An AI support agent is software that answers your customers’ questions on its own, in chat or email, using your help centre, docs and past conversations as its source. When it can answer, it replies in seconds. When it can’t, it hands the conversation to a person with everything it learned attached. The good ones can also take actions, like looking up an account or filing a bug, not just write text.
Here’s how AI support agents work, what they’re good and bad at, how they’re priced, and how to set one up without upsetting customers.
How an AI support agent works
Under the hood, most AI support agents follow the same four steps:
- Read the question. The customer writes in chat or email. The agent works out what they’re asking, in whatever language they wrote.
- Find the answer in your content. It searches your help centre, docs and previous solved conversations for the passages that answer it. This is called retrieval, and it’s what keeps the agent from making things up.
- Write a reply. A large language model writes an answer in plain words from those passages, ideally with links to the sources.
- Act or hand over. If the question needs an action, it does it, within limits you set. If it isn’t confident, or the customer asks for a person, it passes the conversation to your team with a summary.
AI agent vs chatbot vs copilot
| What it does | Who it talks to | |
|---|---|---|
| Old-style chatbot | Follows a decision tree you built. Breaks on anything you didn’t predict. | Customers |
| AI support agent | Understands free-form questions, answers from your content, can take actions and hand over. | Customers |
| AI copilot for agents | Drafts replies, summarises threads and finds answers for your team to send. | Your support team |
Many products now offer both: an agent that answers customers, and a helper inside the inbox for your team.
What AI support agents are good at
- The same questions, every week. Password resets, billing dates, “does it integrate with X?”, “where is the export button?”. For most SaaS products these are a large share of the inbox.
- Speed. Answers in seconds, at 3am, in the customer’s language.
- Showing you the gaps. Every question it can’t answer is an article you should write.
What they’re bad at
- Anything your docs don’t cover. An agent that answers only from your content will say it doesn’t know. One that answers from general knowledge will guess, and sometimes guess wrong.
- Judgement calls. Refunds, exceptions, angry customers and anything legal belong with a person.
- Bugs. “It’s broken” needs someone to look at what happened. The best setups attach a session replay and logs, so whoever picks it up can see the problem.
How AI support agents are priced
There are three common models, and they behave very differently as you grow:
- Per resolution or outcome. You pay each time the AI resolves a conversation, typically somewhere around $0.49 to $2.00 each across the well-known tools. Simple to start; grows with your volume.
- Per seat or session bundles. An AI add-on per agent, or a bundle of AI conversations included in a plan.
- Bring your own key. You connect your own account with an AI provider and pay them directly for what the model uses. Usually the cheapest at volume.
We compared the pricing of eight tools in AI customer support tools for small SaaS teams.
How to set one up without upsetting customers
- Fix your help centre first. The agent can only be as good as what it reads. Write down the answers to your ten most common questions.
- Start with the agent answering first, and handing over easily. Make “talk to a person” one click away from the start.
- Restrict it to your own content. Better “I don’t know, let me get someone” than a confident wrong answer.
- Read its conversations for the first two weeks. You’ll learn where it struggles and which articles are missing.
- Watch three numbers: how many conversations it resolves, how often it hands over, and whether customers who talked to it come back unhappy.
Is it safe to let AI answer customers?
For questions your docs answer, yes, as long as it cites its sources and hands over when unsure. The risk is an agent that improvises. Look for one that answers only from your content, shows where each answer came from, keeps a log of everything it did, and asks a person before doing anything that changes data or money.
How Rectify does it
Rectify’s AI teammate, Quanta, answers first in a shared live-chat inbox. It answers from your help centre and knowledge base with links to its sources, and when it can’t help it brings in your team, alerting you by push and email. You choose the mode: Quanta first with handover, Quanta only, or your team only. It runs on your own AI key, so there’s no fee per answer, and the same teammate also triages bug reports, keeps your roadmap moving and writes your changelog, with your approval on anything sensitive. See the support inbox and Quanta.
Questions founders ask
What’s the difference between an AI support agent and a chatbot?
A chatbot follows a script you wrote. An AI support agent understands the question in the customer’s own words, finds the answer in your content and can take actions or hand over. It handles questions you never predicted; a chatbot doesn’t.
Will an AI support agent replace my support team?
For a small SaaS, it usually replaces the repetitive half of the work, not the people. Your team spends its time on the conversations that need judgement, and nobody has to answer “how do I reset my password?” again.
How long does it take to set one up?
If your help centre is in good shape, an afternoon. Most of the work is making sure your content answers the questions customers actually ask.
Give support its first hire
Let the easy questions answer themselves and keep the hard ones for you. Hire Rectify free and Quanta takes the first shift.
