One in ten financial-services firms has deployed AI agents at scale, according to Capgemini’s World Cloud Report for financial services. The number is low not because the models are missing, but because scale in a bank means integration, control and someone answering for the result every day.

The report is not about contact centres, and we will not pretend it is. But customer contact is one of the places where banks and insurers try to put AI agents to work first, and the obstacles the figure points to are ones we have worked through in production since 2017.

What Capgemini found#

Capgemini published the World Cloud Report – Financial Services 2026 on 12 November 2025. It is based on a survey of 1,100 leaders in the financial-services sector across 14 markets, fielded in June and July 2025. The headline finding we focus on is simple: 10% of financial-services organisations have deployed AI agents at scale.

10%Financial-services firms with AI agents deployed at scaleShare of surveyed organisations reporting AI agents in deployment at scale.Capgemini, World Cloud Report – Financial Services 2026, 12 Nov 2025; survey of 1,100 FS leaders in 14 markets, June–July 2025. Covers all functions, not contact centres specifically.

Three caveats belong next to that number. First, this is a survey of leaders describing their own organisations, not an audit. Second, it covers the whole of financial services and every function, so the 10% is not a measure of AI in customer contact. Third, the summary we worked from does not set out how Capgemini defines “at scale” or “AI agent”; anyone quoting the figure in a board paper should check the report’s own definitions first.

Why scale is hard in a bank#

A pilot is a demonstration that something can work. Scale is a demonstration that it keeps working, across portfolios, months and audits, without anyone hand-holding each case. In financial services the distance between the two is longer than elsewhere, for reasons that have little to do with the intelligence of the model.

Integration#

An agent that cannot read the current balance, write back an outcome or trigger the next step is a conversation without consequences. Other research points the same way. Deloitte’s study of US banks, published on 30 June 2026, found that 77% of the 30 banking executives surveyed cite integrating new technology with other systems as a major modernisation challenge, and 67% cite data security and compliance. That sample is small and US-only and should be read as such, but the pattern matches what we see in European banks.

Control#

A bank must be able to explain what was said to a customer, why, and under which approved wording. That is difficult if replies are generated freely. In regulated processes our dialogue engine answers from a closed, versioned library; large language models, where we use them, classify what the customer means rather than invent what to say back; nothing learns in production. Every call can be reconstructed from recording, logs, statuses and script version.

Accountability#

Somebody must answer for the agent’s behaviour on a Tuesday morning six months after launch. In our experience, many organisations buy a platform, build a first process and then discover that nobody owns the daily work of listening, adjusting and measuring. Scale stalls not at the technology but at the operating model.

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

What scale looks like in customer contact#

We can describe scale in our own field with numbers that carry their scope. Up to 31 October 2026, Vee handled 48,626,959 cases, a case being a record assigned for handling rather than a person or a call, and 38,325,799 connected AI conversations, both cumulative Vee operating data. These conversations are executed by Vee’s AI; Vee employs no call agents, and processes include agreed transfers to the client’s own staff.

Those totals were not reached by switching on an agent across a whole bank. They were reached the slow way: one process or portfolio first, proven on the client’s own metric, then extended to new portfolios, business units and countries. A large international consumer-finance group, for example, has signed further agreements with us for new portfolios. That is what “at scale” tends to mean in practice: not a single launch, but years of expansion that each have to pay back.

Pilot versus scale in regulated customer contact
QuestionIn a pilotAt scale
Who decides which accounts are called?A test listThe bank’s analysts, by portfolio, bucket and customer category
How are replies controlled?Reviewed by the project teamClosed, versioned library; every change approved
What happens at an exception?Handled ad hocDefined paths: identity failure, hardship, dispute, request for a person
How is success judged?Impressions and call samplesComparator on the same buckets; payment, not promise
Who listens to calls?The vendor at launchA quality team every day, in every project

Drawn from Vee’s operating practice in banking and consumer finance.

The path between those two columns is deliberately unglamorous. We start with a diagnostic of goals, data and current results, including listening to the client’s own call recordings, and we say so plainly when a process will not pay back. Then come the contract and data-protection work, a specification in three parts covering dialogue, integration and telecoms, end-to-end tests, and agreed operating rules for call windows, attempts, callbacks and transfers. Only after that does a controlled pilot run on a sample of the base, never the whole of it.

The regulated layer is part of scale, not an obstacle to it#

Financial-services firms carry obligations that do not shrink when a process is automated. Banks audit us as a sub-processor: Vee Center Sp. z o.o. holds ISO/IEC 27001:2022, and the bank’s own GDPR, AI Act and DORA obligations shape how the service is set up. Under Article 50 of the AI Act, whose transparency obligations have applied since 2 August 2026, our conversations always disclose that the customer is speaking with AI. Identity is verified before any debt detail is mentioned, as a platform rule rather than a line in a script.

None of this is decoration. Each item is a reason a bank can say yes to the second portfolio after the first. In our experience, firms that treat control as an afterthought tend to find that their pilot cannot pass the review needed to grow, which is one plausible reading of why so few report AI agents at scale.

Where we would be cautious#

It would be easy to use Capgemini’s figure as proof that the market is behind and that a managed service is the answer. We would not go that far. The report does not measure contact centres, does not show causes, and does not compare operating models. Its value is as a reminder that deployment is not the same as production, and that the gap is wide even in a sector with large technology budgets.

Nor should the other 90% be read as a single group. Some firms will have chosen not to deploy AI agents at all; some will be running pilots that are going well; some will have tried and stopped. A survey headline cannot tell these apart, and a sensible reader will not assume that a firm outside the 10% is behind. In regulated customer contact, a careful decision not to automate a sensitive process is sometimes the right one, and we would rather tell a client that than launch something that cannot be defended in front of an auditor.

It is also worth remembering what an AI agent is for in customer contact. The aim is a measurable business effect, such as a payment arriving on time, a renewal signed or a visit attended. In early-stage collections we have historically seen on-time payment frequency rise by 7–13% versus a human-agent comparator. That is an incremental effect on a defined process, not a claim about agents in general.

Questions to ask before you count yourself among the 10%#

  • Does the agent write back? Ask for case-level statuses delivered daily into your systems, not a dashboard of minutes.
  • Can every conversation be reconstructed? Recording, logs, statuses and the exact script version, available to your auditors.
  • Who owns the process after launch? Name the people who listen to calls and decide changes each week.
  • How are exceptions handled? Ask to see the paths for hardship, disputes, failed verification and a request for a person.
  • What was the comparator? Results without a comparator on the same buckets or portfolios are impressions, not evidence.

One in ten is a low number, but it is an honest one, and it describes a real constraint. The firms that move from pilot to scale will be those that treat AI agents as part of an operated, measured and audited process. That is the kind of process we run for banks and insurers, and the questions above are the ones we expect to be asked.

Follow The First Second

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

Follow via RSS