AI customer service agents: what they resolve on their own and how they are priced (September 2026)

What an AI customer service agent does without a human

The first job is answering questions from approved content. Fin answers on chat, email, voice, Slack and social channels, and it runs on Intercom as well as on other helpdesks such as Salesforce, HubSpot and Freshdesk. Help desks now ship the same function as part of the product: Zendesk includes AI agents in every Suite and Support plan.

The second job is action, and this is where agents differ from the chatbots many teams already had. Gorgias, built for online shops, lists order edits, subscription pauses, cancellations and refunds among its AI Agent actions. Ada gives its agents API, code and MCP tools to complete tasks in business systems. Decagon describes Agent Operating Procedures that define agent behavior in natural language, much like the procedures used to train human staff, and Sierra publishes agents to website chat, apps, WhatsApp, Apple Business Chat, SMS, email, phone and ChatGPT.

The third job is the handoff. An agent has to recognize when a case needs a person and pass it on with the context. How a vendor treats that moment tells you a lot about its bill: Fin counts a configured handoff to a human as a billable outcome, while Zendesk bills only requests the AI agent resolved without escalation.

Sources: [1][2][3][4][5][6][7][8][9]

What counts as a resolution: every vendor measures differently

Fin is the clearest case of a generous definition. A resolution counts when the customer confirms that the answer helped or does not ask for more help after Fin's last answer; if the customer disengages for 24 hours, Fin treats that as an assumed resolution. A Procedure that ends in a planned handoff to a human is also a billable outcome, charged once per conversation.

Other vendors draw the line elsewhere. Zendesk says customers pay only for requests the AI agent resolved successfully and without escalation to a human. Help Scout counts a resolution when the customer receives an AI answer and does not escalate, search the knowledge base, ask more questions or signal that they need more help. Gorgias bills an automated interaction for any request fully resolved without a human, which includes its rule-based Flows and order management as well as the AI Agent.

Some vendors avoid the argument by charging for volume. Freshdesk bills sessions, and an email session covers a 72-hour window regardless of the outcome. Salesforce charges $2 per Agentforce conversation. Decagon offers both models and writes that most of its customers prefer per-conversation pricing, because it avoids arguing over what a resolution is.

For a buyer this means two resolution rates from two vendors are not comparable. A customer who gives up and leaves looks exactly like a satisfied one in a silence-based count. Ask for the counting rule in writing and read a random sample of cases the vendor marked as resolved before you sign.

Sources: [7][10][2][11][12][13][14][15][16]

Published resolution rates are vendor figures

Vendors put high numbers on their websites. Fin states an average resolution rate of 76 percent across more than 12,000 customers, with many above 85 percent, and Salesforce repeated the 76 percent in its acquisition announcement. Ada's homepage shows an automated resolution rate of 84 percent and Gorgias shows 60 percent of tickets auto-resolved, both without saying which customers or which period the figure covers.

The best documented early case is Klarna. In February 2024 the payments company reported that its AI assistant had handled 2.3 million conversations in its first month, two thirds of its customer service chats, and did the equivalent work of 700 full-time agents, with the time to resolve an errand down from 11 minutes to under 2. In May 2025 its chief executive said that investing in the quality of human support was the way of the future for the company.

Analysts expect more automation, but not a service without people. Gartner predicted in March 2025 that agentic AI will autonomously resolve 80 percent of common customer service issues by 2029. In a Gartner survey released in August 2026, 87 percent of more than 3,500 customers said companies using generative AI in service must still offer access to a human agent, and in February 2026 Gartner warned that the cost per resolution for generative AI in customer service will exceed $3 by 2030, more than many offshore human agents cost.

Independent research exists mainly for AI that helps human agents. A study of 5,179 support agents published by the National Bureau of Economic Research found 14 percent more issues resolved per hour on average, and 34 percent for novice and low-skilled workers. It measured an assistant that suggests replies, not an agent that closes cases alone.

Sources: [1][9][4][3][17][18][19][20][21][22]

How AI customer service agents are priced: per resolution, per conversation or on request

Four vendors in this comparison publish a price per resolved case. Fin charges $0.99 per outcome; with another helpdesk there are no seat fees, but a minimum of 50 outcomes a month applies. Help Scout charges $0.75 per resolution for AI Answers and gives new accounts 3 months of unlimited resolutions. Gorgias charges $1.00 per automated interaction on monthly billing and $0.90 on annual contracts. Zendesk bills AI agents per automated resolution beyond the amount included in each plan, but does not print the rate on its pricing page.

Others charge for volume. Freshdesk includes 500 AI agent sessions once per account and then sells 100 sessions for $49. Salesforce offers three models for Agentforce: $2 per conversation, Flex Credits at $500 per 100,000 credits with 20 credits per action (30 for a voice action), or $125 per user a month for flat-fee access. A flat monthly fee exists too: moinAI, based in Hamburg, charges 750 euros a month on annual billing for unlimited conversations on its Business plan.

The platforms built for large companies quote. Sierra calls its model outcome-based: the criteria for each outcome, such as a resolved conversation, a saved cancellation or an upsell, are agreed upfront, and in most cases there is no charge for an unresolved conversation. Decagon sets fixed rates per conversation or per resolution with lower rates for larger commitments. Ada, Cognigy and Parloa publish no prices, and PolyAI prices its voice agents per minute on quote.

The AI fee rarely replaces seat costs; it comes on top. Intercom seats cost $29, $85 or $132 per month on annual billing, Zendesk's Suite Team costs 55 euros per agent a month in Germany and its Copilot add-on another 50 euros, and Freshdesk's Copilot costs $29 per agent. Use the calculator below for the AI part only.

Sources: [10][23][11][12][2][13][14][24][16][15][25][26]

Who sells AI customer service agents: our directory

Our directory lists 61 tools in customer service, counting tools whose main or second niche is support. They fall into five niches: AI agents and support chatbots, voice agents for the phone, helpdesks, knowledge bases and quality assurance, and many tools cover more than one. The United States dominates with 41 tools; Germany follows with 4, among them Cognigy, which NiCE bought in September 2025, Parloa and moinAI.

Prices are hard to find. The largest group, 26 tools, quotes on request or states no model at all, 13 charge by usage or outcome and 11 per seat. Among the six platforms we file mainly as AI agents, only Fin and moinAI publish a price. You can browse all of them in our customer service category or filter the list below.

Security documentation is more common than prices. 41 tools name a SOC 2 report and 21 can keep data in the EU according to our data. Only 8 name ISO/IEC 42001, the certificate for an AI management system: Fin, Sierra, MaestroQA and Observe.AI. For an agent that talks to customers unsupervised, that certificate is a useful question in any tender.

Sources: [27][28][10][24][29][30]

Rules that apply: AI disclosure since August 2026 and liability for answers

Since August 2, 2026, Article 50 of the EU AI Act requires providers to design AI systems that interact directly with people so that those people are informed they are dealing with an AI, unless that is obvious to a reasonably informed person. The European Commission names chatbots, AI agents and avatars explicitly and published guidelines on these transparency duties on July 20, 2026. The grace period until December 2, 2026 covers only the marking of AI-generated content under Article 50(2), not chatbot disclosure.

Breaches of Article 50 can cost up to 15 million euros or 3 percent of worldwide annual turnover, whichever is higher. A support agent that answers questions and processes returns is not on the list of high-risk uses in Annex III, whose rules apply from December 2, 2027 after the AI Omnibus. In Germany, the implementing law of July 22, 2026 makes the Bundesnetzagentur the main market surveillance authority for the AI Act.

In the United States there is no federal disclosure rule, but states have their own. California's BOT Act has made it unlawful since July 1, 2019 to use a bot that misleads people about its artificial identity in order to encourage a sale; a clear and conspicuous disclosure removes the liability. Utah's amended AI law, in effect since May 7, 2025, requires disclosure when a consumer clearly asks, and proactively only for high-risk interactions in regulated occupations.

Liability for what an agent says sits with the company. In Moffatt v. Air Canada, decided on February 14, 2024, a Canadian tribunal held the airline responsible for wrong refund information from its website chatbot and awarded 812.02 Canadian dollars in total. The tribunal wrote that it makes no difference whether the information comes from a static page or a chatbot.

Sources: [31][32][33][34][35][36][37][38][39][40][41]

Choosing and testing an AI customer service agent

Start from your conversations, not from the vendor's rate. Export the last three months of tickets, group them by reason and mark which ones an agent could answer from content and which need an action in another system. The share of repetitive, rule-bound requests is the realistic ceiling for automation, whatever the vendor average says.

If you already run a helpdesk, its own AI agent is the shortest route, because it sees the ticket history and the knowledge base without an integration project; Zendesk and Help Scout are examples. Standalone agents such as Fin, Sierra or Decagon make more sense when you run several systems or need the same agent on chat and phone.

Then run a pilot on real traffic with a hard budget. Compare the vendor's resolution count with your own review of a random sample, track how many customers came back within a week with the same issue, and price the second year, because usage-based bills grow with every new channel you switch on. For employee questions rather than customer questions, our analysis of HR chatbots covers the internal use case.

Sources: [2][11][1][6][5]