An AI support agent is software that answers customer support questions automatically: not by routing tickets or suggesting macros, but by reading your documentation and generating real answers. Used well, it resolves the questions your team answers fifty times a day, frees up human time for the harder ones, and stays available at 3 a.m. without a staffing cost.

Used carelessly, it confidently tells customers the wrong thing and erodes trust faster than no AI at all.

The gap between those two outcomes comes down to one thing: grounding.

What an AI support agent actually is

At its core, an AI support agent combines a large language model with a retrieval layer over your own content: your help center, knowledge base, product documentation, and increasingly, live account data like orders, bookings, or subscription status.

When a customer asks a question, the agent:

Retrieves relevant content from your knowledge base
Generates a response based on that content, not general internet knowledge
(Ideally) cites the sources it used, so the customer and your team can verify the answer
Escalates to a human agent when the question is outside the scope of its knowledge

This is fundamentally different from a generic chatbot. A generic chatbot is trained on internet data and will try to answer anything, often by sounding plausible while being factually wrong. A properly grounded AI support agent works only from what you’ve told it, and should know when to say “I don’t know.”

The distinction matters because customers assume you stand behind every answer your support channel gives, regardless of whether a human or an AI wrote it.

The #1 risk: hallucination

Hallucination is the term for when an AI model generates a confident answer that isn’t supported by the source material, or that’s simply invented. In AI customer support, it looks like:

Quoting a refund policy that doesn’t exist
Citing a feature your product doesn’t have
Giving the wrong steps for a settings change
Confidently answering a billing question with made-up numbers

Teams that deploy AI support agents and then turn them off usually cite this as the reason. One wrong answer, especially a visible one, creates more support work than it saved and damages the customer relationship in the process.

The problem isn’t the technology itself; it’s how the agent is configured and constrained.

How grounding prevents hallucination

Grounding means the agent’s answers are anchored to specific documents in your knowledge base, and any claim in the response maps back to a retrievable source.

A well-grounded agent:

Only uses your content: not the open internet, not training data from other companies’ docs. If it isn’t in your knowledge base, the agent doesn’t know it.
Cites its sources: each answer links back to the specific help article or document it drew from. Customers can verify; your team can audit when something is wrong.
Stays in scope: if a question doesn’t match anything in the knowledge base above a confidence threshold, it doesn’t guess. It flags the gap.

A second layer of protection is account data grounding. Many support questions aren’t documentation questions at all: “where is my order?”, “what plan am I on?”, “when is my next booking?” An agent that can only read your docs will escalate all of these. An agent that can also securely read live account data (via an API or an MCP connection) can answer them accurately, because it’s reading the actual source of truth rather than inferring from general knowledge.

Abstain and escalate: the safety net that matters most

Even a well-grounded agent will encounter questions it can’t reliably answer. The correct behavior in those cases is to abstain: to explicitly not answer, and hand off to a human instead.

An agent that always generates a response, even when uncertain, will hallucinate. An agent with a calibrated confidence threshold will recognize when it’s outside its reliable range and say so.

This is the difference between:

“Your order shipped on June 3rd.” Confidently wrong, with no way for the customer to know that.
“I couldn’t find a reliable answer to this in our documentation. I’m passing this to our team.” An honest escalation that preserves trust.

Suggest mode is an underrated starting point. Instead of sending AI-generated answers directly to customers, the AI drafts a response for a human agent to review and send. This builds confidence in the system, catches mistakes before they reach customers, and lets you verify accuracy at scale before switching to fully automated responses. Many teams start here and graduate to autonomous mode over time.

Understanding AI support pricing

Two pricing models dominate the market:

Per-seat pricing bundles the AI into a platform fee paid regardless of how much the AI actually resolves. If the AI handles very little, you’re paying for unused capacity. If it handles a lot, the unit economics improve.

Per-resolution pricing charges for each query the AI successfully resolves. Intercom’s Fin charges approximately $0.99 per resolution (as of 2026; verify current pricing on their site). The model aligns costs with value, but the numbers add up: 2,000 AI resolutions a month is roughly $2,000 with that pricing.

This model is sometimes called outcome-based pricing, and it’s increasingly common in AI customer support. Worth modeling your expected resolution volume before committing: the bill can look very different at different scales.

Some platforms also offer a free tier while you’re early-stage, which means you can validate whether AI support actually works for your specific support content before spending anything.

How Convot’s Cove approaches this

Cove is Convot’s AI support agent. It was built around the grounding-first principle from day one:

Grounded in your docs and account data: Cove draws from your help center and can connect to live account data via a secure API or MCP server. It doesn’t guess from general knowledge.
Source citations on every answer: each response links back to the specific article or data source it used. Customers see exactly where the answer came from.
Abstain and escalate logic built in: when Cove isn’t confident, it flags the conversation for a human agent rather than generating an uncertain answer.
Suggest mode available: start with a human reviewing every AI draft before it’s sent, then graduate to autonomous when you’re ready.
$0.20 per resolution: compared to Fin’s ~$0.99, that’s roughly a fifth of the cost at similar volume. And Convot is free under $1k MRR, no credit card required.

If you’re evaluating AI support agents and concerned about hallucination risk, the most important question to ask any vendor is: “What does your agent do when it doesn’t know the answer?” The answer tells you more than any feature list.

See also: best Intercom alternatives if you’re comparing Fin to other options, and try Convot free to see Cove in action on your own help center.

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