AI chatbot limitations for questions that need a person

AI chatbot limitations matter most when one wrong answer can cost money, expose an account or put someone at risk. In a 2026 global survey, Gartner found 87% of customers wanted access to a person when companies used generative AI. Start with these six handoffs.

A bot is better suited to repeat policy lookups than disputed decisions, while a person is better suited to choices that change money, rights or safety. That boundary gives the bot useful work without asking it to make promises your team may have to reverse.

Best for small teams: let the chatbot collect the facts on risky cases, then pass the decision and full conversation to one named person. Your customer gets a quick acknowledgment, and your team keeps control of the outcome.

1. Should disputed refunds go straight to a person?

Send disputed refunds to a person so the customer gets a decision based on the order, the policy and the circumstances. A disputed refund is any request where the customer challenges a refusal, asks for an exception or wants compensation beyond the published policy.

Routine return instructions are different. In a US survey published by SurveyMonkey, 59% of respondents said they were comfortable using AI to return an item. A bot can explain the standard window, provide a form and collect an order number, as long as it uses the current policy.

The risk begins when the answer requires judgment. In a Canadian case decided in 2024, an airline chatbot wrongly said a traveler could claim a bereavement discount within 90 days after buying a regular ticket. The tribunal ordered the airline to pay a partial refund worth about $482, plus other costs, according to Manatt, Phelps and Phillips.

Keep billing and account-sensitive topics in draft-only mode so a person can review the response before it reaches the customer. That approach pairs well with a clear process for handling refund conversations once an exception or disagreement appears.

What the bot should say: I can help with the standard return process, but a member of our team needs to review this refund decision. I’ll pass along your order number, the item involved and the outcome you’re requesting.

2. Should your chatbot give legal, medical or financial advice?

Don’t let a customer service chatbot give legal, medical or financial advice, even when the question seems closely related to your product. It can repeat a published fact, but it shouldn’t interpret symptoms, rights, regulations, investments or someone’s personal situation.

Customers already recognize this boundary. SurveyMonkey reported in its 2026 publication that 69% of US respondents were uncomfortable using AI for medical advice and 68% were uncomfortable using it for investment advice. A confident tone won’t make an unsupported answer safer.

Legal questions create the same problem. A 2024 investigation by The Markup and THE CITY found that New York City’s business chatbot gave incorrect answers about workers’ tips, housing vouchers and cash payments. The city later warned users not to treat its responses as legal or professional advice.

Set topic rules that catch phrases about diagnoses, treatment, liability, contracts, compliance, taxes and investment choices. Don’t replace one risky answer with a vague disclaimer followed by more advice. Stop the answer and provide a route to a qualified person.

What the bot should say: I can share our published product information, but I can’t assess legal, medical or financial questions. Please speak with a qualified professional, or tell me if you’d like our team to clarify what our service includes.

3. Should a chatbot handle passwords or identity checks?

Don’t let the chatbot decide who owns an account or reveal private information during a chat. It can direct a customer to the official password reset or verification flow, but a person should handle exceptions involving a lost email address, a changed phone number or suspected account theft.

The danger isn’t limited to data exposure. A bot can also invent a security rule. In April 2025, an automated support bot for a coding tool explained unexpected logouts by describing a login policy that didn’t exist, and the error contributed to subscription cancellations, according to the AI Incident Database.

Keep passwords, payment details and recovery codes out of the conversation. The bot shouldn’t request a password, full payment details, recovery codes or identity documents in an ordinary message. It should send the customer to an approved verification screen and stop if that process fails.

You can still automate low-risk directions. Tell customers where the reset page is, how to recognize an official email and how to contact the security team. For order lookups, use a verified flow that limits what the customer can see, like the safeguards described in this guide to a Shopify order-status chatbot.

What the bot should say: I can’t verify account ownership or view private account details in this chat. Please use the secure reset link. If that doesn’t work, I’ll send your case to a person who can complete the required checks.

4. Should an angry complaint stay with the chatbot?

Move an angry or emotionally sensitive complaint to a person before the bot repeats itself or offers an unsuitable fix. The goal is to give someone authority to listen, choose between options and take responsibility for the next action.

Look for more than angry words. Escalate reports of damaged goods, repeated service failures, discrimination, threats, injury, bereavement or a customer saying nobody has listened. These cases depend on context and empathy, not just the policy paragraph that shares the most keywords.

SurveyMonkey found that among US respondents who preferred human service, 61% said people understood their needs better and 50% said people offered more ways to resolve the issue. That extra room to choose matters when the standard remedy has already failed.

A writing assistant can suggest a reply inside the agent’s reply box, but the agent should review it before sending. This keeps the writing help while leaving the apology, remedy and promise with the person responsible for the case.

What the bot should say: I’m sorry this has been such a frustrating experience. A member of our team needs to review what happened and decide the next step. I’ll pass along this conversation so you don’t have to explain it again.

5. Should the bot set a custom quote or contract term?

Don’t let the bot commit your business to a custom price, discount, delivery date or contract term. It can collect the customer’s requirements and share published prices, but someone with approval authority should make the offer.

Custom quotes often depend on details the bot can’t confirm, such as scope, stock, location, taxes, staffing and deadlines. If one detail changes, the quoted price or schedule may no longer work. A polished answer can still be a promise your business never approved.

Gartner reports that customers still contact companies directly for transactions, account-specific information and harder problems. That makes human review the safer choice when a conversation moves from general information to a binding decision, according to its 2026 customer service research.

The bot’s useful role is intake. Ask what the customer needs, when they need it, where the work will happen and whether there are special requirements. Send those details to the person who prepares quotes, without estimating a final price or promising availability.

What the bot should say: I can collect the details for your quote, but a member of our team needs to confirm the price, timing and terms. Please share the scope, location and preferred date, and I’ll pass them along for review.

6. Should the bot answer when it lacks verified account data?

Don’t let the bot fill gaps with a likely answer when it can’t verify the customer’s account, order or service record. It should say what information is missing and hand the case to someone who can check the source system.

This is where confidence matters. Confidence is the system’s estimate that it found the right information for the question. Gartner advises support teams to let generative AI attempt a resolution only when confidence is high and to keep a clear route to human support in its 2026 findings.

A good knowledge base reduces missing answers, but it doesn’t remove the boundary. Review the questions that trigger handoffs, correct old policies and add answers for repeat requests. This guide explains how to build an AI chatbot knowledge base around information your business can support.

What the bot should say: I don’t have verified information for this account, so I don’t want to give you the wrong answer. I’ll send the conversation to our team with the details you’ve already shared.

Start with disputed refunds because they combine customer frustration, financial impact and pressure for an immediate promise. Once that rule works, add professional advice, security, complaints, custom quotes and unverified account data in that order.

Make the handoff visible and easy. A practical chatbot-to-human handoff preserves the conversation, tells the customer what happens next and avoids another round of questions. The bot has done its job when it recognizes that a person should decide.

Human support remains part of the plan for most service teams. In a March 2025 poll, Gartner found 95% of customer service leaders intended to retain human agents. The useful division is simple: let the bot answer repeat facts, and let people own judgment, risk and exceptions.