Two McKinsey publications from 2026 make the same argument from different angles: agentic AI in customer experience scales only when the plumbing around the conversation is designed, not when the model improves. One is a podcast with a technology vendor; the other is an analytical article. They deserve to be read together, and with different levels of confidence.
We run automated telephone contact for regulated enterprises across Europe, so we read both with one question in mind: what do they say about the conditions under which an automated conversation can carry a whole process?
The two pieces and what kind of evidence they are#
| “Beyond the bot” | “Rewiring customer experience” | |
|---|---|---|
| Date | 6 February 2026 | 14 July 2026 |
| Format | Podcast, about nine pages of transcript | Analytical article |
| Main point | Some firms aim to automate up to 70% of contacts; APIs and UX design limit scale | 41% of customer-facing AI use cases in the cited research are already scaled; scale needs decision orchestration, handovers and shared context |
| Evidence type | Expert interview with a representative of a voice-AI vendor | Analysis drawing on cited research |
| Main caveat | Not a representative survey | Covers all customer-facing functions, not only telephone contact |
Sources: McKinsey, 6 February 2026 and 14 July 2026. Summarised from the publications; read the originals for full context.
The distinction matters. The podcast is a conversation in which a guest from a company that sells voice-AI agents describes what clients are aiming for. It is informed and worth hearing, but it is a practitioner’s view with a commercial interest, not a measurement. The July article is McKinsey’s own analysis and carries a figure from research it cites. Neither is a survey of telephone contact centres.
“Beyond the bot”: ambition and its limits#
The podcast’s headline is ambition: some organisations are on the way to automating as much as 70% of their customer contacts. Read carefully, that is a direction some firms are travelling in, not a result that has been observed, and not an average. It tells you about intentions in the market, not about outcomes.
There is also a question of denominator. Seventy per cent of which contacts? Inbound questions about an order or a balance are a different population from outbound calls to customers who did not expect to hear from the company and may not want to. In our experience the outbound side is harder: someone who calls you will cope even with a clumsy menu, while someone you call may not pick up at all. An automation rate means little until the process, the direction of the call and the comparator are stated next to it.
The more durable part of the conversation is about what stops those ambitions. The vendor guest names willingness to change as the limiting factor, and two practical constraints come up: the availability of APIs into the systems the agent needs to act on, and the design of the user experience. Both are unglamorous. Neither is solved by a better language model.
We agree with the diagnosis. An automated voice that cannot read the customer’s current balance, write back a promise to pay or check a calendar slot is a polite answering machine. And a conversation that is technically correct but tiring to follow loses the customer before the integration ever matters.
“Rewiring”: what 41% does and does not mean#
The July article reports that 41% of customer-facing AI use cases in the research it cites have already been scaled. It then argues that further scale depends on three capabilities: orchestration of decisions across steps, well-designed handovers, and context shared between the agent, the systems and the people involved.
The 41% covers customer-facing functions in general: marketing, sales and service, across channels. It is not a figure for telephone contact, and certainly not for regulated outbound processes such as collections. It would be a misreading to say that 41% of contact-centre automation is at scale.
The three capabilities, however, translate directly to the phone. Orchestration decides what happens after the call: a status sent back, a retry scheduled, an SMS triggered, a case closed. Handovers decide what happens when the conversation should not continue with the AI. Shared context decides whether the customer has to repeat themselves.
Why the plumbing decides on the phone#
In telephone contact the integration problem is sharper than in chat, because the call happens in real time and cannot wait for a back-office queue. If a customer agrees to pay on Friday, that promise must reach the bank’s systems in a form its collections team can act on the next morning. If a patient cancels a visit, the slot must be released early enough for someone else to take it. If a meter reading appointment is confirmed, the technician’s schedule must reflect it before anyone drives out.
That is why our specification before any launch has three parts, not one: the dialogue, the integration and the telecoms. Each is tested on its own and then end to end, often with a friends-and-family round before real customers are called, and the first production run is a controlled pilot on a sample of the base rather than the whole database. The model matters, but it is rarely what fails. What fails is a status that never arrives, a transfer that rings into an empty team or a context field nobody populated.
Empathy is designed, not generated#
The podcast’s title talks about empathetic customer experiences. In regulated contact we would put it more narrowly. Empathy on the phone is mostly a matter of timing, tone, pace and memory: calling at a reasonable hour, not reading out a debt before identity is confirmed, acknowledging that a promise was broken last month without lecturing, and letting someone in hardship reach a person.
These are design decisions. In regulated processes our replies come from a closed, versioned library; language models may help classify intent, but they do not invent what a bank or an insurer says. Separate dialogue paths run in parallel for different situations, and exception paths lead to a transfer, a stop or a referral. That is how a conversation can be considerate and auditable at the same time.
Building a voicebot is not the same as serving a customer. Between the two lies a very long road. – Dawid Wójcicki, Founder & CEO of Vee
What we see in production#
Our cumulative operating data to 31 October 2026 shows 38,325,799 connected AI conversations and 48,626,959 cases (a case is a record assigned for handling, not a person or a call). Those conversations ran inside processes we designed end to end, and that is where McKinsey’s three capabilities show up in practice.
- Orchestration. The contact strategy around the call, including scheduling, retries and text messages, is run by us as part of the service, alongside the conversation itself.
- Handovers. Agreed transfers to the client’s own staff, paced to the team’s capacity and hours, for requests for a person, failed verification, disputes or hardship.
- Shared context. Delinquency stage shapes the base conversation and behaviour history adjusts it; daily case-level statuses go back in the client’s own terms.
- Integration. Data exchange by REST API, SFTP flat files or HTTPS, specified before launch together with dialogue and telephony.
Where we would challenge the framing is the word “agentic”. In regulated sectors the useful question is not how much an agent may decide on its own, but which decisions it is allowed to take, inside which boundaries, and with what evidence afterwards. Allocation strategy, meaning which accounts are contacted and when, stays with the client’s analysts. We run the contact; we do not take over the client’s credit policy.
What to ask a supplier#
- Which systems will the conversation read from and write to, through which interfaces, and who builds them?
- What happens after the call ends: which statuses, retries, messages and closures are triggered automatically?
- Which situations end the automated conversation, and where does the customer go next?
- What context does the person receiving a transfer see?
- Which replies are generated and which are drawn from an approved library?
- When you quote an automation rate, is it a target, a forecast or a measured result, and on which process?
McKinsey’s two pieces are right that the bot is the easy part. The 70% is an ambition and the 41% belongs to a wider population, but the lesson about orchestration, handovers and context holds firmly on the phone. That is the work we have been doing for regulated enterprises for years, and it is where we would start any conversation about scale.
Follow The First Second
Occasional issues on the mechanics and economics of AI-powered customer contact. No noise, no vendor news.
Follow via RSS