November 19, 2026

What leads a customer to trust an AI Agent. Designing conversations that build trust

Trust is earned through precise answers, consistency across channels, and the certainty that every commitment will be kept

‍

A customer trusts an AI Agent when they receive a correct answer, find the same information on every channel, and see every operation the Agent has taken carried through to completion. Courtesy and fluent language matter less than one might think. What matters more is precision, the Agent's ability to state its own limits, transparency about how personal data is used, and the certainty that a request will actually be resolved.

For a company, trust is not a question of image. It determines whether, the next time they need something, the customer goes back to the Agent or looks for another way to get an answer and, in the worst cases, abandons the purchase. Italians are now familiar with AI Agents, but they do not accept them blindly. According to our annual observatory Customer Experience Unlocked 2026, conducted in Italy with Dynata among 1,000 consumers and 150 business decision-makers, 41% of Italians willingly use companies' AI Agents when they are available. Another 38% use them but would rather avoid them. That 38% shows how much room companies still have to win their customers' trust.

Why trust determines the value of an AI Agent

People who prefer to avoid AI Agents do not do so on principle, but out of experience. The reasons they cite most often are that the Agents do not understand their requests (68%) and give inaccurate answers (67%). Further behind come an interaction perceived as impersonal (45%) and procedures that are too long and complicated to reach an answer (37%). Distrust, in other words, stems from conversations that did not work.

The same observatory shows the conditions under which trust is granted. Most Italians (81%) tend to check the information they receive before making a decision, and 76% feel more comfortable when the AI Agent explains the reasoning behind its recommendations. Inconsistency weighs just as heavily, with 82% saying they would abandon a purchase if the information were not consistent across the website, the call centre, the AI Agent and other touchpoints. On personal data, caution has grown. Only 7% say they fully trust how companies handle the data they share with an AI Agent, down from 13% in 2025.

Distrust also carries an operating cost. A customer who has received a wrong answer often does not report it. They check elsewhere what the Agent told them and, the next time, choose another channel. Each inaccurate answer thus generates a new contact and weakens the perceived reliability of the entire service.

Precision before fluency

The first quality a customer expects from an AI Agent is the ability to solve the problem with an exact answer. Courtesy comes second. Anyone designing an Agent needs to take this into account because language models are built to produce plausible text, not necessarily correct text. They always write fluently and convincingly, even when the content is wrong. The customer has no way of telling a correct answer from one that is merely well written until they discover the error at their own expense. That is why an Agent's quality is judged by the accuracy of its answers, not by their elegance.

Precision comes from concrete design choices. Every answer is grounded in verified, up-to-date company knowledge, not in the model's general knowledge. Amounts, deadlines, contractual conditions, and codes are reported exactly as they appear in the source, without paraphrase. When the question concerns the individual customer's situation, such as the status of an order or the amount of the latest bill, the Agents retrieve the data directly from the company's systems instead of offering a generic explanation. Finally, every question receives a focused answer. Additional information comes afterwards, and only if it is needed.

Stating its own limits

An Agent that always has an answer is not more reliable than one that knows when to stop. The opposite is true. The customer checks, compares, verifies. An Agent that makes up a plausible answer is found out sooner or later, and from then on every statement it makes is called into question.

Stating its own limits means, first of all, recognising when a question goes beyond the knowledge available and saying so clearly, indicating who can answer and when. It also means distinguishing a certain fact from an estimate, for example, a delivery that has already been scheduled from a date that is only expected. The same transparency applies to the answers the Agent is able to give. Explaining them, by citing the clause, the tariff rule, or the procedure they derive from, allows the customer to verify them and makes the Agent more credible.

Then there are moments when the customer needs a person, such as a high-value purchase or a decision with significant financial consequences. A well-designed Agent recognises these moments and offers a handover to an operator before the customer has to ask, passing on all the context gathered. Offering it at the right time strengthens trust in the channel, because the customer knows they can count on a person when it matters.

Confirming what has been done

Many conversations end with a phrase such as "your request has been forwarded". The customer does not know whether the operation has been carried out, when it will be, or who is handling it. It is precisely this uncertainty that leads them to contact customer service again in the days that follow.

A conversation that builds trust ends with four elements.

  • An explicit outcome. The customer knows whether the operation has been carried out or not.
  • A verifiable reference. The case number or the return code makes it possible to trace the request at any time.
  • The next steps. The customer knows what will happen and when.
  • A way to follow progress. The customer can check the status of the request without having to start over.

All of this is only possible when AI Agents operate directly on the company's systems rather than merely describing procedures. There is, however, one rule that admits no exceptions. The Agent confirms only what the system has actually recorded. If an operation fails, it says so immediately and proposes an alternative. A confirmation given too early and then contradicted by the facts damages trust more than any incomplete answer. When execution takes time, as with a refund, trust grows if it is the company that notifies the customer once the case is closed, without the customer having to ask.

Consistency across channels and over time

The customer perceives an inconsistency as an error. If two channels give different answers to the same question, at least one of them is wrong, and the customer does not know which.

Inconsistency almost always stems from the organisation, not the technology. Each channel has its own documentation, updated by different teams at different times. A promotion already withdrawn from the website is still being offered in the chat. A revised contractual condition reaches the operators but not the app. The solution is a single source of knowledge for all channels and for the operators, with an owner and a review date for every piece of content.

Consistency also concerns time. A customer who has already explained their problem in the chat should not have to repeat it on the phone, nor to the operator who takes over the conversation. Agents that retain the context of the relationship, from previous requests to operations in progress, give the customer the most valuable feeling of all, that of being recognised.

The right tone for every moment

Those who design an AI Agent tend to confuse empathy with courtesy. Many customers who avoid Agents complain of impersonal interactions, yet few ask for warm answers. The two are not contradictory. Customers do not want stock phrases. They want to feel that their situation has been understood.

A customer reporting a double charge does not need three sentences of apology. They need one sentence that acknowledges the problem and an action that solves it. An upbeat tone or a joke in response to a complaint, on the other hand, conveys the impression that no one is really listening.

That is why tone has to be defined by the type of request, not once and for all. An information request calls for brevity. A complaint calls for acknowledgement and action. A request linked to a bereavement, an insurance claim, or financial hardship calls for restraint and, often, the offer to speak to a person. In every case, the tone remains consistent with the brand's voice, in the chat as on the phone.

Personal data as a condition of the relationship

Trust in how data is handled is the most fragile point of the relationship. A customer who does not know how their information will be used often breaks off the conversation, and sometimes the purchase, too. Transparency is therefore not a formality but the condition for going any further.

In practice, this comes down to a few simple rules. The Agent states from the very first message that it is an artificial intelligence system, as the AI Act also requires. It asks only for the data needed for the operation at hand, explains why it is needed, and does not ask again for what the company already knows. Identity verification is proportionate to what the customer wants to do. A question about opening hours does not require it, a change of IBAN does.

Behind these rules lie architectural and data-handling choices that the customer does not see, but that the company must be able to demonstrate.

How to measure trust

Trust is not measured with a single satisfaction question. It shows in customer behaviour, and much of that behaviour already emerges from conversation data.

  • Spontaneous return. This is the share of customers who, after a first conversation, turn to the Agent again for a different request.
  • Repeat contact on the same issue. A customer who reopens the same matter in the following days, on any channel, either did not believe the first answer or did not see that it acted upon.
  • Requests for an operator right after an answer. These often indicate that the customer did not consider what the Agent said to be reliable.
  • Mid-conversation abandonment. It matters most when it happens at the points where the Agent asks for personal data.
  • Perceived quality once the request is closed. Collected after resolution rather than during the conversation, it measures the judgment on the outcome and not only on the exchange.

These indicators complement resolution and automation metrics and often explain why those figures change. A stable automation rate alongside rising repeat contacts, for example, indicates that the Agents are closing conversations without convincing customers.

Proof from the field

Enercom, a utility with 190,000 customers across Lombardy, Veneto, and Piedmont, has entrusted the first level of support to an AI Agent. Compared with its previous chatbot, built on predefined flows, the number of unique users reached grew by 131%, and sessions handled increased by 187%. Only 2% of conversations require a handover to an operator. These are the numbers for a channel that customers turn to more and more often.

Where to start

Work on trust starts from the conversations that have already taken place.

The first check concerns conversations in which the customer asked for an operator right after an answer from the Agent. Reviewing them shows where the answer was inaccurate, incomplete, or poorly explained. The second is a comparison of the answers given across channels to the most frequent questions, because every difference is an inconsistency that the customer may notice. The third concerns closing messages. For every operation handled by the Agents, the customer must receive the outcome, a reference, the timing, and a way to follow the request.

These checks can start right away, and the corrections that follow change the way customers judge the entire service.

FAQ

Why don't customers trust AI Agents?

In most cases, distrust stems from negative experiences, not from rejection on principle. Customers avoid AI Agents that do not understand their requests or that give inaccurate answers. Trust grows when the Agent answers precisely, states its own limits, and completes the operations it takes on.

Does an AI Agent have to disclose that it is not a person?

Yes. The AI Act requires people to be informed when they are interacting with an artificial intelligence system, unless this is obvious from the context. Beyond meeting the legal requirement, disclosing it from the very first message strengthens trust, because the customer knows who they are talking to and what to expect.

How do you measure customer trust in an AI Agent?

Through behaviour, as well as through satisfaction surveys. The most useful signals are spontaneous return to the Agent for new requests and repeat contact on the same issue in the following days. Requests for an operator right after an answer, abandonment during the conversation, and perceived quality measured once the request is closed also count.

Non crederci sulla parola
This is some text inside of a div block. This is some text inside of a div block. This is some text inside of a div block. This is some text inside of a div block.

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.