The short answer
The best AI for customer support in November 2026 is Fin AI Agent, from Fin, the company formerly called Intercom and owned by Salesforce since 10 September 2026. It works with Intercom and helpdesks such as Salesforce, HubSpot and Freshworks. It answers on chat, email, SMS, WhatsApp and phone, takes actions through Procedures, and lets you test changes with simulations first. Zapier's AI by Zapier is the runner-up for adding AI to an existing helpdesk, and Claude's Free plan is the best free help for drafting replies.
| Our pick | Fin AI Agent (included in all Intercom plans, also works with other helpdesks) |
|---|---|
| Runner-up | Zapier with AI by Zapier and support agent templates |
| Best free | Claude Free plan for drafting replies and articles |
| Fin pricing model | Per outcome, 14-day free trial without a card |
| Channels Fin covers | Live chat, email, SMS, WhatsApp, social and phone |
| Watch out for | Thin help content, and actions taken without approval |
The best AI for customer support in November 2026
AI in customer support does two different jobs. One is answering customers directly, around the clock, and handing over to a person when it should. The other is helping your agents: drafting replies, summarising long threads, writing help-centre articles and spotting why tickets come in. The first needs a dedicated AI agent connected to your helpdesk. The second can start with a general assistant you may already have. Our picks cover both, based on what each maker documents today.
| Tool | Pick it if | The catch | |
|---|---|---|---|
| Our pick | Fin AI Agent | You want AI to resolve customer questions across chat, email and voice | You pay per outcome, so cost rises with volume |
| Runner-up | Zapier (AI by Zapier) | You want AI triage and drafted replies inside the helpdesk you already have | Every AI step and action costs tasks |
| Best free | Claude (Free plan) | Your agents need help writing replies, macros and help articles | It is not connected to your helpdesk, and free chats can be used for training |
Our pick: Fin AI Agent
Why. Fin is a customer-facing AI agent made by Fin, the company formerly called Intercom, which Salesforce finished buying on 10 September 2026. Fin is included in all Intercom plans, but it does not require Intercom. Its pricing page lists Fin with other helpdesks including Salesforce, HubSpot, Freshworks, Dixa, Front, Zoho, Sprinklr and Gorgias, and its homepage says it works with Salesforce, Zendesk, HubSpot and Freshdesk. Fin answers on live chat, email, SMS, WhatsApp, social channels and phone. Beyond answering questions it can take actions such as updating accounts and processing refunds through an API, Data Connectors or MCP, and multi-step workflows run through what Fin calls Procedures. Before a change goes live you can test it with simulations, regression tests and manual inspection, which is the feature support leads should care about most.
Intercom, now renamed Fin, reports on the Fin homepage an average resolution rate of 76% across more than 12,000 customers. Treat that as the maker's number, not a promise for your queue: your rate depends on how good your help content and procedures are.
Who it suits. Support teams with steady volume across several channels, ecommerce and software companies with many repetitive questions, and teams that want automated answers with a clean handover to people.
The catch. Fin is billed per outcome, such as a resolution or a Procedure handoff, with a minimum of 50 outcomes a month when you use it with another helpdesk, so a busy month costs more. Conversations passed to your team without an outcome are not charged. Check the current terms on the Fin pricing page. A 14-day free trial needs no credit card. Fin's homepage names Zendesk among the helpdesks it works with, but its pricing page does not list it, so Zendesk teams should confirm terms before they plan around it. Since the Salesforce purchase, Fin says it keeps working with the helpdesks customers already use.
Runner-up: Zapier with AI by Zapier
Why. Many teams cannot switch helpdesk, and do not need to. Zapier connects more than 9,000 apps, and since October 2026 its AI by Zapier step can reason, call tools and act inside an automation. Zapier offers ready-made Support Email Agent and Zendesk Ticket Agent templates, so a new ticket can be classified, summarised, given a drafted reply and routed to the right queue. A per-tool toggle can require human approval before the AI takes an action, and AI Guardrails by Zapier can detect personal data inside workflows. That combination lets you add AI step by step without letting it email customers on its own.
Who it suits. Small and mid-sized support teams on Zendesk, Gmail or a shared inbox, operations people who already build Zaps, and anyone who wants to test AI triage before buying an AI agent.
The catch. Everything costs tasks: AI steps bill in tasks with higher multipliers for stronger models, so heavy use pushes you up a plan. Some knowledge sources, including Zendesk, HubSpot and Zoom, are not yet supported inside AI by Zapier. The free plan allows only two-step Zaps and 100 tasks a month. Only Enterprise is opted out of training automatically; other plans opt out through a form. See our Zapier guide and Zapier vs Make.
Best free option: Claude
Why. For the agent-assist side of support, a good writer is worth more than a chatbot widget. Claude from Anthropic, on its Free plan with Claude Sonnet and Haiku, is strong at reading a long, angry thread and drafting a calm reply that answers every point, and at turning a messy internal fix into a clear help-centre article. Projects (up to five on Free) can hold your tone guide and policies, so replies stay consistent.
Who it suits. Small teams and founders handling support themselves, and agents who want help with tone and clarity.
The catch. Claude Free is not connected to your helpdesk, so you copy and paste. Limits reset every 5 hours with a weekly cap. Free chats can be used for training unless you turn off Help improve our AI models under Settings, Privacy, and you should remove names, emails and order numbers before pasting customer messages. For customer data at scale, use Claude Team, where content is not used for training by default. ChatGPT Free works similarly; compare them in ChatGPT vs Claude.
How to choose the right AI for customer support
- Should AI talk to customers, or help your agents? Talking to customers needs an AI agent such as Fin connected to your helpdesk and knowledge. Helping agents can start with Claude or ChatGPT and grow into Zapier automations.
- How good is your help content? An AI agent can only answer from what you give it. If your help centre is thin or out of date, fix the top 20 articles first, or the AI will hand over most conversations or answer badly.
- What can the AI do without asking? Decide which actions need a human (refunds, account changes, anything legal) before you switch anything on. Pick tools with approvals and testing, such as Fin's simulations or Zapier's approval toggle.
A support workflow from ticket to better help centre
Here is the whole loop, end to end, for a team adding AI for the first time.
- Find the repeat questions. Export the last month of tickets without personal data and ask Claude to group them by reason and rank them by volume.
- Write or fix the answers. For the top 20 reasons, draft or rewrite help articles with Claude, then have your best agent check every policy detail.
- Set the rules. Write down what the AI may answer, what it must hand over (complaints, legal, refunds above a limit) and how it should sound.
- Test before customers see it. In Fin, run simulations on real past questions. In Zapier, start with drafted replies that an agent approves.
- Go live on one channel. Start with chat or email, not everything at once, and read a sample of AI conversations every day for the first two weeks.
- Close the loop weekly. Look at handovers and bad answers, and turn each one into a better article or a clearer rule.
The step-by-step version is in how to use AI for customer service, and the role view is in AI for customer support teams. For more ready prompts, see prompts for customer support, and for tool sets by team size, /stacks.
Prompts for customer support you can copy
These work in Claude, ChatGPT or an AI by Zapier step. Remove personal data first and replace everything in square brackets.
Prompts to copy
Calm reply to an angry customer
Here is a customer thread: [paste, with personal data removed]. Draft a reply that acknowledges the problem in one sentence, answers every question they asked in order, says exactly what we will do and by when, and stays within this policy: [policy]. Tone: [tone]. Under 150 words. List anything you could not answer from the policy.
You get: A reply that answers every point without overpromising
Thread summary for handover
Summarise this support thread for a colleague taking it over: the customer's problem, what has been tried, what was promised and when, the customer's mood, and the next action. Five bullet points maximum. [paste thread]
You get: A handover note that saves rereading the thread
Ticket reasons from an export
Here are [number] ticket subjects and first messages with personal data removed: [paste]. Group them into no more than 12 contact reasons, count each, and for the top five suggest whether a help article, a product fix or a process change would prevent them.
You get: A ranked list of why customers contact you
Help-centre article from an internal fix
Turn these internal notes about [issue] into a help-centre article for customers: [paste]. Use a title that matches how customers describe the problem, a two-sentence answer first, numbered steps, and a short section on what to do if it still does not work. Plain English, no internal jargon.
You get: A clear article your AI agent and customers can use
Rules for an AI agent
We are setting up an AI agent for [company] support. Based on this policy document [paste], write the rules: what it may answer on its own, what it must hand to a person, what it must never promise, and how it should sound. Then list ten tricky test questions that would check each rule.
You get: Guardrails and test cases before you go live
Macro set for common questions
Write saved replies for these five common questions: [list]. For each, give a short version for chat (under 50 words) and a full version for email, with [placeholders] for order numbers and names. Keep our tone: [tone].
You get: Ready macros for chat and email
Weekly review of AI answers
Here are this week's AI conversations that were handed over or rated badly: [paste, anonymised]. For each, say why the AI failed (missing content, unclear policy, wrong understanding, should have handed over sooner) and the one change that would fix it.
You get: A short fix list that improves the AI every week