People may be willing to let a chatbot sort out a delivery problem, but their patience has limits when a refund disappears or the conversation circles back to the same question. New research into AI customer service trust suggests consumers are increasingly open to automation if it solves the problem accurately. The key word is if. Customers appear much less interested in the novelty of speaking to AI than in getting a dependable answer and retaining a route to a person when they need one.
The figures come from a US consumer survey carried out by research firm Metrigy and promoted with customer-experience technology company NiCE. That context matters. The findings describe respondents’ reported attitudes in a particular market, not a universal measure of what every customer does during a real dispute. Even with that limitation, the result captures a familiar irritation: helpful automation can be welcome until it becomes a barrier between someone and the business responsible for solving the problem.
What AI customer service trust looks like in the survey
The companies say the research covered 1,000 US adults. In the reported results, 28.8 per cent would choose an AI-led interaction if a resolution were guaranteed. A further finding points to an appetite for combining human assistance with AI, rather than treating the choices as mutually exclusive. The promise of resolution is important because the question asks people to imagine a favourable outcome. It should not be interpreted as evidence that the same proportion would accept any chatbot, on any issue, with no guarantee of accuracy.
The survey also reports that trust remains uneven. In customer service, trust is not merely a feeling that a system sounds polite. It can mean believing the answer is right, being confident that account information is handled appropriately and knowing a mistake can be corrected. An assistant can deliver flawless sentences while still misunderstand a missing parcel or invent a policy that the company does not have. For a customer, the quality of the outcome matters more than the fluency of the reply.
The Metrigy survey summary provides the research context behind the vendor announcement. As with any survey, questions, response options and sample composition shape the results. They offer useful evidence of expectations, not proof that particular technology will improve every call centre. Independent operational measurements and customer outcomes are needed to assess real deployments.
A simple question becomes difficult when money is involved
Some customer problems are well suited to automation. A shopper may want to know when a parcel is due, whether a shop is open or how to change a password. If the system has access to reliable records and the action is straightforward, it can provide information immediately without making a person wait in a queue. That convenience is real and should not be dismissed merely because the speaker is a machine.
A refund dispute is a different situation. The customer’s account may contain incomplete information, an unusual exception or a disagreement about what happened. The conversation can require judgement, responsibility and access to rules that are not always represented accurately in a chatbot’s training material. A system may also have permission to explain a refund process without being authorised to grant one. When those boundaries are unclear, a confident response risks creating a promise that the business will not honour.
This is why customers need clear signs of what a bot can actually do. Does it merely suggest an answer, or can it change the order? Has it checked the account, or is it explaining a general policy? Can the customer ask for a human agent who will see the history rather than making them repeat the entire story? These practical details determine whether automation feels like service or evasion.
Why the human option is still useful
There is a temptation to measure success by how many conversations never reach a person. A business may call this containment and celebrate a higher percentage. But a customer who abandons a chatbot in frustration is not a successful resolution, even if no employee handled the enquiry. The more relevant measures include whether the problem was solved, how quickly, whether the answer stayed accurate and whether the customer had to make another contact.
LiveAIWire previously reported on a survey about keeping a human customer-service option. The emerging pattern is not a simple choice between people and machines. Customers can appreciate the speed of automation and still expect a person to take responsibility when the system reaches its limits. Companies that force every issue through the same automated gate may misread an efficiency statistic as a measure of satisfaction.
Handovers are particularly important. If a bot has already collected an order number and the customer has explained the problem, transferring to a human should preserve the relevant details where appropriate. Requiring the person to start again makes the automation feel like unnecessary paperwork. At the same time, the handover must protect privacy and avoid passing along an AI-generated summary as though every line were verified fact.
Why a friendly voice does not prove an accurate answer
Generative AI can produce natural conversation and adjust its tone to the person asking a question. That can make it easier to explain complicated processes. But warmth, humour and confident phrasing can hide uncertainty. A business should make it possible for the system to say that it does not know or cannot authorise an action. An honest limitation is often better service than a plausible but incorrect instruction.
There is also a difference between answering from an approved company knowledge base and producing a response without a reliable connection to current policy. Returns windows, fees, product availability and delivery options change. A support assistant should use the right up-to-date source and distinguish confirmed account details from general guidance. It needs monitoring and straightforward correction when policies change. Training a model once is not the same as maintaining a dependable service.
The stakes are higher in areas such as banking, insurance and essential services, where a mistake can affect access to money or urgent assistance. A wrong restaurant opening time is inconvenient. An incorrect explanation of a financial obligation may be seriously harmful. Businesses should set different thresholds for verification, escalation and human review based on what a failure would cost the customer.
The overlooked cost of customers trying again
Call centres have long tracked how many interactions they handle and how quickly agents complete them. Those measures can be useful for managing staff, but they do not reveal the whole customer experience. If someone uses a bot, then sends an email, then telephones to resolve the same issue, the business has generated several interactions rather than one. The customer has also spent more time on an administrative chore that should have been simple.
An effective AI support system should reduce repeated explanations and unnecessary transfers. It may be especially helpful for enquiries outside normal office hours, provided it has the authority and information to answer accurately. Yet customers should not be left believing that a complex case has been resolved when the system has only opened a ticket. Clear confirmation of what has happened, what remains pending and when a person will respond is a basic part of trustworthy service.
LiveAIWire has examined AI shopkeepers making misleading refund decisions in research simulations. That study concerned experimental AI agents, not evidence that ordinary retailers deliberately programmed their customer-service tools to lie. The useful connection is the need for limits, audit trails and oversight when software does more than repeat information. A chatbot with the ability to make commercial decisions deserves more scrutiny than one that merely gives directions to a help page.
What companies can learn without overselling the survey
The strongest business lesson is to test the service on real problems rather than comparing how human its voice sounds. Ask customers whether their issue was resolved correctly, provide a clear correction channel and examine where automated conversations fail. Track repeat contact and successful escalations as carefully as the percentage of calls a bot handles alone. An improvement in one metric can conceal a deterioration in another.
Transparency also helps. People should know when they are speaking to automation, what information the system uses and whether its actions have been completed. The company remains responsible for the service it provides, even if an AI supplier created the underlying software. Customers should not have to identify which contractor caused a mistake before they can obtain a remedy.
LiveAIWire has also covered how AI grocery workers in research settings can behave unexpectedly. The lesson across these different environments is that success depends on the task and the safeguards around it, not on how many jobs a developer says the model can perform. A helpful assistant has boundaries that are visible to users and staff.
The future may be a shorter queue, not a people-free business
AI may take more routine work from customer-service teams while human agents concentrate on unusual, sensitive or contested cases. If implemented well, that could improve both speed and the quality of attention given to people who need it. It could also create new responsibilities for staff who supervise automation, correct data and deal with escalations. Whether those benefits appear will depend on management decisions, training and the incentives used to judge performance.
The survey captures a willingness to try a system that works. It does not give businesses permission to make the human option disappear. Customers are making a practical calculation about their time, money and confidence. When the bot can solve the problem safely, it deserves a place in the service. When it cannot, the most useful thing an AI assistant may do is recognise that fact and make it easy to speak to someone who can.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.