A better model. A better customer experience?

A customer who has already failed to get help from a chatbot may be difficult to persuade back. Telling that person the model has improved leaves the original grievance unanswered. They wanted a problem resolved and remember having to do more work to get there.

Gartner’s 2 September research makes that gap worth examining now. In a survey of 3,566 B2B and B2C customers, conducted in February and March, only 27% said they would be willing to try a chatbot again after a negative experience. That is a statement of willingness, not an observed churn rate or proof that the remaining customers would leave the brand. Gartner’s findings and survey context

It is nevertheless an awkward finding for anyone planning an AI launch around the promise of better conversation. The quality of the conversation is only part of the experience. What happens after the customer explains the problem may matter much more.

A more capable interface raises the expectation

Sonos announced Sonos 27 on 1 September. Today, 8 September, its external-assistant connection, Sonos 27mcp, enters Early Access, allowing connections with services such as ChatGPT and Claude. Its own music assistant, Sonos 27voice, is due in Early Access later this autumn; custom agents are targeted for 2027. Those are different stages of availability. None demonstrates improved customer satisfaction. Sonos’s 8 September rollout announcement

Music control and customer support are different jobs. Sonos is useful here because its announcement makes the broader promise tangible: a person should be able to ask for what they want in a way that feels natural. With the external connection only just entering Early Access, it is too early to judge the experience it creates.

Consider the simple request to play something suitable for dinner. A system could understand the occasion beautifully and still choose the wrong room, interrupt somebody else’s music or leave the volume too high. These are hypothetical failure cases, not findings about Sonos. They illustrate how understanding an instruction and successfully carrying it out are separate tests.

When a brand invites people to use a more flexible interface, it also invites a wider range of requests. The words “ask me anything” create an especially expensive expectation if the service can act on very little.

Design the failed attempt

A successful demonstration tends to hide the hardest part of a service: recovery. The customer has already supplied information, waited for a response and formed a view of whether the organisation understands the problem. Asking them to begin again throws away some of that effort.

Take a hypothetical customer trying to return a damaged item. A useful assistant might establish which order is involved and collect the relevant information. When the case needs a person’s decision, that colleague should receive enough context to continue the conversation, including the order details and any explanation the customer has already supplied about the damage. Tell the customer who now owns the case. Give them an expected response time they can hold the organisation to.

That sounds mundane beside a new model launch. It is also where a brand’s claim to be helpful becomes observable. A charming explanation of the returns policy will be of limited comfort if the customer cannot obtain a return label or discover why their particular case is stuck.

I would spend part of the launch rehearsal on the requests the system cannot finish. That includes an ambiguous request, unavailable information and a customer who changes their mind halfway through. The aim is to learn how the experience recovers, including whether the person can understand and reverse an action before it creates more work.

Different tasks warrant different boundaries. Choosing background music can tolerate an imperfect first suggestion. Changing a booking or making a payment may require explicit confirmation. Customers should know what will happen before committing to an action, with the amount of explanation reflecting its consequences.

Success needs a customer definition

A session that ends without reaching a human agent can look efficient in a service report. It could also represent someone giving up. The absence of escalation is therefore incomplete evidence of resolution.

Before launch, the team needs an account of what successful completion means for each job. For a delivery query, the customer may need a dependable arrival estimate. For a return, the relevant result may be an accepted request and usable instructions. Measures should follow that work through far enough to distinguish completion from an answer that merely sounded final.

Repeated contact helps expose the difference. If someone returns with the same unresolved issue, the earlier session should not retain an unqualified success label. The time spent repairing a mistaken action belongs in the assessment too, including work passed to colleagues elsewhere in the business.

There is a fair concern that this makes a trial slower and more expensive. It can. Following a customer’s problem over several days, through an initial answer and a follow-up contact with another team, is a more demanding exercise than watching a short demonstration in which the answer arrives immediately. The additional work earns its place when it reveals a defect that would otherwise reach a much larger audience. The appropriate trial size depends on the task and the harm a mistake can cause.

Let the launch claim catch up with the evidence

The marketing team has a role before the launch line is written. It can help choose the situations being tested, listen to the words customers use to describe the result and challenge a claim that extends beyond the proven scope. That is a better use of its judgement than adding warmth to an experience whose underlying problem remains unresolved.

A specific invitation also gives a disappointed customer a more credible reason to return. Explain the job the assistant now handles and make the alternative route visible. A broad assurance that the AI is smarter asks the customer to supply the missing connection between technical progress and personal benefit.

September brings both a fresh warning about customers’ willingness to retry and new examples of AI becoming part of everyday product control. The useful response is to examine the entire task, including the point where the model hands responsibility back. Before inviting more people in, the team should be able to show what happens to the customer whose first attempt still fails.