AI customer service and support: setup and handover rules.
How to set up AI customer support that resolves real questions: the four modes, handover rules written as code, what the Air Canada chatbot ruling teaches, and metrics with clear definitions.
AI customer support is software that answers customers, helps your support team, routes tickets or analyses conversations. For a mid-size company, the best first project is usually one self-service agent on the channel with the longest queue. It resolves status and policy questions from your own content and systems, and hands everything else to a person with the full conversation attached. The agents that fail are rarely the ones that answer badly. They are the ones with no clean way out.
Most people have met the bad version: a chatbot that can’t answer, won’t let you reach a person, and leaves you angrier than before. This guide covers the modes of AI support, what to resolve and what to hand over, the rules that apply as of October 2026, and the measures that show whether it is working.
What is AI customer support, and which mode do you need?
The phrase covers four different kinds of system. They carry different risks, and most companies need only one or two of them at first.
| Mode | What it does | Who talks to the AI | Main risk |
|---|---|---|---|
| Self-service agent | Answers questions and takes simple actions in your systems (look up an order, change an address) | The customer | A wrong answer or action reaches the customer directly |
| Agent assist | Drafts replies, summarises the case, finds the right policy for your agent | Your support team | Agents accept drafts without checking them |
| Routing and triage | Classifies each message, sets priority, tags the topic, sends it to the right queue | Nobody (back office) | An urgent case is mislabelled and waits |
| Analytics and quality review | Reads conversations to find the top reasons for contact, failed answers and gaps in your help content | Support leaders | Summaries that hide the awkward cases |
Agent assist is the lowest-risk start, because a person reads every reply before the customer does. It also has some of the best published evidence. A study of 5,179 customer support agents found that an AI assistant raised issues resolved per hour by 14% on average, and by 34% for novice and low-skilled workers, with minimal impact on experienced and highly skilled workers. That is one study, so treat it as a direction, not a promise.
A self-service agent is where most of the volume, and most of the risk, sits. Gartner forecasts that by 2029 agentic AI will autonomously resolve 80% of common customer service issues without human intervention. That is a forecast, and the word “common” carries the weight. The remainder still needs a person, which is why handover design decides how good your support feels.
What should an AI support agent resolve?
- Status questions: orders, deliveries, bookings, applications and tickets, looked up in your systems and answered directly.
- How-to and policy questions: answered from your help centre, product documentation and policies, with a link to the source.
- Simple changes: updating an address, rescheduling, resending a document or cancelling within policy, done in your systems, not just described.
- Information gathering: collecting the details, photos and documents a case needs before a person picks it up, so nobody has to ask twice.
Shipment status, delivery exceptions and proof-of-delivery questions are covered in AI for logistics companies.
When should it hand over to a person?
Write the handover rules down before launch, and enforce them in code rather than in the prompt. A model can be talked out of an instruction. It can’t be talked out of a rule that runs before it replies. Typical triggers:
- The customer asks for a person. One request is enough, in any wording or language.
- Two failed attempts at the same question. A third try is how doom loops start.
- Nothing relevant was found. If retrieval returns no strong match, the agent says so and offers a person. It does not guess.
- The action is above a limit. Refunds over a threshold, closing an account, changing payment details, anything that commits money.
- Distress, complaints and risk. Anger, a formal complaint, a legal threat, safety, health, bereavement or signs of vulnerability.
- A system is down. If the order system times out, the agent says it can’t check right now and hands over. It does not improvise.
What travels with the handover matters as much as when it happens. The person should receive a short summary, what the customer wants, what the agent already checked and found, the identity checks already passed, the sources it quoted, and anything it promised. The customer stays in the same channel and doesn’t repeat themselves. Out of hours, the agent states when a person will reply and captures a contact route, instead of leaving the customer in a queue nobody is watching.
Regulators have noticed what happens without this. The US Consumer Financial Protection Bureau’s report on chatbots in consumer finance (June 2023) describes customers stuck in loops of repetitive, unhelpful answers with no route to a human. Meta’s WhatsApp Business Messaging Policy allows automation inside the 24-hour customer service window, but requires prompt, clear and direct escalation paths, such as an in-chat transfer to a human agent. Our note on fallback chains covers how to build the handover itself.
The measure of an AI support agent isn’t how many conversations it keeps. It’s how many it actually resolves.
What went wrong at Air Canada, and what it teaches
The clearest public case of an AI support answer going wrong is Moffatt v. Air Canada, 2024 BCCRT 149, decided by British Columbia’s Civil Resolution Tribunal on 14 February 2024. A customer asked the airline’s website chatbot about bereavement fares. The chatbot said the customer could travel first and apply for the reduced fare within 90 days. The airline’s own bereavement page, which the chatbot linked to, said the policy did not apply to requests made after travel was completed. The customer relied on the chatbot, and the airline would not pay the reduced fare.
Air Canada argued it could not be held liable for what its chatbot said. The tribunal called that “a remarkable submission”. It held that the chatbot is “still just a part of Air Canada’s website”, that the company is responsible for all the information on it, and that it “makes no difference whether the information comes from a static page or a chatbot”. It also found that the airline never explained why customers should have to double-check one part of its website against another. The airline was ordered to pay the difference in fare, plus interest and tribunal fees. For the travel and hotel side of this problem, see AI in travel and hospitality.
This was a small-claims decision in one Canadian province, so it doesn’t bind courts elsewhere, and whether a similar claim would succeed in your market is a question for a qualified lawyer. The engineering lessons hold regardless:
- Treat every reply as a published statement. What the agent says is what the company said.
- A link to the right page isn’t a safety net. The wrong answer on screen was the one the customer used.
- Keep policy facts in one maintained source. Eligibility rules, deadlines and refund terms should be retrieved at the time of the question and quoted, not recalled by the model.
- Let code decide anything that commits money. The model can explain a refund rule. A deterministic check should decide whether the customer qualifies.
- Test the policy edge cases. “Can I claim after I’ve travelled?” belongs in your test set, with the correct answer written by someone who owns the policy.
- Log what was said and which passage it came from. If a customer disputes an answer, you need to see exactly what they saw.
How do you keep the answers right?
An AI agent is only as good as what it answers from. Three things matter most:
- Grounding. Answers come from your help content, policies and live systems, never from the model’s general knowledge. Our note on retrieval that knows when to say “I don’t know” explains why abstaining matters.
- Current content. Outdated help articles become wrong answers, sometimes in the very form the Air Canada customer met: two pages that disagree. Fixing the knowledge base is often the first job.
- Testing on real conversations. Hundreds of real past tickets, including the awkward ones, scored before every change (see evals before features).
How does it differ by channel?
The same agent core (knowledge, system access, rules and tests) can sit behind each channel. The interface, and the failure modes, change.
- WhatsApp: customers expect fast, short replies at any hour, and the platform’s rules shape the design: automation inside the 24-hour window, approved templates outside it, and an escalation path throughout. We run a WhatsApp sales agent in production that handles 70% of a solar company’s monthly sales, with 99% uptime and a 4.8/5 customer rating, and the same architecture serves support. See inside a WhatsApp sales agent, and AI for sales teams for the selling side.
- Web chat: the easiest place to start, because you control the page. Handover can be a live transfer, with the transcript already open for the agent.
- Email: slower and longer, with quoted history, attachments and auto-replies. Start with drafts that a person approves, and move to automatic replies only for the question types where the drafts are consistently right. The agent must not answer another auto-reply, or two systems will talk to each other forever.
- Voice: the hardest. Delay is audible, callers interrupt, and speech recognition stumbles on order numbers and names. Confirm key details back to the caller, and make “agent” or “representative” work at any point, with a warm transfer that passes the summary along.
Where AI is the wrong tool in support
- Discretionary decisions. Goodwill refunds, exceptions to policy and retention offers belong to a person with authority. The agent gathers the facts and proposes.
- Upset and vulnerable customers. Speed is not what they need. A person who listens is.
- Eligibility and calculations. Whether a customer qualifies, or what they are owed, is rules work. Use code, and let the model explain the result.
- A broken knowledge base. If your help content is wrong or contradictory, an agent only spreads the errors faster. Fix the content first.
- A question a page answers better. Sometimes a clear order-tracking page beats a conversation. Not every contact needs an agent.
Which rules apply?
As of October 2026, the main ones are these. They describe what the rules say. A qualified professional decides how they apply to you.
- Tell people it’s AI (EU). Article 50 of the AI Act applies from 2 August 2026. The European Commission says chatbots and other interactive AI systems will have to tell users they are dealing with AI, not a human. The text of Article 50 excepts cases where it is obvious to a reasonably well-informed person. Many businesses disclose everywhere, not only where required.
- Financial services (UK). The FCA’s Consumer Duty includes a consumer support outcome: support must meet the needs of retail customers, including those with characteristics of vulnerability, and must not put unreasonable barriers in the way of actions such as making a complaint or cancelling. The rules apply regardless of the channel used, so a chatbot is in scope.
- Consumer finance (US). The CFPB’s report above says institutions should consider the limits of chatbots before using them as their main way of serving customers.
- Untrusted input. Customers, and sometimes attackers, will try to make the agent do things it shouldn’t. Its permissions must be limited in code (see prompt injection is an input-validation problem).
- Personal data. Only the data a conversation needs goes into the model and the logs, handled under the data protection law of each market you serve.
How do you measure AI customer support?
Agree the definitions before launch, and write them down. Two teams can report very different numbers for the same agent depending on what “resolved” means.
| Measure | Definition | What it catches |
|---|---|---|
| Resolution rate | Conversations the AI started and closed with no person involved and no repeat contact on the same issue within seven days, divided by all conversations it started | Whether it actually helps. “The conversation ended” does not count. |
| Escalation rate and reasons | Share of conversations handed to a person, grouped by trigger (asked for a person, no answer found, over limit, distress) | Where the content or rules need work |
| Time to handover | Time from the handover trigger to the first reply from a person | A slow handover undoes the good work |
| Handover quality | Share of handovers in a reviewed sample where the person did not have to re-ask anything | Empty or unreadable summaries |
| Customer satisfaction, AI-handled | Survey score on conversations the AI handled, against your human baseline on similar questions | Whether customers are being served or only deflected |
| Repeat contact | Customers who contact you again about the same issue within seven days | Answers that seemed fine but weren’t |
| Answer accuracy | Share of a weekly random sample, reviewed by someone who owns the policy, that is correct and complete | Drift as products and policies change |
Measure on a human baseline from before launch, by question type. An average across everything hides the topics where the agent is weak.
How do you start?
- Pull the top contact reasons. Take three months of tickets and rank them by volume. Pick the two or three that are common, answerable and low-risk.
- Fix the content behind them. Update the help articles and policies those questions rely on, and remove contradictions.
- Write the handover rules and the definitions. Triggers, the handover summary, out-of-hours behaviour, and the measures in the table above.
- Build a test set from real tickets. Include the awkward ones, the policy edge cases and attempts to push the agent off course.
- Go live on one channel with review. A person reads the conversations daily at first, and the agent’s permissions start narrow.
- Widen only on the numbers. Add question types, actions or channels when resolution and accuracy hold. Keep the weekly sample review running after launch (see our services for how we build and run these).
Questions support leaders ask
Should we buy a support AI product or build?
Buy if your help desk platform’s built-in AI covers your questions well. Build when the agent has to act in your own systems, follow your specific rules, or serve channels and languages the product doesn’t.
Are we responsible if the AI gives a wrong answer?
In the Air Canada case, a tribunal held the company responsible for what its chatbot said, as it would be for any page on its website. The ruling is from one Canadian province, so ask a qualified lawyer how it applies in your market. In practice, assume customers and regulators will treat the agent’s words as yours.
Will it replace our support team?
It takes on the repetitive volume. Teams usually shift towards complex cases, quality review and improving the knowledge the agent answers from.
What is the difference between a chatbot and an AI support agent?
A traditional chatbot follows scripted menus and keywords. An AI support agent understands free-text questions, retrieves answers from your content, and can take actions in your systems within set limits. The second is far more useful, and it needs the handover rules and testing described above.
How fast can we launch?
A focused agent on one channel, answering from good help content, can go live in weeks, with a person reviewing its conversations at first.
Modulus Labs AI builds support agents that resolve questions from your own content and systems, and hand over cleanly to your team. Tell us about your support queue, and we’ll tell you what we’d build first.