.png)
Conversational AI, the umbrella term for chat, voice, and messaging systems built on natural language processing and machine learning, has moved well past the pilot stage in banking. By January 2025, only 37% of surveyed US banking customers said they had never interacted with a bank's chatbot, according to Deloitte's Consumer Banking Survey, meaning roughly two-thirds already had. Adoption was never really the open question in banking's conversational AI story. The CFPB's own review found that among the top ten commercial banks in the country, every one already has some form of chatbot in production.
What most of those interactions actually delivered turned out to be the harder one. Capgemini's most recent World Retail Banking Report found that 61% of bank customers who tried a chatbot ended up contacting a human anyway because they weren't happy with the answer, and more than six in ten rated the whole experience as merely average. Call abandonment, meanwhile, is climbing at Tier I and Tier II banks alike.
That's not a technology failure so much as a deployment one, and it's one of the clearest tests of whether a bank's broader digital transformation and artificial intelligence investment is actually changing how service gets delivered, not just adding a new front end. Call it conversational banking, virtual agents, or simply a smarter contact centre: the seven use cases below, spanning volume deflection, escalation design, next-best-action, channel coverage, audit readiness, fraud detection, and staff-side productivity, are the difference between a chatbot that earns a permanent place in the banking industry's service mix and one that just adds a queue in front of the queue.
Absorbing routine volume is the use case that pays for itself first
The highest-volume, lowest-complexity queries are still the fastest return on any conversational AI deployment.
PostFinance's contact centre handled 240,000+ agent conversations in 2023, run alongside 1 million virtual agent interactions covering the routine volume: balance checks, transaction history lookups, password resets, and the like. That ratio is the point: the bot isn't competing with the agent for the same customer inquiries, it's clearing the ones that never needed a person attached in the first place, while cutting response times on the ones it does handle instantly.
The same pattern shows up industry-wide. At ING in the Netherlands, a classic rule-based chatbot resolved 40 to 45% of the bank's 85,000 weekly customer conversations before the bank rebuilt it with generative AI. The remaining volume, more than 16,500 conversations every week, had nowhere to go but a live agent. The value of this use case has nothing to do with how clever the bot's answers sound. It's about which conversations never need a human at all, freeing exactly the capacity that use case two depends on.
Most banks haven't actually captured that value yet, which is worth naming directly. Capgemini's research found that only 6% of retail banks have built an enterprise-wide roadmap for scaling AI-driven transformation, against a majority still running disconnected pilots. PostFinance's deployment is a useful reference point precisely because volume deflection isn't a pilot there; it's been running long enough, and at high enough scale, to show up as a structural change in how the bank's customer service operation is staffed.
The real value is in the escalation, not the deflection
A chatbot that can't hand off cleanly creates more risk than it removes.
Financial institutions should avoid using chatbots as their primary service channel when it's reasonably clear the bot can't meet the customer's need, the CFPB warned in its 2023 review, after a rising share of complaints described customers stuck in unresolved loops with no route to a person.
Swiss Post's deployment is a clean counter-example. Its bot service rate climbed from 55% to 72% within roughly 18 months, achieved not by making the bot smarter in isolation but by running it alongside Live Chat and WhatsApp with a built-in human handoff. The chatbot doesn't try to be the whole answer. It's one stage in a conversation that a person can pick up without the customer starting over. That design choice, not the underlying model, is what sets the ceiling on how much volume a bank can safely automate.
There's a trust and customer experience dimension here too, and it compounds over repeated bad experiences. Capgemini's research found that 17% of bank customers simply distrust chatbots outright and prefer a human agent, independent of how well the bot actually performs. A clean, visible escalation path doesn't just resolve the immediate query for that segment of customers. It's what keeps customer engagement intact after a bad first attempt, willing to try the bot again next time, rather than routing around it to a call centre by habit.
A service conversation becomes the moment for the next best action
The same conversation that resolves a query is also the best-timed moment for a human advisor to make a relevant offer.
At BPER Banca, average handling time on Live Chat and AI Chatbot conversations dropped 20%, with 700 advisors working from Sidekick, integrated into the bank's Salesforce CRM, alongside bank-approved knowledge documents. A shorter, better-informed conversation doesn't just close faster. It leaves room, inside the same interaction, for the advisor rather than the bot to raise the next relevant product, a personalised recommendation timed to the conversation rather than a generic campaign, while the context is still live.
That sequencing matters more than the automation itself. The bot's job is to clear the informational overhead; the advisor's job, freed of that overhead, is the judgment call a next-best-action moment actually requires. That kind of integration, CRM and core banking systems data feeding the same conversation an advisor is already having, is what separates a genuine next-best-action prompt from a generic cross-sell banner.
This is consistent with where Capgemini sees the wider market heading: banks pairing advanced customer data with AI to deliver more tailored offers are seeing that personalisation translate directly into revenue growth, not just satisfaction scores, as competition for wallet share intensifies. The mechanism is the same one BPER's numbers point to. A faster, well-informed conversation isn't just cheaper to run. It's the moment a bank has the most context and the most attention, which makes it the wrong moment to waste on a query the bot could have handled thirty seconds earlier.
Customers pick the channel; conversational AI has to follow
Deploying conversational AI on one channel while customers live on three others defeats the purpose.
Swiss Post now handles 200,000+ customer support inquiries a month across chatbot, Live Chat, and WhatsApp, with WhatsApp alone logging more than 4,000 conversations within weeks of launch. Twenty percent of inquiries that used to arrive by phone or email now happen over live text channels instead.
None of that shift happened because WhatsApp is a superior channel to phone support. It happened because that's where the customers already were. A conversational AI use case scoped to a single channel is really just a pilot with better PR; the version that changes contact-centre economics is the one built to follow the customer, not the one that waits for the customer to come to it.
Channel coverage cuts both ways, though. The same population that increasingly reaches for WhatsApp or in-app chat still includes the 17% who would rather speak to a person than a bot at all, referenced above. Following the customer means keeping a live agent reachable on those same channels, not just a bot, so that channel choice and escalation design end up solving the same problem from two directions.
The channel list keeps growing, too. McKinsey's research found 23% of consumers already use generative AI for financial tasks at least monthly, a channel that barely existed in most banks' contact strategy two years ago. A conversational AI deployment built to extend across channels, rather than own a single one, is the version that doesn't need rebuilding every time customer behaviour shifts again.
Every AI-assisted conversation still has to survive an audit
A conversational AI deployment that can't produce a clean record is a compliance liability waiting to surface.
The CFPB's 2023 review was explicit that inaccurate or incomplete chatbot responses can constitute a violation of the Consumer Financial Protection Act, regardless of whether anyone at the bank intended the harm. That risk doesn't disappear with a better model; it's a regulatory compliance question as much as a product one. It's addressed at the platform level, through what gets logged, retained, and made producible on request against whatever regulatory requirements apply.
This is where audited certification, not marketing language, does the convincing work, whether the requirement in question is SOC 2, ISO 27001, or PCI DSS for anything touching payment card data. Unblu's platform holds SOC 2 Type 2 certification, audited by BDO against the AICPA Trust Service Criteria, alongside ISO 27001, 27017, and 27018 certification covering cloud application security and data privacy. For a compliance stakeholder evaluating conversational AI against those regulatory requirements, that's the use case underneath the use case: not whether the bot can answer a question, but whether the bank can reconstruct exactly what it said and why, months later, on demand.
The audited scope matters as much as the certificates themselves. Unblu's SOC 2 Type 2 attestation covers the specific infrastructure the conversations run on, Google Cloud Platform and Swiss Open Telekom Cloud deployments, backed by a 99.75% committed SLA on the Unblu Financial Cloud. That's the difference between a vendor citing a certification in the abstract and one whose auditors have actually looked at the environment a bank's customer conversations pass through.
Catching fraud before the customer has to raise it
Fraud detection is one of the use cases bank employees themselves are most enthusiastic about automating, and conversational AI increasingly sits on the front line of it.
Bank employees report more enthusiasm for generative AI copilots automating fraud detection than for almost any other application, according to Capgemini's research, ahead of data visualisation and drafting personalised customer content. In practice, that means the same conversational layer handling a balance enquiry can also surface a fraud alert or security alert the moment a transaction pattern looks wrong, prompting the customer or the advisor before a dispute ever gets raised.
This overlaps deliberately with the audit-readiness use case above. A fraud alert delivered through a logged, auditable channel does two jobs at once: it flags the anomaly, and it creates the record a compliance team would need if that flag turns into a formal dispute later.
The productivity case extends past the customer, to the people behind the bot
The most durable conversational AI use case isn't customer-facing at all. It's what the bot frees the advisor or agent to do next.
McKinsey estimates generative AI could add $200 billion to $340 billion in annual value to banking, equivalent to 9 to 15% of the industry's operating profit, and the bulk of that is projected to come from productivity gains rather than new revenue. Conversational AI's customer-facing use cases are the visible ones. This is where most of the value actually accrues.
In wealth advisory contact centres specifically, Capgemini found conversational AI can cut human-handled calls by 26% and service costs by up to 30% within 12 to 24 months, with one anonymised British investment bank lifting chatbot containment from 30% to 65 to 70% in the process. PostFinance's own deployment reports efficiency gains equivalent to roughly 40 full-time roles, achieved through automation across every channel it runs.
Numbers like that rarely show up on a customer-facing dashboard, which is exactly why this use case gets underweighted in most conversational AI business cases. The bot's most measurable output isn't the conversation it has. It's the conversation it removes from someone else's day, and PostFinance's 40-FTE figure is one of the few places that abstract McKinsey estimate above turns into a number a contact-centre manager can actually plan a headcount around.
What ties these seven together
None of these use cases are really about the bot. They're about where a bank chooses to put the handoff: which conversations it lets go entirely, which ones it hands to a person and when, and what gets logged along the way. Get that sequencing right and a conversational AI deployment compounds value across service cost, advisor time, and audit readiness at once. Get it wrong and it just relocates the same bottleneck one screen earlier, which is exactly the outcome behind Capgemini's 61% figure at the top of this piece.
Not all of this requires moving to a fully autonomous AI agent, either. Each of these seven patterns works with a conversational layer paired with human oversight at the right moment, well short of the agentic AI territory some deployments in the banking industry are now reaching for.
PostFinance, BPER Banca, and Swiss Post are running fairly different deployments, retail volume absorption, advisor-assisted service, and multi-channel orchestration respectively, but each treats the handoff design as the actual product, not an afterthought bolted onto a chatbot. That's the pattern worth copying, more than any specific vendor or model choice.
Unblu supports this across Secure Messenger, Live Chat, and Co-Browsing, deployable on cloud, Swiss sovereign cloud, or on-premise depending on a bank's data residency requirements.
