Insights/Guides

How to automate customer support without trapping your customers

What to automate, what to keep with a person, what the law says about AI disclosure, and how to tell whether automation is really working.

Thinkiyo team··12 min read

Automate the questions with one correct answer. Keep a person on anything involving money, safety, the law or a customer who is clearly upset. Always leave a way out. Do those three things in that order and most of the failures in the news don't happen to you.

What customers actually object to

Most people will use a bot that works. They object to dead ends.

Gartner found in 2026 that 87% of customers say an option to reach a human is essential when a company uses generative AI for service, while 50% say AI makes interactions easier. Both can be true. Its follow-up finding is harder: only 27% would try a chatbot again after a bad experience. You get roughly one chance.

Vendor surveys point the same way, so read them as vendor surveys. Verint, which sells customer service software, reported in June 2026 that 61% of 5,000 US consumers prefer human agents, and that 69% of those would switch to self-service if it fully resolved their issue. Qualtrics, which sells experience software, surveyed more than 20,000 consumers in 14 countries and found nearly one in five who used AI for service got no benefit. A study by Twilio with YouGov across seven Asia Pacific markets listed AI not understanding questions, having to repeat themselves and scripted answers as Australian respondents' top frustrations.

Peer-reviewed work adds two design points. Crolic and colleagues found a humanised chatbot lowers satisfaction, firm evaluation and purchase intent when customers arrive angry. So don't give the complaint bot a cheerful name and face. And Luo and colleagues, in a sales call experiment with more than 6,200 customers, found that disclosing the bot before the conversation cut purchases by more than 79.7%. That is a sales setting, so it may not carry over to support. Expect disclosure to cost you something, and do it anyway, for the reasons below.

When it went wrong in public

  • Air Canada, 2024. A chatbot told a grieving customer he could claim a bereavement fare after travel. Air Canada argued the bot was responsible for its own words. The British Columbia tribunal called that "a remarkable submission", held the airline responsible for everything on its website and found negligent misrepresentation. Damages were CAD 650.88, plus interest and fees. The case is small, but it is the one every lawyer now cites.
  • Cursor, 2025. A support bot invented a login policy that didn't exist. Some customers cancelled, and the cofounder apologised for an "incorrect response from a front-line AI support bot".
  • Commonwealth Bank, 2025. The bank reversed 45 role cuts after a voice bot rollout and said "this error meant the roles were not redundant". The union said calls had gone up and team leaders were answering them.

The pattern: the company owns what the bot says, and removing people early is expensive to undo.

The rules

The principle that holds almost everywhere: what your bot tells a customer counts as what your business told them, and consumer rights don't shrink because software answered. Specific duties to disclose AI or explain automated decisions vary a lot. This is general information and depends on your circumstances. Check your own market with a lawyer.

MarketTelling people it is AIAutomated decisions
USNo federal chatbot law that DLA Piper identified. The FTC Act is the baseline. Maine, Utah (on request) and New Jersey have state rules, and California's is narrow.California's ADMT rules apply to significant decisions by 1 January 2027, with notice, access and a right to opt out, unless the business offers an appeal to a human reviewer who can overturn the decision.
UKWe could not confirm an AI-specific chatbot disclosure rule. Check current guidance.Articles 22A to 22D require notice, a chance to make representations, human intervention and a way to contest fully automated significant decisions. In force from 5 February 2026.
EUAI Act Article 50 says people must be informed they are interacting with AI, unless obvious. Applies from 2 August 2026 and the Digital Omnibus did not delay it. The duty sits on the system's provider, which can be you if you built the bot or put it into service under your own name. If you bought it, ask your vendor how it is met.GDPR Article 22 gives a right not to be subject to solely automated decisions with legal or similarly significant effects.
AustraliaThe privacy regulator says public-facing AI should be clearly identified.From 10 December 2026, privacy policies must describe programs, including chatbots, that make or inform decisions significantly affecting people. That can apply even when a person reviews the output.
IndiaWe found no dedicated chatbot disclosure rule. Check for updates.We found no rule to cite. Check for updates.
CanadaThe Competition Bureau said in September 2026 there is "no AI exception" to the law. A federal consultation on AI transparency closed that month.We found no rule to cite. Check for updates.

Whether a refund refusal counts as a "significant decision" depends on the market and the facts. Don't assume. Keeping a person on refunds removes most of the question.

General consumer law reaches bots too. As one example, Australia's ACCC says any information from staff must be accurate and intent makes no difference, and Treasury's 2025 review described that law as technology-agnostic. Telling someone they have no remedy when they do is misleading, whether a person or a bot says it.

Step 1: Count why people contact you

Export three months of tickets. Tag them by reason. Note how often each reason appears and whether it has one correct answer. You will usually find a short list of reasons accounts for much of the volume. Only automate from that list.

Step 2: Sort everything into three groups

SituationWho owns itDo firstNever
Order status, hours, how to change an addressBot, with a way outAnswer from your live systemGuess when the data is missing
Delivery complaint, warranty question, most reviewsPerson, using an AI draftCheck the draft against policySend without reading
Refund or paymentPersonAcknowledge, confirm ownerLet a bot refuse or promise
Legal threat, safety, health, regulatorSenior personShort hold reply, escalateLet the bot argue
Vulnerable customer or bereavementPersonSlow down, skip scriptsRoute to self-service
Angry customerPerson, quicklySkip the botMake them prove it to a chatbot first
Bot doesn't understand twicePersonHand over with historyLoop back to the menu

Be strict about the last five rows. They are where the news stories come from.

Step 3: Write down the answers

AI gives answers as good as what it can read. Put refund rules, delivery commitments and warranty terms in one current place with an owner for each. Where policy is vague, your team has been filling the gap with judgement. A bot will fill it with a guess. Make every automated answer traceable to a source someone can check.

Step 4: Build the way out first

Do this before you switch anything on.

  • Let people ask for a person at any point, in plain words.
  • Pass the full conversation across, so nobody starts again.
  • Don't put a login in front of the handoff.
  • Say when no one is available and when someone will reply.
  • Tell people they are talking to AI.

Gartner's advice is similar: launch narrow, tell customers what the bot can and can't do, and treat it as a bridge to people.

Step 5: Use these messages

Bot opening line:

I'm an automated assistant for [company]. I can help with [two or three things]. Type "person" at any time and I'll pass you to the team.

Handoff, first reply from a person:

Hi [name], I'm [your name], a person this time. I've read what you told the assistant about [issue], so no need to repeat it. I'm checking [thing] and will update you by [time].

Public reply when someone complains about the bot:

[Name], you shouldn't have had to fight a bot to reach us. I'm [your name] and I have your case. Please message [private channel] and I'll pick it up from there.

Update when there is no news:

[Name], still waiting on [team]. Nothing new yet. I'll update you by [time] either way.

Closing reply:

[Name], [fix] is done. If it isn't right by [date], reply here and it comes back to me, no bot.

The hard case, when the bot got it wrong:

[Name], the assistant told you [what it said], and that was wrong. Sorry. Our policy is [correct policy]. Since you acted on what we told you, here is what we can do: [remedy].

Honour the bot's mistake where you reasonably can. It is usually cheaper than the Air Canada route.

Step 6: Start with drafts

Run the AI in draft mode first, with your team approving each reply. Their edits show you where it is wrong before a customer sees it. Automate one reason at a time, starting with the one your team changes least.

Time it. If reviewing drafts takes longer than writing replies, the automation is adding work. Agents will tell you this if you ask.

A rollout you can adopt on Monday

These timings are a suggested starting point, not a benchmark.

  • Week 1: tag three months of tickets. Agree the three groups.
  • Week 2: write down policy answers. Build the handoff.
  • Weeks 3 to 4: drafts only, every reply reviewed. Log edits.
  • Week 5: automate the one reason with the fewest edits, for a slice of traffic.
  • Week 6: check the measures below before adding a second.
  • Before any headcount decision: three months of data.

Mistakes people actually make

  • Counting closed chats as resolved. A customer may have given up or moved to another channel.
  • Removing people before the bot has proven itself.
  • Hiding the human route to protect the deflection number.
  • Letting the bot make up policy when a question falls just outside its data.
  • Naming the bot and giving it a face, then using it for complaints.
  • Reporting a "resolution rate" without checking whether those people came back.
  • Using the bot for the very customers who got hurt by it, such as after an outage.

What to measure

  • Repeat contact on the same issue within 7 days of an automated answer.
  • How many ask for a person, and how long they wait.
  • How many move to a public post or review afterwards.
  • Edit rate on drafts, by reason.
  • Handle time for what people still handle, since it often gets harder.

It is working when repeat contact stays flat or falls while automated volume rises, and the wait for a person holds. If containment rises and repeat contact rises with it, you have moved the work to the customer.

What runs in Relay today

Relay leans towards the draft-first model. Messages arrive in one queue, get a category, urgency and owner, and money or safety topics go to a person. Relay doesn't send anything without a person pressing send, and every ticket keeps an event log. Reopen rate is computed from ticket history.

Limits: there is no customer-facing chatbot or web chat, so it doesn't answer on its own. Drafts come from rules per category and need editing, because reading your policy documents into each draft is not wired up yet, and the autonomy settings on screen don't change that behaviour. Order and refund actions in other systems stay manual. See how Relay works, or contact us. For a worked example of the money path, read the double charge post.

Common questions

What should I automate first?

Frequent questions with one correct answer and little risk. Order status is the usual starting point. Automate it only if the bot reads live data, because a stale answer is worse than a wait.

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

In the EU the AI Act requires providers to build chatbots so people are told they are dealing with AI, from 2 August 2026 unless it is obvious, and some US states have their own rules. Elsewhere it may not be a specific duty, but Australia's privacy regulator recommends it. A plain line at the start is the safest global default.

Are we liable for what a chatbot says?

In Moffatt v Air Canada a Canadian tribunal held the company responsible for its chatbot's information. That is one small claims decision, not a global rule. Most consumer laws look at what your business told the customer, so plan on owning it.

Will automation let me reduce my team?

Plan on the work changing first. Simple questions go away, and what remains is harder and takes longer per case. Decide on team size after you have measured the results, and read the Commonwealth Bank reversal first.

How do we know the bot is really solving problems?

Check whether people come back on the same issue within a week. Silence after a bot reply doesn't prove anything. Follow a sample of closed chats by hand each month.

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.