What is an AI support agent? (And how to keep it from hallucinating)

An AI support agent is software that answers customer support questions on its own by finding the answer in your documentation, and sometimes your customers’ account data, and writing a reply from it. A good one answers routine questions at any hour, shows where each answer came from, and hands the conversation to a person when it is not sure. A careless one gives confident wrong answers, which is why grounding and the ability to abstain matter more than any other feature.

This guide explains how an AI support agent works, how it differs from a chatbot, why it hallucinates, how to prevent that, how pricing works and what to ask before you choose one.

How does an AI support agent work?

An AI support agent combines a large language model with a search layer over your own content. When a customer asks a question, it goes through four steps:

1. Retrieve. It searches your help center, documents and saved answers for the passages most relevant to the question.
2. Answer. It writes a reply based on those passages, not on general internet knowledge.
3. Cite. It can show the customer, and your team, which article or document it used.
4. Escalate. When the question is outside what it can answer reliably, it hands the conversation to a person.

Many agents can also read live account data, such as orders, bookings or subscription status, through an API or an MCP connection, so they can answer “what plan am I on?” from the real record.

AI support agent vs chatbot: what is the difference?

An AI support agent writes answers from your own content and stops when it does not know, while a traditional chatbot follows scripted decision trees or answers from general knowledge. The difference shows up in the questions each can handle and how each fails.

Customers will reasonably treat every answer your support channel gives as yours, whether a person or an AI wrote it. That is why the “when it does not know” row matters most.

Why do AI support agents hallucinate?

AI support agents hallucinate when the language model fills a gap with plausible text instead of admitting it has no answer. A model is built to produce fluent, likely-sounding replies, so without constraints it will answer a question its sources do not cover. In customer support, that looks like:

• Quoting a refund policy that does not exist.
• Describing a feature the product does not have.
• Giving the wrong steps for a settings change.
• Answering a billing question with made-up numbers.

One visible wrong answer can create more support work than it saves, and it can damage trust with that customer. The fix is not a better model alone; it is how the agent is constrained.

How do you prevent AI hallucination in customer support?

Prevent AI hallucination in customer support with four controls: ground every answer in your own content, show the source, set a confidence threshold below which the agent hands off to a person, and fence off sensitive topics such as billing disputes and legal questions. The first control, grounding, does most of the work.

What is grounding in AI customer support?

Grounding in AI customer support means the AI writes each answer only from sources it retrieved, such as your help center, uploaded files and live account data, and declines to answer when none match. A grounded answer can be traced to a named article; an ungrounded one comes from the model’s general training. A well-grounded agent:

• Uses only your content, not the open internet or other companies’ documentation.
• Shows its sources, so customers can check the answer and your team can audit mistakes.
• Stays in scope, declining to answer when nothing in your content matches well enough.

Grounding quality depends on your content: an agent can only be as accurate as the articles it reads, so a clear help center, with one question per article and the answer near the top, is the most effective thing you can do for AI answer quality. The guide to building a help center covers how.

Account data grounding is the second layer. Questions like “where is my order?” or “when is my next booking?” are not documentation questions. An agent that can securely read the real record answers them correctly instead of escalating or guessing.

What should an AI support agent do when it does not know?

An AI support agent should abstain when it does not know: say it cannot find a reliable answer and hand the conversation to a person, with the context, so the customer does not repeat themselves. This is the safety net that matters most.

The difference looks like this, in an illustrative example:

• “Your order shipped on June 3rd.” Confident, wrong, and the customer has no way to know.
• “I couldn’t find a reliable answer to this in our help center, so I’m passing you to our team, who will reply here.” Honest, and it keeps trust.

Most agents decide this with a confidence threshold: below it, the agent escalates. A stricter threshold means fewer wrong answers and more handoffs; a looser one answers more but risks more mistakes. Good agents also let you list topics that always go to a person, such as billing disputes, cancellations or anything legal.

Should AI support start in suggest mode?

AI support should start in suggest mode, a setting where the AI drafts a reply and a person reviews, edits or rejects it before it reaches the customer. It catches mistakes before customers see them, shows your team how good the answers are, and builds a record of how often drafts are accepted.

Most teams move in stages:

Moving to automatic replies outside working hours first is a low-risk step: customers get an answer at night instead of waiting until morning, and your team reviews the results each day.

Which questions should an AI support agent handle?

An AI support agent should handle questions that have one right answer in your documentation, and leave judgment calls to people. A simple split:

Measure the split once the agent is live. If the agent keeps escalating a type of question your docs should cover, write the missing article. If it answers questions it should not, fence them off.

How do you measure an AI support agent?

Measure an AI support agent by how much it answers, how good the answers are and what it costs:

• Deflection rate = conversations the AI answered without escalating ÷ conversations it handled.
• Answer quality, from customer feedback and from grading its conversations, such as with AI quality grading.
• Escalation reasons, to find gaps in your documentation.
• Cost per resolved conversation, to compare with the time it saves.

Watch quality before deflection. A high deflection rate built on wrong answers is worse than a modest one built on correct answers.

How is AI support priced?

AI customer support is priced either per seat, as a platform fee that includes the AI, or per resolution, as a fee for each conversation the AI resolves. Per seat is easier to budget; per resolution scales with use. For what each vendor charges per resolved conversation, see AI customer service agent pricing compared.

As a worked example, not a quote: Intercom’s Fin is priced from $0.99 per outcome (published price, October 2026), so 2,000 outcomes a month would cost at least $1,980 before other charges. At Convot’s $0.20 per resolved conversation, the same 2,000 would cost $400. Vendors define a resolution differently, so compare definitions as well as prices, and check whether drafting, translation or grading are billed separately.

What does an AI support agent need from your team?

An AI support agent needs three things from your team to work well: good source content, clear rules about what it may answer, and someone who reviews its work each week. None of these take long, and skipping them is a common reason AI answers disappoint.

1. Source content. One help article per common question, answer first, with exact names for settings and screens. Add short hand-written answers for questions where the wording must be exact, such as a refund policy.
2. Rules. A confidence threshold, a list of topics that always go to a person, and a business context that says what the product is and who the customers are.
3. A weekly review. Read a sample of its answers and every escalation, then fix the articles behind the misses. An agent improves as fast as its sources do.

What questions should you ask an AI support vendor?

Ask every vendor what their agent does when it does not know the answer. The reply tells you more than any feature list. Then ask:

1. Where do answers come from? Only your content, or also general knowledge?
2. Can customers and your team see the source of each answer?
3. Can you set the confidence threshold and fence off topics?
4. Is there a suggest mode, and how do you graduate from it?
5. Can it read account data securely, and what can it change?
6. How is a resolution defined and billed?
7. How do you find and fill gaps in its knowledge?

How does Convot’s Cove AI prevent hallucination and price resolutions?

Cove AI is Convot’s AI support agent, built around grounding from the start:

• Grounded in your content. It answers from your help center, web pages, uploaded files and hand-written answers, and can read live account data through your own REST endpoint or MCP server.
• Sources shown. It can show the article or page each answer came from; citations are a setting you control.
• Abstains and escalates. A confidence threshold, a list of topics that always go to a person, and a cap on bot replies per conversation decide when it hands off.
• Suggest mode first. Your team reviews drafts, then moves to automatic replies outside working hours or all the time.
• Priced per resolved conversation. Cove is on Convot’s paid plans, which start at $49 a month, at $0.20 per resolved conversation in which it sent an automatic reply, charged once per conversation. Fin is priced from $0.99 per outcome, and the two define a resolution differently, so compare definitions as well as prices. Drafts your team reviews and sends draw on the AI credits included in your plan instead.

Common questions about AI support agents

Can an AI support agent replace a support team? An AI support agent can answer routine questions that have one right answer, but it should not replace people for bugs, refunds, upset customers or judgment calls. Most small teams use it to answer simple questions at any hour and give people more time for the hard ones.

How long does it take to set up an AI support agent? Setting up an AI support agent takes as long as getting your documentation in order. Connecting the sources is quick; writing clear articles for your most common questions is the real work, and it pays off for customers who read them directly too.

Do customers mind talking to an AI? In our experience, customers mind wrong answers and dead ends more than they mind talking to an AI. An agent that answers correctly, shows its source and hands off cleanly when it cannot help is far better received than one that guesses.

Choosing and running an AI support agent, in seven steps

1. Write clear help articles for your most common questions.
2. Choose an agent that answers only from your content and shows its sources.
3. Set a confidence threshold and fence off sensitive topics.
4. Start in suggest mode and review every draft.
5. Move to automatic replies outside working hours first.
6. Measure answer quality before deflection.
7. Write the missing articles the escalations reveal.

Start free with chat and a help center while you are under $1,000 MRR, and add Cove AI on a paid plan.

About the author

Tarang Agarwal is the founder of Convot. Convot is built by Sidepanda, a small studio that runs several Shopify apps of its own, including Appointo, Depo and Panda Bundle, and supports all of them from one Convot inbox.