AI in customer service is reshaping jobs far more often than it is removing them, and some of the organisations that removed them are already hiring back. That is what two Gartner surveys published in 2026 say when you read them together, and it matches what we have seen in telephone contact for more than a decade.

The practical conclusion is not “keep everyone” or “automate everything”. It is that the division of work between the AI and people has to be designed at the start of a programme, with the same care as the dialogue and the integration. Cutting first and redesigning later is the expensive order.

What Gartner found#

In November 2026 Gartner published a survey of 275 customer service and support leaders, conducted in July and August 2026. Only 21% said they had reduced human-agent headcount because of AI. Among those who had, the average reduction was 16.7%. And 7% of leaders said they had already rehired.

In April 2026 Gartner had published a separate survey of 321 leaders, carried out in September and October 2025. It found that 85% were expanding the responsibilities of their human agents, despite widespread expectations of mass AI layoffs. In the same survey 31% said they had made, or planned, AI-linked layoffs by the first quarter of 2027.

Two Gartner surveys of service and support leaders
FindingValueSurvey
Reduced human-agent headcount because of AI21%275 leaders, Jul–Aug 2026
Average reduction among those who reduced16.7%Same survey, reducers only
Already rehired7%Same survey
Expanding human agents' responsibilities85%321 leaders, Sep–Oct 2025
Made or planned AI-linked layoffs by Q1 202731%Same survey

Gartner press releases of 5 Nov 2026 and 28 Apr 2026. Self-reported survey answers; the two samples differ and should not be combined.

Read the numbers carefully. The 16.7% applies only to the minority that cut, not to the whole sample. The two surveys have different samples and dates, so the 85% and the 21% cannot be added or subtracted. And these are declarations by leaders, not payroll data. Even so, the direction is consistent: AI has mostly redistributed service work, and where it removed people quickly, some of that work came back.

Why the work comes back#

The surveys tell us what happened, not why. From our side of the table, having run millions of automated cases in banking, insurance, healthcare and energy, three patterns explain most of the rehiring we hear about.

The exceptions were never counted#

A business case for automation usually counts the average call. The real workload sits in the exceptions: the customer in financial hardship, the disputed debt, the bereaved relative, the complaint that turns into a legal matter, the request to erase personal data. These cases are rarer, longer and more sensitive. When the routine work is automated, they do not disappear; they arrive at a smaller team, without context and often without a defined route.

Automation creates new work#

A process that reaches more customers, faster, generates more of what follows: more interested leads to call back, more promises to monitor, more rescheduled appointments, more data to review. If nobody planned capacity for that downstream work, it piles up, and the organisation hires to clear it.

Someone has to own the machine#

An automated process needs people who design dialogues, listen to calls, approve content changes, analyse results and decide on the next base. When that role is not defined, it falls to whoever is left, usually on top of their old job. The 85% expanding agents' responsibilities are, at least in part, discovering this.

How we divide the work#

Vee employs no call agents. The conversations in our processes are executed by Vee's AI, and the people at Vee design, listen, analyse and improve. That does not mean people disappear from the process. It means the division of roles is explicit, and it is written down before launch.

A typical division of roles in an automated telephone process
WhoWhat they do
Vee's AIReaches the customer, verifies identity where required, runs the conversation from an approved script, records the outcome, schedules retries and SMS, and transfers when a stop condition is met.
The client's own staffHandle transferred cases: hardship, disputes, complaints, sales closing, anything that needs judgement or authority to decide.
The client's analysts and process ownerDecide allocation: which accounts, buckets or leads go to the process and when; approve content; own the budget.
Vee's teamDesigns dialogues, listens to calls daily, analyses results, runs tests, proposes changes and fixes its own mistakes.

Illustrative division based on our operating model; the exact split is agreed per process.

Each row has a reason. The AI takes the work where consistency and persistence matter most: the hundredth call of the day sounds like the first, and a retry happens when the data says it should, not when someone has time. People take the work where judgement and authority matter: deciding on a repayment plan, settling a complaint, closing a sale.

Designing the human route#

The handover to a person is not a fallback; it is part of the product. In banking, for example, we design exception paths before go-live: the customer asks for a person, identity verification fails, recording consent is withdrawn, a death or legal dispute is mentioned, an erasure request is made, or financial hardship is described. Each leads to a defined action: a transfer, a stop, or a referral.

  • Stop conditions are explicit. The AI does not improvise in sensitive territory; it recognises the situation and follows the agreed route.
  • Transfers carry context. The person receiving the call should not have to ask again what the customer already said.
  • Transfers are paced to capacity. In sales processes we time live transfers to the hours and capacity of the receiving team, so interested customers are not left in a queue.
  • Outcomes come back. What the person decided is recorded against the case, so the next contact knows.

This is what makes it possible to automate a large share of the volume without leaving customers stranded. It is also what stops an organisation from discovering, six months after a headcount cut, that nobody can handle the cases the AI correctly refused to handle.

We have seen this cycle before#

The fear that automation will empty the contact centre is older than generative AI. In 2017 a Polish business daily ran a headline saying our company would lay off consultants. What actually happened in our early projects was different. An information line we launched in one of those early projects was designed with a route to a consultant from the first day, because some questions should never be answered by a machine alone.

The pandemic in March 2020 showed the other side. Demand from our large clients rose sharply within weeks, and we scaled by adding telephony capacity, not by recruiting and training people. Scale in an automated process is an infrastructure question. Judgement is still a people question. Both lessons point to the same design rule: decide in advance which work belongs where.

What has changed since then is the breadth of what the AI can handle reliably, not the principle. Each generation of technology moved the line between routine and exception a little further. None of them removed the line.

What changes for the people who stay#

In our view, the 85% finding is the more important of the two. Agents' work is becoming broader and more demanding, because the simple cases go first. The people who remain in service teams handle a higher share of difficult conversations, and they need more authority, better information and different training to do it.

There is also new work around the automation itself: deciding which bases to send, reviewing outcomes, approving content, listening. Some of it sits with the supplier, some with the client. In our model the analysis and continuous improvement sit with us, while allocation strategy and final approvals stay with the client. Being clear about that boundary is what prevents duplicated work on one side and a vacuum on the other.

For a head of collections or customer service this has a budgeting consequence. The saving from automation is real, but part of it should be reinvested in the people who handle exceptions and in the ownership of the automated process. In our experience, organisations that treat the whole saving as headcount reduction tend to find the cost again later, in rework, complaints and rehiring. Gartner does not say why the 7% rehired, but their answer is consistent with that pattern.

We are not a software house. We don't build a process, switch it on and say: now it will fly by itself. – Dawid Wójcicki, Founder & CEO of Vee

Questions to settle before you cut anything#

  1. Which conversations will the AI handle end to end, and which will it transfer? Write the stop conditions down.
  2. What volume of exceptions do you expect, based on your own recordings rather than the average call?
  3. Who receives transfers, during which hours, and with what information on screen?
  4. What new work will automation create downstream (callbacks, promises to monitor, rebookings), and who will do it?
  5. Who owns the automated process day to day: content changes, listening, analysis, the next base?
  6. How will you measure the result: outcomes in your systems, or minutes and calls?

If a supplier cannot help you answer these, the risk is that you join the 7%. If they can, automation becomes what it should be: a way to move repetitive work to a system that does it consistently, and to give people the cases that need them. That is how we have built processes since 2012, and the division of roles is one of the first documents we write with a new client.

Follow The First Second

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

Follow via RSS