AI agents for customer service: how to evaluate them

AI agents for customer service answer and resolve customer requests across chat, email, and voice, and hand the conversation to a person with its context when they can't finish. To evaluate one, test it against your own ticket mix: which requests it resolves on its own, which actions it can take in your systems, and how it hands off the rest.

As of September 2026, Airframe's Customer Support AI market report tracks 135 products from 128 companies. The report tags 89 of those products as core AI support agents and 46 as adjacent, a group that includes helpdesk, contact-center, and CRM platforms that have added agents. The report draws on 6,137 vendor case studies and counts 12,508 customer organizations using these products.

Airframe's narrower AI Customer Support Agents report (August 2026) covers ten dedicated agent products and splits them into three groups:

• Standalone agent platforms: Sierra, Ada, Decagon, Netomi, and Maven AGI.
• Helpdesk-native agents: Intercom Fin, Forethought, and eesel AI.
• Regulated and complex-workflow agents: Lorikeet and Gradient Labs.

The same report rates this group an emerging category: KPI results vary widely and no vendor has pulled ahead. In the wider market, Zendesk, Salesforce Service Cloud, Freshdesk, Genesys, NiCE, and other platforms sell agents inside products many teams already run. Voice is a market of its own; Airframe's Voice AI Agents report tracks 42 vendors and its contact center (CCaaS) report tracks 73.

Many of the agent companies are young. Among the 135 products, 51 come from companies founded in 2020 or later, and 38 from companies founded in 2022 or later, including Sierra, Decagon, Lorikeet, Crescendo, and Parahelp.

What has changed since mid-2025

Several vendors have been bought or agreed to be bought. The Customer Support AI report lists these deals:

Sierra also bought Fragment, a Paris company, in April 2026.

Buyers are also switching helpdesks. Of the 6,137 case studies, 2,350 state the customer's prior solution, and 332 record a named switch between vendors the report tracks. Zendesk is the most common prior vendor, with 157 named switches away, and Zendesk to Gorgias is the most common single move (64). The products gaining the most named switches are Gorgias (76), Zendesk (36), DevRev (23), Intercom (22), and Pylon (21). Case studies are published by the vendor that won, so the counts track documented moves only and do not measure market share.

What vendors report

The figures below are medians of results as published in vendor case studies. Vendors publish their successes, so read them as the upper range of what is reported, with small samples behind most of them.

Results vary widely. In the wider report, the middle half of automated resolution results runs from 35% to 79%. The result for your team will depend on your ticket mix, the state of your knowledge base, and which actions the agent is allowed to take.

Build or buy

The Customer Support AI report sets out three options:

• Keep support led by people when volume is low, accounts are high value, or the product resists standardization. The costs are slower responses, paid after-hours coverage, and headcount that grows with volume.
• Build only a narrow agent, such as grounded answers plus a small set of approved account or order actions. The lasting work comes after launch: curating knowledge, encoding policies and exceptions, maintaining integrations and access controls, and running monitoring and fallback paths.
• Buy a managed agent when you need broad channel coverage, integrations, and governance. The work shifts to configuring and governing the product, and you still have to test conversations, exception paths, action permissions, and escalation context yourself.

The report's advice is to build only where a proprietary workflow justifies owning the system permanently.