Insights/Guides

What AI customer service chatbots can and can't do in 2026

What modern AI agents handle well, what must stay with a person, the disclosure and liability rules by market, how to test a bot and how vendors charge.

Thinkiyo team··12 min read

A modern AI customer service agent can read a messy question, find the answer in your help content or order system and reply in plain language. It is good at questions with one correct answer and bad at judgement, money and distress. In most markets you are responsible for what it says, and in the EU you must now tell people it is AI. Use it for the first group, keep people on the second and test it hard before anyone sees it.

For what to automate across your whole support operation, read how to automate customer support without trapping customers. This post is about the chatbot itself.

Old bots and new agents

The chatbots most people hated were decision trees: a flowchart of buttons and keywords. Match a branch and you got an answer. Miss and you got the menu again. An AI agent built on a large language model works differently.

Decision-tree botAI agent
UnderstandsButtons and set keywordsFree text, typos, several questions in one message
Answers fromScripts someone wroteYour help content, policies and connected systems
Can take actionsOnly what was hard-codedWhatever you connect, such as order lookup or address change
Fails byLooping back to the menuGiving a confident, fluent wrong answer

The last row matters most. An old bot failed visibly. A new agent can fail invisibly, with an answer that sounds right and isn't.

What suits automation

A question suits an AI agent when all four of these hold:

  1. There is one correct answer your business has written down.
  2. The answer comes from data the bot can read live, or from policy that rarely changes.
  3. A wrong answer is cheap to fix.
  4. The customer is not upset, at risk or asking for an exception.

Order status, opening hours, how to reset a password, what a plan includes, how to start a return and where to send a document all pass.

What should never be automated

Some contacts fail the test on every point.

  • Refunds, chargebacks and payment errors. Money decisions need someone who can say yes, and refusals carry legal weight. A bot that promises a refund creates an obligation. A bot that refuses one may be wrong about the customer's rights. See our post on handling a double charge.
  • Complaints with legal weight. Threats of legal action, regulator complaints, safety issues, injury, data breaches and discrimination claims. Anything the bot says may end up in evidence.
  • Vulnerable customers. Bereavement, illness, financial hardship, domestic abuse, someone who says they are scared. Scripts go badly here, and the Air Canada case below began with a grieving customer.
  • Exceptions and goodwill. "Normally no, but in your case yes" is a person's call.
  • Anyone who asks for a person. Honour it the first time.

The rules

The principle that holds almost everywhere: what your bot tells a customer counts as what your business told them, and consumer law applies the same way whether a person or software answered. Specific duties to disclose AI vary by market. This is general information, it depends on your circumstances and it changes, so check your own market with a lawyer.

The case everyone cites is Moffatt v Air Canada, decided by British Columbia's Civil Resolution Tribunal in February 2024. Air Canada's chatbot told a grieving customer he could book a full fare and claim the bereavement rate within 90 days. The airline's real policy didn't allow that. Air Canada argued it could not be held liable for information from its representatives, including a chatbot. The tribunal found the airline "did not take reasonable care to ensure its chatbot was accurate" and awarded CAD 650.88 plus interest and fees. It is one small claims decision, but the reasoning travels.

MarketTelling people it is AIAccuracy and liabilityReaching a human
EUAI Act Article 50: providers must design systems that interact with people so they are informed it is AI, unless obvious, "at the latest at the time of the first interaction". Applies from 2 August 2026; the Digital Omnibus did not delay it.General consumer law.GDPR Article 22: where solely automated decisions with legal or similarly significant effects are allowed, people get "at least the right to obtain human intervention".
UKThe CMA's guidance on AI agents says if knowing it is AI "might affect people's decisions then you should tell them".Same guidance: "you are responsible for what an AI agent does in the same way you are responsible for what an employee does", even if a supplier built it. Fines can reach 10% of worldwide turnover.Articles 22A to 22D, in force from 5 February 2026, require human intervention and a right to contest for significant decisions made solely by automation.
USNo federal chatbot law that DLA Piper identified. Maine, New Jersey, Utah (on request) and a narrow California rule apply.The FTC said in Operation AI Comply (2024) there is "no AI exemption from the laws on the books".No general rule. For financial firms, the CFPB warned in 2023 about "doom loops" with no "offramp" to a human, and said chatbots "must comply with all applicable federal consumer financial laws".
AustraliaNo AI-specific chatbot law we could confirm.The ACCC says information from staff must be "accurate, truthful and based on reasonable grounds" and "it makes no difference whether a business intends to mislead". False claims about a consumer's rights or remedies are covered.No general rule we could confirm.
IndiaNo dedicated chatbot disclosure rule found.General consumer law.Under the Consumer Protection (E-Commerce) Rules 2020, e-commerce entities need a grievance officer who acknowledges complaints within 48 hours and resolves them within one month, per Lakshmikumaran and Sridharan. A bot can't be that officer's only face.
CanadaNo federal rule we could confirm.Moffatt v Air Canada, above.No general rule we could confirm.

The EU duty in Article 50(1) sits on the provider. If you bought the bot, ask your vendor in writing how disclosure is met. Whether a refund refusal is a "significant decision" under GDPR or UK law depends on the facts; keeping a person on refunds removes most of that question.

A decision table

SituationWho owns itDo firstNever
Order status, hours, password resetAI agentAnswer from live data, cite the sourceGuess when data is missing
What a plan includes, how returns workAI agentAnswer from current policy textAnswer from an old help article
Delivery complaint, warranty questionPerson, AI draftCheck the draft against policyLet the bot decide eligibility
Refund, chargeback, payment errorPersonAcknowledge, confirm owner and timelineLet the bot promise or refuse
Legal threat, regulator, safety, data breachSenior personShort holding reply, escalateLet the bot argue or admit fault
Bereavement, illness, hardship, fearTrained personSlow down, drop the scriptRoute to self-service
Bot fails to understand twicePersonHand over with the full transcriptSend them back to the menu
Customer asks for a personPersonHand over at onceAsk them to try the bot first

How to test a bot before launch

The CMA puts it simply: "Testing is a crucial part of training AI agents", and test results should be used to improve the agent "before you deploy it". A workable process:

  1. Build a test set from real tickets. Pull 200 to 300 recent conversations, anonymised, across your top contact reasons. Include typos, two questions in one, anger, sarcasm and every language your customers use.
  2. Write the correct answer for each, with the policy line it rests on. This is slow and it is the whole job.
  3. Add trap questions. Ask about policies you don't have, products you don't sell, a refund rule from a competitor, last year's prices. The right answer is "I don't know, here's a person". Anything else is a hallucination.
  4. Add escalation cases. Bereavement, a legal threat, "I want a human", a payment taken twice. Pass only if the bot hands over every time.
  5. Score every answer. Correct, partly correct, wrong, or invented. Treat any invented policy as a launch blocker, whatever the overall score.
  6. Check the handover. The person should get the full transcript, and the customer shouldn't repeat anything or log in first.
  7. Rerun after every change. New help content, new model, new prompt. Keep the test set and grow it with every real failure.

The pass mark is yours to set. A suggested starting point, not a benchmark: no invented policies, every escalation case handed over, and a support lead happy to sign their name under the rest.

How AI customer service software is priced

Three models dominate. All prices below are from each vendor's own page, as of November 2026, and change often.

ModelExampleWhat the page says
Per resolutionIntercom Fin$0.99 per resolution, defined as "no further help is requested after Fin's last answer". Minimum 50 outcomes a month with other helpdesks.
Per resolution, inside seat plansZendeskAI agents included in every plan, billed per automated resolution. 5 to 10 included per agent per month depending on plan, then US$2.00 each pay as you go, less if committed. Seats from US$19 per agent per month billed annually.
Per sessionFreshdeskFirst 500 Freddy AI Agent sessions included once per account, then $49 per 100 sessions. Seats from $19 per agent per month billed annually.
Per conversation or per actionSalesforce Agentforce$2 per conversation, or Flex Credits at $500 per 100,000, with a standard action using 20 credits.

What to check before you sign:

  • How "resolved" is defined. Per-resolution pricing usually counts a conversation where the customer didn't ask for more help. A customer who gave up and emailed instead may count as resolved.
  • Volume spikes. An outage raises usage bills along with contacts.

We break these down further in Intercom Fin pricing explained and Zendesk pricing explained. If a vendor beats your own team on cost per correct answer in your test set, it is the better choice for that slice of work.

Mistakes people make

  • Testing on the vendor's sample questions instead of your own tickets.
  • Stale content. An old help article with last year's returns window will be quoted confidently.
  • Hiding the human route to protect the containment number.
  • Giving the bot a name, a face and a cheerful tone, then using it for complaints.
  • No owner for the content the bot reads, so policy changes reach it late.
  • Treating "the bot said it" as a defence. Air Canada tried that.
  • Paying per resolution without auditing what counts as one.

What to measure

Containment versus real resolution. Containment is the share of conversations that never reached a person. Real resolution is the share where the problem was solved. Only the second matters, so check a sample of contained conversations by hand each month.

Reopen and repeat contact. How many customers come back about the same issue within 7 days, on any channel. If containment rises and repeat contact rises with it, the work has moved to the customer.

Handover rate and wait. How often people ask for a person, and how long they then wait.

CSAT, split by bot and human. Survey both. Compare like with like, by contact reason, because the bot gets the easy questions.

Wrong answers found in reviews and complaints. The target is zero.

It is working when real resolution rises, repeat contact stays flat or falls and the wait for a person holds.

A worked example

This example is made up to show the method. The company, people and numbers are invented.

Kestrel Home sells kitchenware online in the UK, Germany and Singapore. Its support lead, Amara Okafor, tags three months of tickets and finds order status, returns and product care cover most of the volume.

Her 250-question test set catches the first vendor trial inventing a 60-day returns window for Singapore, where Kestrel's policy is 30 days. The cause is an old help article still live there. Amara deletes it, assigns content owners per market and reruns the set clean.

Kestrel launches for order status only, with an opening line saying it is an automated assistant and how to reach a person. Refunds, damaged goods and anything mentioning a bereavement or a lawyer go straight to the team.

After six weeks, her monthly sample shows some "resolved" chats where the customer then emailed. Repeat contact within 7 days tells her which. She fixes the order-lookup gap behind them before adding returns.

What runs in Relay today

Relay is a support inbox, and it has no website chat widget or customer-facing bot. Incoming messages are classified by category, urgency and sentiment, including sarcasm, so a "great service, thanks" on a one-star review isn't auto-thanked. Refunds, payments, explicit requests for a person and low-confidence cases are routed to a person. Relay drafts a reply for each message, but drafts come from category rules or a general language model and are not yet grounded in your policy documents, so they need checking. A person presses send on every reply. Reopen rate is computed from ticket history. See how Relay works or contact us.

Common questions

Is an AI customer service agent the same as a chatbot?

An AI agent is a chatbot built on a language model that reads free text and can be connected to your systems to take actions. Older chatbots followed fixed decision trees. The practical difference is how they fail: decision trees loop, AI agents can give fluent wrong answers.

Do I have to tell customers they are talking to AI?

In the EU, Article 50 of the AI Act requires it from 2 August 2026 unless it is obvious, and several US states have their own rules. The UK CMA says to tell people if it might affect their decisions. A plain opening line is the safest default everywhere.

Are we liable for what our chatbot says?

In Moffatt v Air Canada the airline was held liable for its chatbot's wrong answer, and the UK CMA treats an AI agent like an employee. Plan on owning every answer, and honour mistakes a customer acted on where you reasonably can.

Can an AI agent handle refunds?

It can explain your refund process and collect details. Keep the decision with a person, because a bot that promises or refuses a refund can create obligations or misstate the customer's rights.

How do I stop a chatbot from making things up?

Restrict it to content you control, keep that content current and test it with questions it shouldn't be able to answer. The correct reply to those is a handover. Rerun your test set after every change to content, model or prompt.

Sources

See what Relay can do for your customer service

Get a personalised 30 minute demo on your own channels and policies. We show you Relay handling your real conversations, then set it up with you.

  • Built and connected for you
  • Can run in your own cloud
  • A person on every refund

Brands we work with

  • Bolt
  • Dave & Buster's
  • ETV Win
  • Malachite Cars
  • Space Angel
  • Infinity Quest
Made by Thinkiyo© 2026 Thinkiyo. Relay is a Thinkiyo product.Examples use Lumen Parcel, a fictional brand.