NTT DATA’s 2026 report on AI in contact centres finds a gap between how useful bots are perceived to be and how customers actually experience them. That gap is not a technology shortfall. It is what happens when automation is deployed without an operating model that owns the result.

The study is a useful counterweight to vendor optimism, provided it is read for what it is. Below we set out the scope, the caveats and what the finding means for regulated enterprises that run customer contact by telephone.

What the report is#

“Realising the power of AI in contact centres” was published by NTT DATA in 2026. It combines two populations: 1,600 consumers and 560 business leaders who commission or manage contact centres, both in the UK and Ireland, surveyed online in October and November 2025. The report is available as a direct PDF from NTT DATA’s site.

NTT DATA, “Realising the power of AI in contact centres”, 2026
ItemDetail
PublisherNTT DATA
Date2026
Sample1,600 consumers and 560 contact-centre leaders
GeographyUK and Ireland
Core findingBusinesses rate chatbots as more useful than the customers who use them do
ContextSurvey run with AWS; fieldwork by Coleman Parkes, October–November 2025

Source: NTT DATA, 2026. Denominators differ by question; check each figure against its base in the original.

Three caveats apply before any figure from the report is quoted. First, the geography: the UK and Ireland are not the EU, and their regulatory setting differs from the one our clients work under. Second, NTT DATA is a technology services firm and produced the survey in partnership with AWS, so the report sits within a commercial ecosystem that sells contact-centre technology. Third, the report surveys two populations, and the base changes from question to question. A share of consumers and a share of leaders should never be set side by side as if they were one measurement.

The gap, in plain terms#

The report’s central finding is a mismatch. The usefulness that bots are expected or perceived to have is not matched by what customers experience when they actually use them. Put simply, the promise and the call are two different things.

We do not need the report’s exact percentages to recognise the pattern. It is the oldest problem in automated customer contact. A system that tests well in a demonstration, with clear speech, a cooperative caller and a short script, meets real customers who interrupt, mumble, ask two things at once, call from a noisy street or simply do not want to talk. The distance between those two situations is where customer experience is won or lost.

Other research points the same way. Gartner found in 2026 that only 27% of customers would try a chatbot again after a negative experience, in a survey of 3,566 B2B and B2C customers conducted in February and March 2026. A bad first encounter is expensive because it is rarely repeated.

Why the two sides see it differently#

There are ordinary reasons why leaders and customers would read the same bots differently, and none of them requires bad faith. Leaders tend to see dashboards: containment, volumes handled, cost per contact. Customers see a single conversation, usually at a moment when something has gone wrong or when they did not expect a call. A dashboard can look healthy while a meaningful share of individual calls ends in frustration, because averages hide the calls that matter most to the customer.

Take a typical early-collections call. The bank sees a contact attempted, a conversation connected and a status returned. The customer hears a voice that either recognises they paid yesterday, or does not; that either accepts a promise for next Friday, or insists on today; that either lets them ask for a person, or loops. Each of those moments is a design decision, and each is invisible in an aggregate unless someone listens to the call.

Why the gap is an operating problem#

The title of the report talks about realising the power of AI. Our reading is that power without an operating model stays latent. An automated conversation is only as good as the decisions around it: which customers are called, at what hour, from which number, with which voice, after which SMS, with how many attempts, and what happens when the conversation should stop.

Most of those decisions are not made by the model. They are made by people who design the process, listen to calls and change what does not work. When nobody owns that work, the gap between perceived and actual usefulness opens by default, because nobody is watching the actual.

Companies have learned this lesson many times: they buy good software and it stays in a drawer, because nobody can design the process on their own. – Dawid Wójcicki, Founder & CEO of Vee

This is also why we built Vee as a managed service rather than a platform. We design, launch, run and keep improving the process, and we answer for it. Conversations are executed by Vee’s AI; Vee employs no call agents. The people at Vee design, listen, analyse and improve.

What closes the gap in production#

Over tens of millions of cases we have found a short list of practices that keep the experience close to the design. None of them is specific to one sector, and all of them apply with particular force in banking, insurance, healthcare, telecoms and energy.

  • Voice chosen per process. In cold contact, studio recordings often convert better than synthesis; we choose the best speech engine per language rather than per vendor.
  • Listening before launch. The diagnostic includes listening to the client’s own call recordings, so the design starts from real customers rather than an idealised script.
  • A pilot on a sample. The first production run covers part of the base, never the whole database, with Vee Analytics live from the first hour.
  • Daily human listening. In every project our QA team listens to selected calls each day, especially unusual ones.
  • One change, every call. Because we do not manage a floor of agents, an approved change reaches every live conversation at once.
  • A route out. Agreed transfers to the client’s staff when the customer asks for a person or the situation requires judgement.

The measures we use to check the experience start with the customer’s behaviour, not with the system’s confidence. Did we reach the customer? Did they continue after hearing the company name, or hang up? Did we obtain the key information? Did the business effect follow? That four-rung ladder makes the gap visible, because a hang-up rate after the greeting is the customer’s verdict on the experience in the first few seconds.

38,325,799Connected AI conversationsConversations in which the customer was connected to Vee’s AI.Cumulative to 31 October 2026, Vee operating data.

Where we read the report differently#

Studies of this kind often frame the answer as better AI. Our experience suggests that the larger gains come from the operating side: who is called, when and how, and what happens next. The model is a tool, not a solution. In regulated processes we keep it on a short leash: replies come from a closed, versioned library, and there is no learning in production.

We would also caution against reading consumer scepticism as a reason to avoid automation. The more precise reading is that customers judge automation by its worst moments. A process designed around exceptions, with a clear disclosure that the caller is speaking to AI and a clear route to a person, earns more tolerance than one designed around the happy path. Since 2 August 2026 that disclosure is also a legal matter in the EU under Article 50 of the AI Act.

What to ask before you deploy#

  1. Who owns the result after go-live: a supplier, an internal team, or nobody in particular?
  2. How often are real calls listened to, by whom, and how quickly can a fix go live?
  3. What share of customers hang up in the first seconds, and how is that tracked?
  4. Is the first run a pilot on a sample with a comparator, or the whole base at once?
  5. Which exceptions lead to a person, and has each one been tested?

NTT DATA’s study describes the gap honestly. The way to close it is not another round of technology selection but an operating model that watches the real call every day. That is the service we run for regulated enterprises.

Follow The First Second

Occasional issues on the mechanics and economics of AI-powered customer contact. No noise, no vendor news.

Follow via RSS