An AI customer support agent is a chatbot that answers customer questions in your own words, drawn from your help articles, policies and past replies. Done well, it takes the repetitive questions off your team and answers them at any hour. Done badly, it invents a refund policy you never had and your business is left to honour it. This guide explains what these tools actually do, how they are priced, how to set one up responsibly and how to tell whether it is working.

What an AI support agent automates, and what it hands off

Most tools in this category do three jobs. They answer questions from a knowledge source you supply, they take simple actions through integrations (checking an order status, for example), and they pass the conversation to a person when they cannot help. The third job matters as much as the first two.

The questions an AI agent handles best are frequent, factual and stable: opening hours, delivery times, how to reset a password, what the returns window is. It struggles with anything that needs judgement, an exception to policy, or access to a system it is not connected to. Vendors are open about this. Tidio, for instance, says its Lyro agent is designed to recognize when it cannot handle a query and pass complex or urgent issues to human agents, and that even at a high automation rate it will sometimes need a human operator for edge cases.

So before comparing products, sort your incoming questions into three piles: those an AI could answer from written material alone, those that need a lookup in another system, and those that need a person. The size of the first two piles is the realistic ceiling for automation in your business, whatever a sales page promises.

How pricing works: resolutions, conversations, sessions and seats

The biggest difference between tools is not the chat widget but the billing unit. Read the definition on the pricing page, because the same word can mean different things.

  • Per resolution. You pay only when the AI closes a conversation without a human. Zendesk bills its AI agents in automated resolutions, charged only when an issue is resolved without escalation to a human agent, and says an LLM verifies conversations flagged as resolved. Help Scout's pricing page describes its AI Answers as pay per resolution, counting one resolution per conversation and none when the customer asks for a person.
  • Per automated interaction or ticket. Gorgias charges for its AI Agent per automated interaction, a request fully resolved without a human agent, and its helpdesk is priced by ticket volume rather than by seat. It reclassifies an interaction as an ordinary ticket if an agent replies within 72 hours.
  • Per conversation. Tidio's pricing page counts a Lyro conversation as any customer interaction with at least one reply from the AI agent, whether or not it ends in a resolution. Its top tier offers pay-per-resolution billing as an alternative.
  • Per session. Freshdesk sells Freddy AI Agent in sessions, where an email session is a 72-hour window from the customer's first email, however many AI replies it contains. Its agent-assist Copilot is a separate per-agent add-on.

Per-resolution billing ties cost to results, but you are trusting the vendor's definition of "resolved". Per-conversation billing is more predictable but charges for conversations the AI started and a human finished. Seat-based helpdesk plans still apply on top of most of these, because someone has to handle the hand-offs.

A simple way to estimate cost

  1. Count last month's customer conversations across chat and email.
  2. Estimate the share that falls in your "answerable from written material" pile. Be conservative.
  3. Multiply by the vendor's unit price for the unit it actually bills (resolution, conversation, session or interaction), then add the helpdesk seats or ticket allowance you still need.
  4. Repeat with a pessimistic and an optimistic share. If the tool only makes sense in the optimistic case, keep looking.

Prices change often, so take unit prices from the vendor's own pricing page on the day you decide, and check whether a spending cap is available. Help Scout, for one, lets customers set monthly spending caps for AI Answers.

Training it on your help content

An AI support agent is only as good as what it is given to read. Most tools let you point them at your help centre or website, add question and answer pairs by hand, or import files. Tidio describes four routes for Lyro: scanning website URLs, adding Q&As manually, uploading a file on its Plus plan, and suggestions drawn from past conversations.

Practical preparation pays off more than any setting:

  • Fix contradictions first. If two pages give different returns windows, the AI may quote either.
  • Write down the unwritten. Policies that live only in your team's heads cannot be answered.
  • Exclude what should not be public. Do not feed it internal notes, pricing exceptions or anything containing customer data.
  • Review generated answers. Where a tool turns web pages into Q&A pairs, read them before going live.
  • Keep it current. Every policy change needs a matching knowledge update, or the AI keeps quoting the old rule.

Accuracy, hallucination and who is responsible

Large language models can produce fluent answers that are wrong. In customer service the risk is not embarrassment but liability. In Moffatt v Air Canada, a Canadian tribunal held the airline liable for negligent misrepresentation after its website chatbot told a customer he could claim a bereavement fare after travel, contrary to the airline's own policy page. The tribunal rejected the argument that the chatbot was responsible for its own actions, calling it "a remarkable submission" and treating the chatbot as part of the company's website. Laws differ between countries, but the practical lesson travels: assume you own what your bot says.

To reduce the risk:

  • Restrict the agent to your own knowledge sources where the tool allows it, rather than general web knowledge.
  • Route money matters (refunds, compensation, cancellations, pricing exceptions) to a person, or to a fixed scripted flow.
  • Test with the awkward questions your team actually receives, not just the easy ones, before switching it on for everyone.
  • Read a sample of AI conversations every week, especially in the first months.

Escalation to a human

A good hand-off is the difference between an AI that helps and one that traps people. Check that customers can reach a person by asking for one, that the human agent sees the full conversation so the customer does not repeat themselves, and that outside staffed hours the AI says honestly when someone will reply. Check, too, how escalations affect your bill: under per-resolution plans an escalated conversation is normally not charged as a resolution, while under per-conversation plans it usually is.

Set clear triggers for escalation, such as a customer expressing frustration, mentioning a complaint or legal matter, or asking the same question twice. Vulnerable customers and anything involving health, safety or money should reach a person quickly.

Data protection and telling customers it is AI

Support conversations contain personal data: names, emails, order details and sometimes more. Using an AI tool to process them is covered by data protection law. In the UK, the ICO's guidance on AI and data protection covers accountability (including data protection impact assessments), transparency, lawfulness, accuracy, fairness, and individuals' rights, including the rules on solely automated decisions. Its chapter on transparency in AI says privacy information should explain your purposes, retention periods and who you share the data with, provided at the time you collect the data.

In the EU, Article 50 of the AI Act requires that AI systems intended to interact directly with people are designed so that people are informed they are interacting with an AI system, unless that is obvious to a reasonably informed person. The European Commission's overview of the AI Act puts it plainly: people using chatbots should be made aware they are interacting with a machine. Check the Commission's page for when each obligation applies.

Practical steps: label the bot clearly as AI, update your privacy notice, check where the vendor stores and processes data and whether it uses your conversations to train its models, sign its data processing agreement, and decide how long transcripts are kept. None of this is legal advice; if you are unsure, read the regulator's guidance directly or ask a qualified adviser.

Measuring whether it works

Vendor dashboards usually lead with an automation or resolution rate. Treat it as a starting point, since each vendor defines resolution its own way. Track a few measures of your own before and after launch:

  • Genuine resolution: sample conversations marked resolved and check whether the customer came back about the same issue.
  • Escalation rate and reasons: the questions that escalate most are your next knowledge articles.
  • Customer satisfaction for AI-handled conversations compared with human-handled ones.
  • Time to first human reply for escalated cases, which should fall as the queue shrinks.
  • Total support cost per conversation, including AI usage and seats.
  • Error log: every wrong answer you find, what caused it and how you fixed the source.

Run a trial on real traffic where possible. Several vendors offer free plans or trials, which is the cheapest way to see how a tool handles your own questions.

Comparing the options

The tools on our ranking take different approaches. Some are built around e-commerce stores, some are full helpdesks with AI added, and some are standalone AI agents that sit on top of an existing helpdesk. Decide first which billing unit suits your volume, which systems the agent must connect to, and how you will hand off to people. Then use our ranking of the best AI customer support agents to compare free plans, trials and key features side by side.