McKinsey’s survey of 440 customer-care leaders says that organisations further along with AI report much better customer-experience results than those at the start. The more useful finding sits next to that headline: a large share of the laggards do not trust AI to handle an interaction end to end, and that distrust is the thing to work on.

We read the study as practitioners who have run automated telephone contact for banks, insurers, healthcare providers, telecoms and utilities since 2012. Below we set out what McKinsey reports, how far the numbers can be pushed, and where they match what we see in production.

What McKinsey published#

The article is titled “Building trust: How customer care leaders pull ahead with AI” and was published by McKinsey on 23 February 2026. It rests on McKinsey’s State of Customer Care Survey 2025, with 440 customer-care leaders responding. McKinsey groups respondents by maturity, with leaders as the top 10% and laggards as the bottom 30%, and compares the groups on the outcomes they report.

Two figures carry the piece. Among the leaders, 40% report significantly improved customer-experience scores over the past 12 months, against 12% of the laggards. And 37% of the laggards say they are not at all comfortable with AI handling interactions end to end, against 4% of leaders.

McKinsey, “Building trust”, 23 February 2026: the two headline figures
FindingFigurePopulation
Report significantly improved CX scores in the past 12 months40% vs 12%Leaders (top 10%) vs laggards (bottom 30%)
Not at all comfortable with AI handling interactions end to end37%Laggards
Sample440Customer-care leaders, international

Source: McKinsey, 23 February 2026, State of Customer Care Survey 2025. Self-reported survey answers; groups defined by McKinsey’s own maturity methodology.

How far the numbers go#

This is a survey of leaders describing their own organisations. The 40% and the 12% are declarations, not measured customer outcomes, and the respondents who report gains are the same people who led the programmes. That does not make the answers worthless. It makes them self-assessment.

The comparison is correlational. Mature organisations report better results, but the study does not show that maturity caused the improvement. Organisations with better data, cleaner processes and bigger budgets are likely to be both more mature and more successful, and a survey cannot separate the two.

The groups themselves come from McKinsey’s methodology. A reader should not carry “leader” or “laggard” over to another study’s definition, and should not add the 37% to figures from other surveys with different populations. The study covers customer care in general, not telephone contact in particular.

Trust is the finding that matters#

The 37% figure is the one we would put in front of a steering committee. It names a reason why organisations stall: not the absence of models, but the absence of confidence that an automated conversation will do the right thing every time without someone watching.

In regulated customer contact that doubt is reasonable. A bank calling a customer in early delinquency cannot accept a conversation that discloses debt details before identity is verified, invents a repayment term or forgets that the customer asked for a person. An insurer or a healthcare provider faces the same exposure with different words. Trust in that setting is not a feeling. It is evidence that the process behaves as designed and can be shown to have done so.

That is why we read “building trust” as an operating discipline rather than a communications exercise. The laggards in McKinsey’s sample are not wrong to be cautious. In our view, they are missing the controls that would let them stop being cautious.

How trust is built in production#

Our own route to trust has been built over tens of millions of cases. By 31 October 2026 we had handled 48,626,959 cases cumulatively (Vee operating data; a case is a record assigned for handling, not a person or a call), with every conversation executed by Vee’s AI. Vee employs no call agents. What made that scale possible in regulated sectors was a set of controls, not a single model.

  • A closed library of replies. In regulated processes our replies come from a closed, versioned library. Language models may help classify what the customer said; they do not generate what the bank says back.
  • No learning in production. The system does not change its behaviour on its own between approved versions.
  • Full reconstruction. Every call can be rebuilt from its recording, logs, statuses and the script version in force at the time.
  • Daily human listening. Our QA team listens to selected calls in every project every day, especially unusual ones.
  • A designed human route. Agreed transfers to the client’s own staff for requests for a person, failed verification, disputes, hardship and other exceptions.
  • Versioned, approved changes. Every content change in a regulated set-up is versioned and approved by the client before it goes live.

Every day, in every project, our people listen to selected calls, especially unusual ones, to hear with their own ears what technology still cannot. – Dawid Wójcicki, Founder & CEO of Vee

None of these controls is exotic. Together they turn “do we trust the AI end to end?” into a narrower and answerable question: do we trust this process, with these boundaries, these exceptions and this evidence trail? In our experience, risk and compliance teams can answer that one.

Where the study and our experience agree#

McKinsey’s leaders report better customer experience. We would add that the gains we can show come with scope attached, because one number never travels across processes. In soft collections for an insurer, 61.9% of contacted cases in scope were closed at the first successful contact (FCCR, historical). In inbound FAQ for a mobile-payments provider, 73.5% of contacts were resolved first time (FCR, historical). Neither figure is a market benchmark. Each is the result of one designed process, measured on the client’s terms.

61.9%First contact closure rate (FCCR)Cases closed at the first successful contact, divided by contacted cases in scope.Scope: soft collections, insurance, historical.

We also agree that maturity is not bought. It accumulates through a delivery path: a diagnostic that includes listening to the client’s own call recordings, a three-part specification of dialogue, integration and telephony, a controlled pilot on a sample rather than the whole base, then managed production with an intensive, weekly and later monthly rhythm. Each step produces evidence that the next step relies on.

Where we would push back#

Surveys of leaders describe customer experience from the inside. In telephone contact we prefer measures that start with the customer’s behaviour: whether we reached them, whether they stayed on the line after hearing the company name, whether we obtained the key information and whether the business effect followed. A leader’s sense of “significant improvement” is useful context, but it does not replace a comparator on the same portfolios and buckets.

We would also be careful with the idea that trust grows simply with maturity. In our experience it grows with transparency to the customer. Vee always introduces itself as AI, and since 2 August 2026 the transparency obligations of Article 50 of the AI Act have applied to AI interactions, with Commission guidelines published on 20 July 2026. We help clients assess their own configuration against those rules; we do not certify it.

There is a further point about who is being asked. A customer-care leader answers for the whole function: chat, e-mail, self-service, human teams and voice. Telephone contact behaves differently from the rest. Someone who rings a company will usually tolerate a weak menu to get an answer; someone the company rings may not pick up, and may hang up within seconds of hearing the company name. Trust has to be earned in those first seconds, in a channel where the customer did not choose the timing. That is a narrower and harder test than a satisfaction score across channels, and it is the one our clients care about most.

What to ask before you trust a supplier end to end#

  1. Which replies can the system generate freely, and which come from an approved, versioned library?
  2. Can you reconstruct any single call, with recording, logs, statuses and the script version in force?
  3. Who listens to real calls, how often, and what happens to what they hear?
  4. Which exceptions trigger a transfer to our staff, and how is each one tested before launch?
  5. What is the comparator for the results you report, and on which buckets or portfolios?
  6. Who approves a content change, and how fast does it reach every live conversation?

McKinsey’s survey is a fair reminder that trust, not technology, is where many customer-care programmes stall. The answer is not to trust harder. It is to run the service so that trust can be checked, call by call. That is the work we do for regulated enterprises every day.

Follow The First Second

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

Follow via RSS