Top 20 Customer Support AI Agents Compared (2026)

A customer support AI agent answers your customers on its own and can resolve a conversation without a person. The 20 agents here are built in one of four ways: inside a help desk (Fin, Zendesk AI agents, Agentforce), as a separate AI agent platform connected to your help desk (Decagon, Sierra, Ada), in an agent builder (Botpress, Chatbase), or as your own agent running on a support platform's API (Plain, Front and Pylon name the feature). Beyond testing and price, what separates them is what each agent can reach, and whether the agent that investigates is kept apart from the one that talks to customers. Plain is the agentic infrastructure for customer work, and this guide is published by Plain, a B2B customer support platform (help desk) with its own AI support agent.

Disclosure. Plain publishes this guide, and Plain's Customer Agent (formerly Ari) is one of the 20. It falls short in four places. Microsoft Teams is on Horizon and up and Discord is on Frontier; which plans include connected agents is on Plain's pricing page. And it works on threads in Plain, so choosing it means choosing the help desk.

How was this list built?

An agent is on the list if it talks to customers on its own, can resolve a conversation without a person, and its maker documents it on the maker's own site. Agent-assist copilots, which help a person write the reply, are covered in our guide to AI copilots for customer support. Agents built mainly for ecommerce and retail brands are out.

Every vendor fact comes from that vendor's own pages, checked in November 2026, and links to them. "Not documented" means the pages we checked don't say it; it does not mean the agent can't do it, so ask during your evaluation. Decagon, Sierra, Maven AGI and Forethought keep their product docs behind a login, so their rows rest on public pages only. The list is grouped by how each agent is built, not ranked, and Plain's entry uses the same fields as the other 19. This guide compares the agent, not the help desk underneath it.

The top 20 AI agents at a glance

Checked November 2026.

What are the four ways customer support AI agents are built?

The architecture decides how many systems you run, where a handoff lands and who sends you the AI bill. It usually decides whether you keep your current help desk.

Checked November 2026.

Fin is built into Intercom and, per fin.ai, also runs on Salesforce, Freshdesk, HubSpot and Zendesk. The trade-off is usually one more system against more control. As one buyer told us, "I didn't want to build a Frankenstein system."

Should the agent that investigates be separate from the agent that talks to customers?

It depends on what the agent can touch. Know what each agent can reach. Read access to code, logs and internal notes is safer in a team-facing agent with no customer channel; actions a customer asks for need per-action controls on the agent that takes them.

The OWASP Top 10 for LLM applications lists Excessive Agency as a risk, with excessive permissions and autonomy among its root causes, and recommends minimum privileges plus human approval of high-impact actions. Plain's own docs make the same split: "An internal agent normally doesn't write to customers directly, allowing you to give it access to sensitive tools and data sources" (Plain docs).

Three things decide whether a vendor supports that split: a separate team-facing agent, a way to pass work between agents, and a way to run your own agent on the same queue. The table uses four tiers for the team-facing agent:

• Investigating agent: a team-facing agent that reads internal systems (threads, code, docs, tools) and can act for your team, with no channel to customers.

• Operations agent: improves or configures the customer-facing agent, or analyses support data.

• Internal employee agent: answers your own staff's questions (IT, HR, internal knowledge).

• Reply copilot: drafts or suggests replies for a person to send.

Who investigates, who talks to customers. Checked November 2026.

On the public pages we checked, only Plain and Pylon document all three. Agentforce documents agent-to-agent at the platform level, though Salesforce's pages disagree on whether A2A is generally available or in pilot, and Zendesk's workflow-to-agent step is in early access.

Plain. Sidekick, the team-facing agent, reaches threads, CRM, issue tracker, docs, codebase, analytics and call recordings, and "cannot send messages to customers on its own" (Sidekick). The Customer Agent answers from knowledge sources and "currently is unable to carry out actions" (Plain docs). A workflow sends work to Sidekick or to your connected agent, including after the Customer Agent hands off (workflows). Your own agent connects through a machine user, an API key and webhooks (Plain docs), at 15 credits per assignment plus 15 per discussion; plan availability is on Plain's pricing page.

Pylon. Assist Agent, the team-facing agent, works privately with your team, investigates internal systems and drafts replies (Pylon). The Support Agent answers customers from approved, customer-agnostic knowledge, and Pylon says private account data, internal conversations and logs are not exposed to the customer. A Skill can hand an issue from Pylon's agent to your own (Pylon docs). Your own agent connects through webhooks, the API or MCP, on every plan; no per-unit price is published.

Some teams arrive with that agent already built; as one buyer at an AI company told us, "We're going to offload the heavy investigation to our own agent." Plain's docs name Decimal, Duckie and Parahelp as third-party agents that run on Plain, plus Inkeep's Keep, a copilot (Plain docs). Front's External Agents page names Claude, Codex, Decagon and Sierra. Vercel's Case Retro Agent is Vercel's own agent, not Plain's: in the words of the Vercel case study, "Plain didn't build the Case Retro Agent. Vercel did, using Plain's API and support for bring-your-own agents." The case study reports that the agent saves 65% of ticket volume. Plain's approach to running several agents on one queue is on its agentic infrastructure page.

Can you test an AI support agent before customers see it?

Yes. Almost every vendor documents a preview (16 of 20) or simulations (18 of 20). Far fewer let you see the agent on your own traffic: five document a bulk replay of your past tickets, seven a mode that runs on live conversations and sends nothing, and five document both. Plain has the no-send mode but no past-ticket replay.

Testing before launch. Checked November 2026.

Fin moved new workspaces from Batch tests to Evals on November 5, 2026. Zendesk is rolling out its AI agent launchpad from November 5 to November 30, 2026, and retires the legacy "Test AI agent" button on December 10, 2026 (Zendesk). Freddy's replay and shadow mode apply to the Email AI Agent only.

Which AI support agents take actions, who approves them, and how do they hand off?

Most of these agents take actions through a named feature, and the control on those actions differs. Four vendors document approval by your staff and Lorikeet a partial version; three document confirmation by the customer, two an identity check, and four an allowlist of tools.

Actions and handoff. Checked November 2026.

MCP shows up in two directions, and they need different controls. When the agent calls your MCP servers, it can act in your systems. When a vendor exposes an MCP server, Claude, ChatGPT or Cursor can work with your support data, with whatever permissions you grant.

On handoff, Plain leaves an internal note listing the reason, the search terms and the documents read; Agentforce transfers the conversation history; Decagon and Sierra pass a summary; Ada documents transcript summaries for specific help desks.

What do you pay for with each AI support agent?

Compare the billable event before the price. The same conversation can count as a resolution, an outcome, a session, an engaged conversation or a few credits, depending on the vendor, and some vendors bill a handoff while others bill nothing when a person takes over. Four vendors publish no price on the pages we checked. Per-event pricing grows with the conversations the agent handles, while seat- and plan-based pricing grows with team size, so which costs less depends on your volume; for a worked model, see what AI support costs at scale.

Pricing. Checked November 2026.

Agents built into a help desk also need that help desk's seats, which this table leaves out; for seat prices, see our comparison of AI customer support platforms for B2B SaaS.

Escalations cost nothing at Lorikeet, under Decagon's per-resolution model, and at Fin when a default rule hands off. A handoff is billable at Front, and at Fin when a Procedure you configured ends in a handoff. Plain charges its Customer Agent per resolution from the plan's monthly credits (Plain vs Fin), and your connected agents draw on the same monthly allowance (Plain docs).

Which AI support agents answer customers in Slack, Microsoft Teams, Discord and email?

Every one of the 20 answers email. Ten agents document answering customers in Slack. In Microsoft Teams, Fin is the only agent documented answering customers; Maven AGI and IrisGPT position Teams for internal help. Five document Discord.

That gap matters for B2B teams. In 1,929 conversations with 1,165 companies choosing a support tool (June 2025 – October 2026), 85% of the startups in our conversations support customers on two or more channels (694 of 812 with channel data), and 69% have Slack as a support channel (560 of 812). Plain offers Slack and email on every plan, Microsoft Teams on Horizon and up and Discord on Frontier; see multi-channel support.

Agent answers customers vs help-desk channel only. Checked November 2026.

Plain's Discord entry rests on its docs keeping agent replies under Discord's 2,000-character limit (Plain docs); Pylon's Discord entry is a 2025 changelog line that predates its current Support Agent.

Which 20 agents are in this guide?

The 20 are grouped by how they are built. Each entry uses the same fields.

Agents built into a help desk

Plain Customer Agent (Plain). The customer-facing AI agent built into Plain's help desk. Runs on: Plain only. Reads: help centers and added knowledge sources, plus the thread's recent messages and the customer's name and company; not internal notes. Test before launch: Playground, staging-workspace evals and Test mode; no past-ticket replay. Acts and hands off: takes no tool actions; hands off on the same thread with an internal note. You pay for: a resolution, 100 credits (Plain docs). Consider it if you want the investigating agent kept apart from the customer-facing one, with your own agents on the same queue.

Fin (Fin, formerly Intercom; owned by Salesforce). The company renamed itself from Intercom to Fin on May 12, 2026, and Salesforce completed its acquisition on September 10, 2026 (Fin). Runs on: Intercom, Salesforce, Freshdesk, HubSpot and Zendesk. Reads: enabled knowledge sources, including internal articles customers can't see, plus user, company and connector data; not past conversations. Test before launch: preview, Evals and single-conversation replay. Acts and hands off: Procedures and Data Connectors, with a staff-approval step. You pay for: $0.99 per outcome. Consider it if you want one agent across several help desks, including Slack, Teams and Discord.

Zendesk AI agents (Zendesk). Zendesk's AI agents, built in the AI agents workspace (Zendesk). Runs on: Zendesk. Reads: Zendesk help centers and connected knowledge sources, plus CRM data through actions; past tickets are not a knowledge source. Test before launch: test widget and a conversation simulator in early access. Acts and hands off: configured actions and action flows; MCP tools generally available since August 10, 2026. You pay for: automated resolutions beyond those included, at $1.50 committed or $2.00 pay-as-you-go. Consider it if you run Zendesk Suite or Support, where 5 or 10 automated resolutions per agent a month are included.

Agentforce Service Agent and Help Agent (Salesforce). Salesforce's two customer-facing service agents; Help Agent is the prepackaged one (Salesforce). Runs on: Salesforce Service Cloud. Reads: a Data Library of knowledge articles, files and web sources, plus records the agent's user can access. Test before launch: Builder preview and Testing Center. Acts and hands off: agent actions with an optional customer-confirmation setting; hands off through Omni-Channel with conversation history. You pay for: $2 per conversation or Flex Credits; Help Agent $2 per resolution. Consider it if you want your own bot routed alongside Agentforce agents through Omni-Channel.

Freddy AI Agent (Freshworks). Freshworks' customer-facing agent for Freshdesk, built in AI Agent Studio (Freshdesk). Runs on: Freshdesk and Freshdesk Omni. Reads: solution articles, uploaded files and public URLs, plus contact properties and API data; the email agent reads the full thread. Test before launch: preview, a bulk Test tab, and, for email, ticket replay and shadow mode. Acts and hands off: Agentic Workflows and API actions, with customer confirmation for high-impact actions. You pay for: sessions, $49 per 100 after the first 500. Consider it if you run Freshdesk and want shadow mode on email.

HubSpot Customer Agent (HubSpot). HubSpot's agent that answers from your content and takes configured actions (HubSpot). Runs on: HubSpot. Reads: knowledge base, synced web pages, files and Q&A generated from resolved tickets, plus CRM properties you grant. Test before launch: test preview with Testing Insights and batch testing. Acts and hands off: up to 20 actions with an identity check each; hands off to HubSpot's inbox or help desk. You pay for: 50 HubSpot Credits per resolution, at $0.010 per credit. Consider it if you want each action gated by an email match or verification on the HubSpot contact.

Kustomer Concierge (Kustomer). Kustomer's customer-facing AI, which responds, follows procedures and uses connected tools (Kustomer). Runs on: Kustomer. Reads: the Kustomer knowledge base and added web sources, plus customer, company and conversation records through its tools. Test before launch: Test Console and AI Evaluations. Acts and hands off: tools in Procedures, including MCP, with a passcode check; routes to a Kustomer queue. You pay for: $0.75 per engaged conversation, an add-on to Enterprise and Ultimate. Consider it if you want the agent on Slack Connect conversations a rep opens, as well as email and chat.

Scout (Help Scout). Scout is part of Help Scout's Envoy and answers customers across Beacon and email, per Help Scout's docs (Help Scout). Runs on: Help Scout. Reads: Docs sites, selected web pages and text sources; no customer records documented. Test before launch: Scout > Test in chat or on chosen email conversations. Acts and hands off: answers from knowledge sources; no actions documented; email escalations go to your Inbox. You pay for: $0.75 per resolution. Consider it if you're new to Help Scout: new accounts get 3 months of free, unlimited resolutions.

Pylon Support Agent (Pylon). Pylon's customer-facing agent, which answers from approved knowledge and escalates with account context (Pylon). Runs on: Pylon. Reads: approved knowledge; Pylon's pages disagree on account data: the product page says it loads account context, while older help articles say its AI agents don't reference account fields. Test before launch: Custom Chat, backtesting on past issues and QA mode. Acts and hands off: Skills and Connectors, with a per-agent tool allowlist; escalations stay on the Pylon issue. You pay for: credits per run, $0.30 each pay-as-you-go. Consider it if you want a team-facing Assist Agent kept apart from the customer-facing one.

Front Autopilot (Front). Front's AI agents, which own conversations in your inbox through normal assignment (Front). Runs on: Front. Reads: selected knowledge base articles, websites, PDFs and Playbooks, plus third-party data through Playbooks and MCP Connectors. Test before launch: Resolve page, Compass Simulations from your conversations and a drafting-only agent. Acts and hands off: Playbooks and MCP Connectors, each action off by default; stays in the Front conversation. You pay for: $0.39 per handoff or $0.89 per resolution. Consider it if you also want External Agents on the same inbox.

AI agent companies

Decagon (Decagon). An AI agent platform whose agent behaviour is written as Agent Operating Procedures (Decagon). Runs on: Zendesk, Intercom and Salesforce, among others. Reads: help centers, internal docs and ticketing systems, plus synced customer data and API lookups. Test before launch: Simulations, Message Replays and test cases from your transcripts. Acts and hands off: AOPs with tools, including MCP; hands off with a summary. You pay for: per conversation or per resolution; no published price. Consider it if you want test cases generated from conversations where your agents struggled.

Sierra (Sierra). A conversational AI platform whose agents are built in Agent Studio or the Agent SDK (Sierra). Runs on: its own platform; no help desk named. Reads: Help Center content, FAQs and policies, plus account history from systems of record. Test before launch: Preview Chat, simulations, regression tests and Voice Sims. Acts and hands off: agent actions, MCP tools and publishing into ChatGPT; hands off with a summary. You pay for: outcomes; no published price. Consider it if you want one agent across chat, voice and ChatGPT.

Ada (Ada). An AI agent platform aimed at large companies in regulated, high-volume industries (Ada docs). Ada and Dialpad announced a combination on November 4, 2026, expected to close around November 12. Runs on: Zendesk, Salesforce, Kustomer and other help desks, and contact centres. Reads: synced knowledge bases, imported websites and Ada articles, plus your systems through API tools. Test before launch: Interactive Testing and Simulations. Acts and hands off: Ada Computer tools with an ASK step. You pay for: conversations or automated resolutions; no published price. Consider it if you want simulations that cover voice as well as chat and email.

Forethought Solve (Zendesk). Solve is the customer-facing agent in Forethought AI agents by Zendesk; Zendesk completed the acquisition on March 26, 2026 (Forethought). Runs on: many help desks, with or without Zendesk. Reads: knowledge base articles and Autoflows, trained on historical support data, with connections to your CRM. Test before launch: Test Suite. Acts and hands off: Autoflows and custom actions; Triage routes the handoff to your help desk. You pay for: an outcome-based charge plus a platform fee; quote only. Consider it if you want Zendesk-owned AI without moving to Zendesk.

Lorikeet Concierge (Lorikeet). An AI agent platform aimed at complex and regulated businesses (Lorikeet). Runs on: seven help desks, including Zendesk, Intercom and Salesforce, or with no help desk. Reads: help docs, SOPs and policies, plus ticket history and customer details in your help desk. Test before launch: simulations from your imported tickets, and internal-note rules per channel. Acts and hands off: actions via APIs, webhooks and MCP; hands over inside your help desk. You pay for: Start $2,100 a month plus $0.99 per chat, email or SMS resolution; escalations free. Consider it if you want simulations built from your own imported tickets, with tools mocked.

Maven AGI (Maven AGI). An enterprise AI agent platform for support across chat, email, voice and web (Maven AGI). Runs on: Zendesk, Salesforce, Intercom, Freshdesk and Genesys. Reads: help center articles, historical tickets and internal docs, plus account data through Actions. Test before launch: simulation and regression testing in Agent Designer. Acts and hands off: Actions in external systems; hands off with full context. You pay for: resolutions; Maven's blog cites a $25,000 minimum for 12 months. Consider it if your handoffs go to Genesys as well as a help desk.

eesel AI (eesel). An AI agent that works tickets inside the help desk your team already uses, or in its own chat (eesel docs). Runs on: 11 help desks, including Zendesk, Freshdesk and Intercom. Reads: help center, macros and past tickets, plus the customer, company and other tickets through help-desk actions. Test before launch: simulation over resolved past tickets, then draft mode. Acts and hands off: actions with On / Ask first / Off per action. You pay for: credits, from $299 a month. Consider it if you want per-action staff approval inside your current help desk.

IrisGPT (IrisAgent). The customer-answering agent in IrisAgent's support-automation platform (IrisAgent docs). Runs on: Zendesk, Salesforce, Intercom, Freshworks, HubSpot, Zoho Desk and Gorgias. Reads: knowledge and past tickets, plus live data from your APIs mid-conversation. Test before launch: chatbot tester, draft modes, and a run on a sample of past conversations set up by IrisAgent. Acts and hands off: Smart Operating Procedures that call APIs; hands off with full history. You pay for: Standard from $500 a month. Consider it if you want an overlay you can start on a $0 tier.

Agent builders

Botpress (Botpress). A builder with Studio, the ADK framework and Desk, its own help desk (Botpress). Runs on: Desk, Zendesk or Intercom. Reads: Studio knowledge bases and tables, plus billing, CRM or order data once connected. Test before launch: Studio Emulator and ADK Evals. Acts and hands off: Studio Actions, ADK Tools and Desk tools, with Always Allow / Needs Approval / Never Allow in Desk. You pay for: conversations; Plus $150 a month billed annually. Consider it if you want to build the agent and need Discord or Slack as customer channels.

Chatbase (Chatbase). A builder for agents trained on your own business data (Chatbase docs). Runs on: its own Helpdesk, or hands off to 9 other help desks. Reads: help center, files, links and Q&As, plus contact data after identity verification and data from Actions. Test before launch: Playground and Tests. Acts and hands off: Actions such as Stripe, Shopify and custom APIs; escalations create a ticket. You pay for: message credits; Hobby $32 a month billed yearly. Consider it if you want written test scenarios graded by an AI judge before launch.

How should you choose a customer support AI agent?

Work through these steps in order.

1. Decide what each agent can reach. If investigation needs code, logs or internal notes, look for a team-facing agent with no customer channel: Plain, Pylon and Kustomer document one. If you want to run your own agent on the same queue, Plain, Zendesk, Agentforce, Pylon and Front name the feature. Plain and Pylon document all three pieces.

2. Put per-action controls on the agent that acts. Staff approval is documented at Fin, eesel AI and Botpress Desk, and for Plain's Sidekick and connected agents. Customer confirmation is documented at Agentforce, Freddy and Ada; HubSpot and Kustomer check identity.

3. Test on your own traffic. Bulk replay of past tickets is documented at Freddy (email), Pylon, Front, Lorikeet and eesel AI. A no-send mode on live conversations is documented at Plain, Freddy (email), Pylon, Front, Lorikeet, eesel AI and IrisGPT.

4. Compare the billable event. Escalations are free at Lorikeet and under Decagon's per-resolution model; a handoff is billable at Front, and at Fin when a configured Procedure hands off. Decagon, Sierra, Ada and Forethought publish no price.

5. Check your channels. Use the channel table to confirm the agent itself answers where your customers are.

Most buyers compare more than one option: among the companies we spoke with, about half were weighing two or more vendors (253 of 482 startups that named a vendor besides Plain and their incumbent).

FAQ

What is the best AI customer support agent?

There is no single best AI customer support agent; the right one depends on how it is built and what you already run. Agents built into a help desk, such as Fin, Zendesk AI agents, Agentforce and Plain's Customer Agent, suit teams staying on that help desk. AI agent companies such as Decagon, Sierra, Ada and Lorikeet layer over your existing help desk. Builders such as Botpress and Chatbase suit teams that want to build the agent themselves.

Can you test an AI support agent without customers seeing it?

Yes. Most AI support agents offer a private preview chat or simulations, and some run on live conversations without sending anything. As of November 2026, Plain, Freddy (email), Pylon, Front, Lorikeet, eesel AI and IrisGPT document a no-send mode on live traffic, and Freddy (email), Pylon, Front, Lorikeet and eesel AI document a bulk replay of past tickets.

Can you bring your own AI agent to a customer support platform?

Yes, on some platforms. As of November 2026, Plain (connected agents), Front (External Agents), Pylon (Bring Your Own Agent), Zendesk (Sunshine Conversations bots, and Bridge in early access) and Salesforce Agentforce (Bring Your Own Bot for Messaging) document a named way to run your own agent on the same queue as people. Plan gates and pricing differ by platform.

Should the agent that investigates be separate from the agent that answers customers?

It depends on what the agent can touch. Know what each agent can reach. Read access to code, logs and internal notes is safer in a team-facing agent with no customer channel; actions a customer asks for need per-action controls on the agent that takes them.

Which AI support agents can take actions, not just reply?

Most of them. Fin (Procedures), Zendesk (action flows), Agentforce (agent actions), Freddy (Agentic Workflows), HubSpot (up to 20 actions), Kustomer (tools), Pylon (Skills and Connectors), Front (Playbooks), Decagon (Agent Operating Procedures), Ada (Ada Computer), eesel AI, Lorikeet and others document actions. Plain's Customer Agent answers from knowledge and does not take tool actions; Plain routes action work to Sidekick or connected agents.

What counts as a billable resolution or outcome?

It differs by vendor. Plain counts a resolution when its Customer Agent replied, no teammate was involved and the thread stayed Done for 7 days. Fin bills an outcome, which includes a Procedure that ends in a handoff. Zendesk checks each automated resolution with an LLM, Freddy bills 72-hour email sessions, and Kustomer bills engaged conversations.

What are the alternatives to Fin for B2B support?

It depends on whether you replace the help desk or keep it. If you replace it, agents built into other help desks include Plain's Customer Agent, Zendesk AI agents, Pylon Support Agent and Front Autopilot. If you keep your help desk, including Intercom, agent companies that run on it include Decagon, Lorikeet, eesel AI and Maven AGI. As of November 2026, Fin is priced at $0.99 per outcome and is owned by Salesforce.

What does the human see when an AI support agent hands off?

It depends on the agent. As of November 2026, Plain leaves an internal note with the handoff reason, the search terms it used and the documents it read. Agentforce transfers the conversation history, Decagon and Sierra pass a summary, and Pylon and Front keep the handoff on the same issue or conversation.

Appendix: how we checked

Every vendor fact on this page was read on the vendor's own pages between November 10 and 11, 2026, and each table links the page it relies on. Where a vendor's pages disagreed, the page says so. This page is rechecked monthly against each vendor's current pages.