Capgemini’s study of customer service describes an operation that most executives think is inefficient, most agents do not enjoy, and most agents say generative AI has already lightened. The figures are striking, but they come from three separate groups of respondents, and the gap they reveal is less about technology than about how the work is designed and run.
This article reads the report as published, sets out what each figure can and cannot carry, and then compares it with what we see running telephone customer contact for banks, insurers, healthcare providers and energy companies.
What Capgemini published#
The report is “Unleashing the value of customer service”, published by Capgemini on 18 March 2025, with a public summary page and a downloadable PDF. Its fieldwork covered three populations: around 9,500 consumers, 500 customer-service agents and supervisors, and 1,000 executives, with fieldwork in November and December 2024. It is a survey, not a forecast, and the numbers below are what respondents reported about their own experience and expectations.
| Finding | Figure | Population |
|---|---|---|
| Executives who say their customer service faces low operational efficiency | 65% | About 1,000 executives |
| Agents satisfied with their job | 16% | Front-line agents surveyed |
| Agents who say generative AI has reduced their overall workload | 70% | Front-line agents surveyed |
Capgemini, “Unleashing the value of customer service”, 18 March 2025. Self-reported survey data; the populations differ and their denominators should not be combined.
The report’s own framing is operational: legacy systems, weak integration, customer experience, workforce strain and return on investment run through it. That is why we read it as a study of operating models rather than of AI products.
Three populations, three denominators#
The most important caveat is also the easiest to miss. The 65% comes from executives; the 16% and the 70% come from front-line agents; the consumer sample sits alongside both. These are not three views of one organisation. They are three samples, each answering about its own world.
It is tempting to write a sentence such as “two thirds of firms are inefficient, so their agents are unhappy, so AI will fix it”. The data does not support that chain. Nobody measured whether the agents who are dissatisfied work in the organisations whose managers see low efficiency. Nor does a declared lighter workload tell us anything about cost, quality or customer outcome.
The consumer sample deserves the same discipline. Around 9,500 people answered about their experience as customers, not about any particular operation. Their views tell you what customers expect from service in general; they do not tell you how a given contact centre performs, and they cannot be set against the managers’ or agents’ answers as if the three groups had been asked about the same calls. For a regulated firm, the practical lesson is to collect the equivalent evidence for your own process: what your customers say after a call, measured on your own cases.
Low efficiency is a design problem before it is a staffing problem#
When 65% of executives say their service is not efficient, the usual response is to look for capacity: more agents, more shifts, a new outsourcer, a tool that answers faster. In our experience of regulated customer contact, that rarely reaches the cause. Inefficiency in a collections or renewals operation usually lives upstream of the call: in which accounts are dialled and when, how many attempts a case receives, whether the number shown is one people pick up, and whether the result of the call returns to the client’s systems in a form someone can act on the next morning.
We describe this as a reachability strategy with six parameters: outbound number, destination number, time window, channel and content, attempt intensity and intervals, and persona and voice. None of these is fixed by hiring. All of them are fixed by design, testing and daily analysis, and allocation decisions stay with the client’s own analysts.
Measurement is the second half of the problem. An operation that counts minutes and handled calls can be busy and inefficient at the same time. We use a four-rung ladder instead: reach, continuation, obtaining the key information, and the real business effect such as a payment or an attended visit. If managers cannot see where on that ladder cases are lost, “low efficiency” stays a feeling rather than a finding.
What the 16% says about the job#
Only 16% of the front-line agents surveyed said they were satisfied with their work, against 68% of their managers and supervisors. That is a statement about the job, and it deserves to be taken seriously on its own terms. Repetitive, scripted, high-volume calls with a long tail of refusals are hard work. In surveys we have run, the pattern is familiar: a person loses enthusiasm after many refusals in a row, while an automated conversation starts every call the same way.
It would be wrong to read the figure as an argument that agents should simply be replaced. It is an argument for being deliberate about which conversations people should hold. Our own position, consistent since 2016, is that repetition belongs to the machine and judgement, empathy and complexity belong to people. In our processes that takes a concrete form: agreed transfer rules to the client’s own staff for a customer who asks for a person, a hardship case, a dispute or a failed identity check.
The 70% is a self-report, not a measured saving#
Seven in ten agents say generative AI has reduced their overall workload. That is a meaningful signal about acceptance on the floor: staff are not uniformly hostile to the technology. It is not evidence that generative AI reduces cost or improves outcomes, and it records what agents perceive rather than what anyone measured.
In regulated processes the question is also narrower than “use generative AI or not”. We run several variants of our dialogue engine, from fully deterministic to a large language model used only to classify what the customer means. In regulated processes the replies come from a closed, versioned library, nothing learns in production, and every call can be reconstructed from the recording, logs, statuses and script version. A lighter workload bought at the cost of an answer nobody approved is not a saving in a bank.
We don’t manage people, so we control the service far better: one change and every live conversation already reflects it. – Dawid Wójcicki, Founder & CEO of Vee
How it compares with what we run#
Our model removes one of Capgemini’s three populations from the equation. Vee employs no call agents; the conversations are executed by Vee’s AI, and the people at Vee design dialogues, listen to calls every day, analyse results and run tests. Up to 31 October 2026 we handled 48,626,959 cases, cumulative Vee operating data, where a case is a record assigned for handling rather than a person or a call.
That does not make the efficiency problem disappear. It moves it. Instead of agent satisfaction and shift planning, the constraints become integration with the client’s systems, the quality of the base, the contact strategy and the discipline of measurement against a comparator. Where we have measured incremental effect properly, the gains are specific rather than sweeping: in early-stage collections, historically, we have seen on-time payment frequency rise by 7–13% versus a human-agent comparator on the same kind of portfolios.
The other thing Capgemini’s framing gets right is that value sits in the whole process. Our delivery path starts with a diagnostic, including listening to the client’s own recordings, runs a controlled pilot on a sample rather than the whole base, and only then moves into managed production with weekly and later monthly reviews. Efficiency is earned in that rhythm, not installed.
What to take from it#
Capgemini’s study is useful precisely because it does not pretend the problem is a missing tool. If you are using it to make an internal case, use it accurately and ask the questions it raises:
- Where is efficiency lost? Map your cases against reach, continuation, key information and outcome before deciding what to automate.
- Which conversations should people hold? Write down the transfer rules first, then design the automated path around them.
- What will be measured, against what? Agree a comparator on the same buckets or portfolios, and separate a declaration from a result.
- Who answers for the process? Ask any supplier whether they run and improve the service, or hand over software and leave.
- How are content changes controlled? In a regulated set-up every change should be versioned and approved before it reaches a customer.
Treat the 65%, the 16% and the 70% as three honest signals from three groups, not as a business case. The business case is the one you can measure in your own systems, on your own customers, after a pilot on a sample. That is the work we do with clients, and it is where the gap between ambition and scale actually closes.
Follow The First Second
Occasional issues on the mechanics and economics of AI-powered customer contact. No noise, no vendor news.
Follow via RSS