech support.

Frequently asked questions

How much does an AI customer service agent cost?

The market has converged on $0.90–$2.00 per resolution. Gorgias is $0.90 on annual plans, Intercom Fin $0.99 per outcome, Sierra around $1.50, and Zendesk about $2 pay-as-you-go. Enterprise platforms add a floor — Decagon charges $50,000 a year before usage, Forethought roughly $59,500.

What counts as a "resolution" in AI support pricing?

It varies by vendor and it materially changes your bill. Intercom charges only when the customer confirms resolution or a workflow completes. Zendesk and Gorgias charge when no human touches the ticket. Salesforce bills a 24-hour "conversation" that may contain dozens of actions. Always model your own ticket mix rather than comparing headline rates.

Is Intercom Fin better than Zendesk AI agents?

Fin is cheaper per resolution ($0.99 vs about $2), uses a stricter billing definition, and will run standalone on top of Zendesk. Zendesk wins only when you are already on Zendesk and want a single vendor. On pure economics Fin is the benchmark the others are measured against.

Should we build our own AI support agent instead?

Above roughly 50,000 tickets a month the economics invert — an in-house agent costs cents per ticket against a dollar or more per resolution. What you take on is evaluation harnesses, guardrails, escalation logic and continuous QA, which is exactly what the platforms sell. Below that volume, buying is almost always cheaper than the engineering.

What percentage of support tickets can AI resolve?

Well-scoped deployments report containing the majority of tickets, but the number depends far more on your documentation quality and how many actions the agent can take than on the model. Teams that stall usually have stale help centres or no API path to issue refunds and changes — not a weak LLM.

Which AI support agent is best for e-commerce?

Gorgias, on both price and fit. At $0.90 per resolved conversation it is the cheapest transparent rate in the category, and because it is built around Shopify, order lookups, returns and refunds work without custom integration.

Other AI agent verticals

Sources

Pricing read from vendor pages and reporting in August 2026: Intercom Fin, Salesforce Agentforce, Zendesk per-resolution, Gorgias, Decagon pricing, Ada & Forethought estimates. Funding and traction: Decagon $250M at $4.5B, Sierra $950M at $15.8B, Decagon revenue. Contact-centre voice-share figures are reported industry estimates, not a Flowpicker benchmark.