How AI Answers Customer Questions: Grounding, Accuracy, and Cost
AI answers customer questions by retrieving your own documented sources and replying from what it finds, not from open-ended memory. How grounding works, what resolution rates are realistic, the Air Canada ruling, and what it costs.
AI answers a customer question by retrieving the relevant passages from a defined set of your own sources, help articles, policies, past tickets, and then writing a reply grounded in what it found, rather than generating an answer from open-ended memory. The systems worth using resolve the conversation end to end when they are confident, and hand it to a human the moment they are not. That grounding step is the whole ballgame: it is what separates a useful support agent from a confident, plausible-sounding liability.
This guide walks through how that actually works, what resolution rates are realistic versus marketed, why a wrong answer is a legal problem and not just an embarrassing one, when the human should take over, and what the whole thing costs.
Jump to a section:
- How does AI answer a customer question?
- What resolution rates are realistic?
- The hallucination problem and the Air Canada ruling
- When the human should take over
- What does it cost to have AI answer questions?
- How SupportWire answers customer questions
How does AI answer a customer question?#
The popular mental model of a support AI is a chatbot that "knows" your product because it was trained on the internet. That model is wrong, and the difference matters for every point below.
A good support AI answers in two steps. First it retrieves: given the customer's question, it searches a defined set of sources you control and pulls back the passages most likely to contain the answer. Then it generates: it writes a reply grounded in those retrieved passages. The technique has a name, retrieval-augmented generation, or RAG.
IBM's definition of RAG puts it plainly: it "anchors LLMs in specific knowledge backed by factual, authoritative and current data" rather than relying solely on what the model absorbed during training. That anchoring is the point. When a customer asks about your 30-day return window, you do not want the model reciting a generic return policy it saw a million times in training. You want it reading your policy, the one that says 30 days, and answering from that.
Retrieval, not recall
The practical consequence is that a support AI is only as good as the sources it retrieves from. If your refund policy lives in a current, well-written article, the AI can answer refund questions accurately. If that policy is out of date, contradictory, or missing, the AI has nothing solid to ground on, and the quality of its answer falls off a cliff.
This is why "how does the AI resolve a question" and "how good is your documentation" are really the same question. The retrieval step decides everything the generation step can safely say. It is also why raising the share of questions an AI can close is mostly a content problem, a point we go deep on in what is ticket deflection.
What resolution rates are realistic?#
Here is where marketing and reality diverge, so it is worth separating what vendors claim, what independent analysts forecast, and what you should actually plan around.
Start with a vendor claim, labeled as such. Intercom's Fin product site states a 76% average resolution rate across its customer base, "with many seeing over 85%." That is a vendor-stated figure from the company selling the product, on its own marketing site. It is useful as a ceiling that the best deployments reach, not as a number you should expect on day one.
Independent forecasts are more measured. Gartner, in a March 5, 2025 press release, predicted that "by 2029, agentic AI will autonomously resolve 80% of common customer service issues without human intervention, leading to a 30% reduction in operational costs." Read that carefully: it is a forecast for 2029, and it is scoped to common issues, not all of them.
Salesforce's State of Service report gives a nearer-term view from surveyed service organizations. Adoption of AI agents in customer service rose from 39% in 2025 to 66% in 2026. The report expects 50% of service cases to be resolved by AI by 2027, up from 30% in 2025. And 70% of organizations that adopt AI agents report measurable value within 60 days.
The adoption jump is worth sitting with for a second. Two-thirds of surveyed service organizations now run AI agents, up from well under half a year earlier, which means the open question for most teams is no longer whether to let AI answer questions but how well it does so and how much it costs. That reframes the whole evaluation around resolution quality and per-resolution price rather than novelty.
Put those together and the honest planning number is this: half of your common, well-documented questions is a realistic target as the technology and your content mature, not a starting point. Anyone quoting you 80% out of the box is quoting a marketed ceiling. Independent observers regularly point out that real-world resolution rates lag the headline figures vendors advertise, and while we will not put a specific counter-number on that here because we could not source one to a primary study, the direction is well established: plan for the middle of the range, not the top.
The reason the honest number matters so much comes down to what happens to every question the AI does not resolve. It flows to a human at full cost. So the resolution rate is not a vanity metric, it is the single biggest lever on your support economics.
The hallucination problem and the Air Canada ruling#
Grounding lowers the risk of a made-up answer. It does not remove it. IBM is explicit on this in the same reference above: "RAG can reduce the risk of hallucinations, it cannot make a model error-proof." A retrieval system can pull the wrong passage, miss a passage, or the model can still stitch together a confident sentence that the sources do not support.
When that happens in customer service, it is not a harmless quirk. It is a liability, and there is now case law to prove it.
In Moffatt v. Air Canada (2024 BCCRT 149, decided February 14, 2024 by the British Columbia Civil Resolution Tribunal), Air Canada's website chatbot told a passenger he could apply for a bereavement fare retroactively after booking. That was false under Air Canada's actual policy. When the passenger sought the refund the bot had promised, Air Canada refused, and argued in tribunal that the chatbot was "a separate legal entity" responsible for its own statements.
The tribunal rejected that argument outright. It ruled that Air Canada owed a duty of care to users of its chatbot and "did not take reasonable care to ensure its chatbot was accurate," which amounted to negligent misrepresentation. The passenger was awarded $650.88 CAD in damages, plus interest and fees.
The dollar figure is small. The precedent is not. A company is responsible for what its AI tells a customer, exactly as if a human agent had said it. That single ruling is the strongest possible argument for two design choices: ground every answer in current, authoritative sources so the AI has less room to invent, and build a clean escape hatch to a human for everything the AI is not confident about. Which brings us to handoff.
When the human should take over#
The best support AI is not the one that answers the most questions. It is the one that knows the difference between a question it should answer and a question it should not touch.
The clearest primary-source guidance on this comes from vendor documentation. Zendesk's own help-center articles on AI agent design describe escalation built around a few triggers: complexity beyond the procedures the bot has been given, an explicit customer request for a human, and confidence thresholds where the AI escalates rather than guesses. Their recommended flow structures a conversation as greet, verify, diagnose, then resolve or escalate, and collects the customer's details in a structured form before handoff so the human agent does not have to re-ask everything.
That last part is the piece teams underrate. A handoff that dumps a cold conversation on an agent with no context is barely better than no AI at all. The agent has to read the whole thread, re-establish what the customer wants, and effectively start over. A good handoff carries the full conversation, what the customer asked, what the AI tried, what it retrieved, and why it stepped back, so the human picks up mid-stride instead of from zero.
So the rule of thumb for when a human takes over is straightforward:
- The customer asks for a person. Never make them fight the bot. An explicit request is an immediate handoff.
- The AI's confidence is low. If retrieval turned up nothing solid, guessing is the Air Canada failure mode. Escalate instead.
- The request is outside defined procedures. Refund exceptions, account-specific edge cases, anything requiring judgment or authority the AI does not have.
- Sentiment is going wrong. A frustrated customer is a human's job, not a place to keep trying automated replies.
A tool that gets handoff right turns the resolution rate conversation from scary to safe. You are not betting the customer relationship on the AI being right every time. You are letting it handle what it can prove it can handle, and routing the rest to a person with everything they need.
What does it cost to have AI answer questions?#
AI support is priced differently from traditional helpdesks. Instead of paying only per seat, you mostly pay per resolution: a metered fee each time the AI fully closes a conversation on its own. That shift changes how you should compare tools, because the per-resolution price, multiplied across thousands of conversations a month, becomes the largest line on the bill.
Here is where the market sits, from vendor pricing pages:
- Intercom Fin: $0.99 per outcome, on top of per-seat plans. Confirmed on Intercom's pricing page. An "outcome" includes a resolution, and also billable events like a procedure handoff.
- Help Scout AI Answers: $0.75 per resolution, billed on top of seat pricing, per Help Scout's pricing page. One resolution is billed per conversation regardless of how many AI replies it took.
- SupportWire: $19 per seat per month, with 50 AI resolutions included each month, then $0.49 per resolution after that. Full pricing is on the pricing page.
SupportWire's first 50 resolutions are included with the $19 seat, then $0.49 each. Help Scout and Intercom rates are their published per-resolution prices as of August 2026, billed on top of their own seat fees.
The gap is not cosmetic. At $0.49, SupportWire's per-resolution rate is exactly half of Fin's $0.99. Because you pay that rate on every resolved conversation, the difference compounds fast at real volume. We ran Intercom's own published ROI model, same conversation count, same resolution rate, same human cost, and changed only the per-resolution price in the full AI support cost breakdown. The one number that changed took thousands of dollars off the monthly bill for an identical outcome.
The lesson is the same one Intercom itself makes: compare per-resolution prices only at the same resolution rate, because a cheap AI that resolves less just shoves more conversations onto humans at full cost. Pick the AI that resolves well, then pay as little as you can for each of those resolutions.
How SupportWire answers customer questions#
Full disclosure, since the section is about our product: we built SupportWire, so treat the bias as declared. Here is how our AI, Kal, does the two-step process this whole guide describes.
Kal grounds every answer in your Knowledge Store, which is the repository of sources Kal reads from: your help articles, policies, product docs, and past resolved conversations. When a customer asks a question, Kal retrieves from that repository and writes a reply grounded in what it finds, the retrieval-then-generate pattern from the top of this guide, not open-ended generation. If the sources do not cover the question, Kal does not invent an answer. That is the Air Canada lesson built into the default behavior.
For conversations that need a person, Kal drafts the First Response so your team never starts from a blank box, and when it is time to escalate, Handover passes the human the entire context: the customer's question, what Kal retrieved, and what it already tried. The agent picks up mid-conversation instead of re-reading a cold thread. That is the clean handoff that makes a high resolution rate safe rather than risky.
If you want to go deeper on any piece of this:
- The AI agent is the resolution engine, how Kal grounds, resolves, and escalates.
- The AI support team is the fuller picture of Kal working alongside your humans.
- Auto-resolve is the mechanism that closes a conversation when it is genuinely done and hands off cleanly when it is not.
And if you are currently on Intercom and weighing a move, the practical steps, data export, messenger swap, and cancellation timing, are in how to switch from Intercom.
The short version of the whole guide: AI answers customer questions well when it retrieves from sources you control, resolves at a rate your documentation can actually support, hands off cleanly the moment it is unsure, and charges you a per-resolution price that does not quietly become your biggest bill. Get those four right and AI stops being a gamble and starts being the cheapest, fastest first responder your team has.
Frequently asked questions
A support AI answers a customer question by retrieving the most relevant passages from a defined set of sources, help articles, policies, and past tickets, then writing a reply grounded in what it retrieved. This is called retrieval-augmented generation. It is different from an open-ended chatbot that generates from training data alone, and the grounding step is what keeps the answer tied to your actual policies rather than a plausible guess.
Yes. Even a grounded AI can produce a wrong answer, and the company is liable for it. In Moffatt v. Air Canada (2024 BCCRT 149), a tribunal ruled that Air Canada owed a duty of care for what its chatbot told a customer and awarded $650.88 CAD after the bot gave false refund information. Grounding in current sources lowers the risk, but IBM notes it cannot make a model error-proof, which is why a clean handoff to a human still matters.
Vendor-stated rates run high. Intercom's Fin site claims a 76% average resolution rate. Independent forecasts are more measured: Gartner predicts agentic AI will autonomously resolve 80% of common customer service issues by 2029, and Salesforce's State of Service report expects 50% of service cases to be resolved by AI by 2027, up from 30% in 2025. A realistic rate today depends heavily on how well your sources cover the questions you get.
Most AI support tools bill per resolution, not per seat. Intercom's Fin charges $0.99 per outcome, and Help Scout's AI Answers charges $0.75 per resolution, both on top of seat fees. SupportWire is $19 per seat per month with 50 AI resolutions included, then $0.49 per resolution after that. You can run the full per-resolution math in our AI support cost guide.
Updated July 2026
