The honest answer to when not to use AI for customer service is simpler than the vendor decks suggest. Not on tickets where a wrong answer creates legal, financial, or emotional harm the AI cannot repair. Not on tickets where the customer has already asked for a human. Not on tickets where your own written source cannot answer the question. Everything else is a design problem, not a philosophical one.
Why the question matters more in 2026
Gartner's August 2026 survey of 3,566 customers found 87% say it is essential to be able to reach a human agent when a company uses generative AI in service. A separate SurveyMonkey study puts the figure at 79% of Americans who strongly prefer a human over an AI agent. The pattern is not "AI is bad". It is that customers accept AI on their terms, and the terms include a clean route out.
The stakes have moved too. In February 2024 the British Columbia Civil Resolution Tribunal ordered Air Canada to pay CAN$812.02 to Jake Moffatt after its chatbot invented a bereavement-fare policy. The airline argued the chatbot was a separate legal entity. The tribunal disagreed and held the company liable for what the bot said, the same as a static page. The EU AI Act's high-risk provisions apply from 2 August 2026 and require human oversight, logging, and disclosure for a growing set of AI systems in consumer decisions. When not to automate is now also a question of when not to invite a lawsuit or a regulator.
What is a ticket you should never automate?
Seven categories, in our experience.
Legal or regulatory questions. Consumer rights, dispute rights, refund entitlements under national law, insurance coverage, medical dosing, immigration status, tax filings, or employment law. Get one wrong and the misrepresentation is the company's, per Moffatt. Keep the AI on definitions and process, escalate the moment the customer asks whether they are entitled to something.
Bereavement, medical crises, and safeguarding. A message that opens with a death, a serious illness, or a safeguarding concern is not a place for a machine reply. There is no upside to routing this through the bot. Route direct to a human, and if you are open 9 to 5, add an out-of-hours acknowledgement written by a person, not generated.
High-stakes financial decisions. Refund amounts above a threshold you set (around 200 euros for consumer stores, higher for B2B), chargebacks in flight, disputes over billing, subscription cancellations tied to a churn moment. The AI can gather context. The decision belongs with a human whose name the customer can later cite.
Retention and cancellations. The moment a customer says "cancel", the AI's job is to hand off cleanly and log the reason, not to run a retention playbook. Automated retention feels manipulative and burns the goodwill you had left. If you want to save the account, put a person on it.
Complaints with emotional content. A damaged wedding gift. A missed birthday delivery. An outfit that arrived after the interview. The customer needs to be heard before they need policy. AI answers here read as flippant no matter how well written, because the customer wanted a witness.
Anything the AI's own source cannot answer. If the retrieval layer returns no relevant chunk, or returns one with low confidence, the AI should not answer. It should say so and hand off. Bots that invent an answer here are the reason customers say they hate chatbots.
Reputation-critical accounts. Named investors, press contacts, enterprise buyers in an active deal, a customer with a large social following who has already had a bad experience. Put a human on these even when the AI could technically answer. The cost of a small AI slip on a visible account is disproportionate.
When is AI a good fit?
To keep this honest, here is the mirror. AI is a strong fit for a defined slice of ticket types, and customers actively prefer it there. Industry surveys in 2025 found 74% of customers prefer chatbots for simple questions, and 82% prefer them to waiting for a human on simple, fast transactions.
The comparison, at a glance:
| Ticket type | Good AI fit | Keep human | |---|---|---| | Order status ("where is my order?") | Yes, if live order data is available | Only if the order is lost or damaged | | Return eligibility check | Yes, from a clear policy | Refund amount decisions above a threshold | | Product spec, sizing, compatibility | Yes | Custom-quote questions | | Password reset, account access | Yes, with correct identity checks | Any account-takeover suspicion | | Refund request under a small threshold | Yes | Anything contested or unusual | | Complaint with emotion | No | Always | | Legal or entitlement question | No | Always | | Bereavement, medical, safeguarding | No | Always | | Cancellation or retention | No | Always | | Anything the AI's sources cannot answer | No | Always, with a clean handoff |
Two rules do most of the work. If the AI's source cannot answer, do not answer. If the customer asks for a human, give them one.
How should the AI decide to escalate?
Confidence-based routing plus a short list of hard rules.
The AI should carry a numeric confidence on every draft answer. Below a threshold (a good starting point is 0.7 on a normalised scale), it does not send. Above, it can send. In between, post the draft to a human as a suggested reply. That middle mode is underused and works well on teams of five to fifty.
On top of that, hard rules based on intent, sentiment, and account tags. Sentiment below a threshold triggers escalation regardless of confidence. Certain intents (cancel, refund, dispute, legal) route direct to human. Certain account tags (VIP, at-risk, in-dispute) route direct to human. The word "human" or "agent" in the customer message routes direct to human, always.
The failure mode we see repeatedly is a bot that keeps trying past the point of usefulness. It closes tickets it should have escalated and creates worse second tickets two days later, to a human agent who now sees a customer that already feels unheard. That second ticket is more expensive than the first would have been. Design against it.
What does the AI actually do on tickets you do not automate?
It still helps. On a ticket routed to a human, the AI can extract the question, order number and account context in one line, suggest a first-draft reply the human accepts or rewrites, retrieve two or three relevant policy snippets with citations, and log the reason for escalation so you can measure it. That is not "AI doing the ticket". That is AI making the human faster on the tickets that need a human. On a small team this is almost always the higher-return use of the model. For the underlying arithmetic see our support staffing math piece.
The legal and regulatory backdrop
Two things to know if you set support policy. First, the Moffatt precedent. A company is liable for what its AI chatbot tells a customer, on the same footing as a static page. Grounding, refusal, and logging are legal controls, not just quality controls. Second, the EU AI Act high-risk provisions apply from 2 August 2026. Article 50 requires that users be told they are interacting with an AI system, and high-risk systems require human oversight and traceability logs. For a fuller EU-specific take see our EU AI Act primer for SMBs. If your bot's answer could plausibly steer a consumer decision, assume the regulator will ask to see the source it used and the logs of what it said.
How Keloa approaches this
We build AI agents that answer from your own sources with citations and decline when they cannot. We route straight to a human on cancellation, dispute, refund above the threshold you set, and any message flagged by sentiment or explicit request. Escalations land in the unified inbox with the full conversation, the customer context, and the AI's draft reply if one existed. Confidence-based routing lives in the flow builder, so the rules that fire escalation are visible and adjustable, not buried in a model.
Per-reply pricing makes the incentive line up: we do not get paid for a reply the AI should not have sent. That is the design point.
Frequently asked questions
Is there a rule of thumb for the "confidence threshold" for AI replies? A starting point is 0.7 on a normalised scale. Sample the replies at 0.7 to 0.8 weekly for a month, then move the threshold up or down based on how many of them a human would have edited. Most teams end up between 0.7 and 0.85 depending on the ticket mix.
What if the customer asks for a human but the AI could easily answer the question? Give them a human. The correct answer to "let me talk to a person" is a person, not a better bot reply. Gartner's August 2026 survey put the customer expectation for a human option at 87%. You do not want to be in the losing 13%.
Are there ticket types where a wrong AI answer creates legal risk? Yes. Anything touching entitlements, refunds under consumer law, medical dosing, insurance coverage, immigration, and tax. The Moffatt v Air Canada ruling in 2024 established that a company is liable for what its bot says. Keep the AI on definitions and process, escalate the entitlement question.
How should we log AI decisions for compliance? Store the customer message, the retrieved sources, the AI's draft, the confidence score, the escalation decision, and the human's edit if any. Retention aligned to your regulator's expectation. For high-risk contexts under the EU AI Act, this is not optional.
Does refusing to answer look bad to customers? Less bad than a wrong answer. A short "I do not have enough to answer that, putting you through to a colleague" reads well when it comes with a clean handoff and a quick human reply.
Is there a case for automating cancellations? Only the mechanical step. Confirming the cancellation is fine. The conversation that led to it is not. Route the reason-for-leaving thread to a human and log the sentiment.
Ready to draw the line for your own team? Book a demo and we will walk through your ticket mix and where the confidence and escalation lines should sit.