Last month, two significant acquisitions landed within eight days of each other. Salesforce announced it would acquire Fin - formerly Intercom - for $3.6 billion. Next, Backbase acquired Kasisto. ServiceNow had made the same bet six months earlier, completing its acquisition of Moveworks for $2.4 billion.
Three major technology companies, six months, one shared conviction: whoever controls the layer where intent becomes action controls the future of banking.
Aurelie L'Hostis at Forrester named this in a piece published following the Kasisto acquisition news: technology providers are racing to control what she calls the "agent runtime" - the layer that interprets intent, determines next best actions, and coordinates workflows across systems.
Her conclusion: conversational AI is no longer a front-end capability. It is a core execution layer.
That framing changes how banks need to think about what they are building right now.
Conversational AI vs. chatbots
Conversational AI in banking is not a chatbot. The distinction matters and it is worth stating plainly before going further.
A chatbot responds. It answers a question, retrieves information, and presents options. The customer then decides what to do and navigates the bank's systems to do it. The chatbot is a search interface with a friendlier face.
Conversational AI - specifically the agentic kind now being deployed by leading banks - acts. A customer states an intent in natural language. The system understands it, checks the customer's state against the bank's policies, and executes the work. The card gets blocked. The transfer goes through. The dispute gets filed. No navigation or human-in-the loop is required for routine actions.
That shift - from a system that responds to a system that acts - is what the current wave of acquisitions and deployments is actually about. The front door of banking is changing from an app you navigate to a conversation that executes.
From recommendations to actions
For the past few years, the dominant model of AI in banking has been advisory. It surfaces insights, recommends the next best actions and flags anomalies. A human then decides what to do and navigates the systems to do it.
That model is giving way to AI that does not just recommend, but also act. Take for example a customer expresses intent in natural language. The system understands it, checks the customer's state, applies the right policies, and executes. The card gets blocked. The product gets opened. The dispute gets filed. The work happens.
I have been watching this transition accelerate across the industry over the past six months, and the evidence is no longer confined to announcements. Banks are in production.
- Starling Bank launched what it describes as the UK's first agentic AI financial assistant in March 2026 - an in-app tool called Starling Assistant that carries out banking tasks on a customer's behalf, from setting up savings goals to organising bill payments, through voice or text.
- Santander and Mastercard completed Europe's first live end-to-end payment executed by an AI agent in March 2026, with Mastercard Agent Pay integrating AI agents into the payment flow as visible, governed participants.
- OCBC unveiled OCBC WoW, what it describes as Southeast Asia's first AI-native mobile banking app, currently in beta for wealth clients - featuring two AI avatars, Wendy and Wayne, modelled after the bank's own staff and trained on its proprietary research, offering personalized wealth management advice around the clock through voice or text.
- Citi Wealth unveiled Citi Sky in April 2026 - an AI avatar that engages Citigold clients directly through real-time voice and video conversation, working alongside financial advisors.
The shift from AI that advises to AI that acts is already underway, and it is happening faster than most banking leadership teams have acknowledged internally.
Why the acquisitions signal something structural
The Salesforce, ServiceNow, and Backbase moves are worth examining together because they reveal different strategic bets on the same underlying thesis.
Salesforce's $3.6 billion acquisition of Fin brings an AI agent that resolves customer queries end-to-end across live chat, email, WhatsApp, SMS, phone, and Slack into the Agentforce ecosystem. It is a horizontal bet: an agent runtime built to work across industries and use cases.
ServiceNow's $2.85 billion acquisition of Moveworks was the earliest of the three moves. Moveworks lets employees get things done across IT, HR, and operations through natural language rather than navigating portals and ticketing systems. A request becomes a completed action.
The Backbase acquisition of Kasisto is a different kind of bet - not horizontal, but vertical. Deeply, specifically vertical.
Kasisto has spent ten years building the infrastructure that connects natural language intent to govern execution inside a regulated bank. It is not a general-purpose agent. Instead, it is an agent runtime built from the ground up for the compliance requirements, the audit trails, the policy constraints, and the customer state complexity that banking specifically demands. Dozens of banks are already running their conversational and agentic capabilities on that technology.
That distinction matters more than it might appear. A general-purpose conversational layer can answer a question and draft a response. What it cannot do - not safely, not in a regulated environment - is execute work on a customer's behalf, inside a bank, with a full audit trail, under the right policy constraints, drawing from a unified view of who that customer is and what they are entitled to do.
That is an architectural problem. It requires an architectural solution.
The execution layer is the strategy
The LLM is not the differentiator. Open-source, open-weighted models are closing the quality gap with frontier labs faster than most executives have internalized. Within six to twelve months, model quality will be effectively commoditized - every bank will have access to roughly equivalent intelligence at declining cost. I have been making this argument publicly for a few months now and I am becoming more, not less, convinced of it.
What will not be commoditized is the execution layer underneath the model, i.e.:
- The unified view of the customer.
- The policy engine that governs what any actor - human, AI agent, or digital channel - is entitled to do.
- The audit infrastructure that makes every action explainable and traceable.
- The orchestration layer that takes intent and turns it into completed work across the bank's systems.
That is what the race to find the new front door is really about - not who has the best conversational interface, but who has built the foundation that makes conversational AI safe, accountable, and genuinely operational inside a regulated bank - regardless of which model is running on top of it.
The banks building that foundation are not just preparing for the next product cycle. They are defining the operating model that will govern how their entire frontline works for the next decade.
What this means in practice
The temptation is to treat the conversational AI shift as a feature decision. Pick a vendor. Deploy an assistant. Add it to the app. Announce it.
That is the wrong frame. The banks that approach it that way will find themselves in the same position as the institutions that treated mobile as a UI project - capable of the surface, unable to capitalize on what was underneath it.
The right frame is an operating model decision, which does not look at which conversational interface to deploy, but how the conversational interface connects to the rest of how the bank operates. Important questions include:
- Does the conversation share the same customer state as the advisor workspace?
- Does it apply the same policies as every other channel?
- Does every action it takes produce a traceable, auditable record?
- Does it draw from a unified view of the customer that updates in real time, regardless of which surface last touched them?
If the answer to any of those is no, the bank has a conversational feature. It does not have a conversational operating model.
This is precisely what the Unified Frontline is designed to address - the operating model where customers, employees, and AI agents work as one because they all operate from the same foundation.
The conversational interface is one surface of that model, not a standalone capability. When it is connected to the same customer state as every other surface, every conversation compounds the relationship and the bank's ability to act. When it is not, it adds a new seam to an existing architecture of seams.
The window to act
The institutions moving fastest are not always the largest. Starling and Discovery Bank are not the biggest banks in their markets. They are among the most architecturally prepared - they built the foundation that makes this kind of deployment possible without years of integration work.
The larger institutions are moving too - Citizens, Citi, HSBC - but the race is not between large and small. It is between architecturally unified and architecturally fragmented. That is a race that does not respect size or legacy market position.
The new front door is being built right now. The banks that understand it as an operating model decision, not a feature decision, are the ones that will own it.
Frequently asked questions
What is the new front door in banking?
The new front door in banking is the shift from the mobile app as the primary customer interface to conversational AI that executes work on a customer's behalf. For the past decade, the mobile app was how customers accessed their bank. The emerging model replaces menu navigation with natural language - a customer states what they want, and the AI understands the intent, applies the bank's policies, and completes the action. Banks including Starling, OCBC, and Discovery Bank have already deployed early versions of this model. The race among technology companies to acquire conversational AI infrastructure - Salesforce acquiring Fin, ServiceNow acquiring Moveworks, Backbase acquiring Kasisto - reflects a shared conviction that controlling this layer is the defining competitive move of the next five years.
What is the difference between conversational AI and a chatbot in banking?
A chatbot responds to a customer's question and presents options. The customer still has to decide what to do and complete the action themselves. Conversational AI - specifically agentic conversational AI - goes further: it interprets the customer's intent, checks their account state and entitlements, applies the relevant policies, and executes the action on their behalf. The difference is between a system that informs and a system that acts. Most banks currently have chatbots. The banks moving fastest are deploying conversational AI that produces actions as outputs, not recommendations.
Which banks have already launched agentic AI assistants for customers?
Starling Bank launched Starling Assistant in March 2026, described as the UK's first agentic AI financial assistant, capable of setting up savings goals, organising bill payments, and carrying out routine banking tasks through voice or text.
OCBC launched the beta of what it describes as Southeast Asia's first AI-native mobile banking app, featuring two AI avatars - Wendy and Wayne - offering wealth clients 24/7 personalized investment insights and portfolio tracking, with human relationship managers still handling complex products and advice.
Discovery Bank's Discovery AI resolves up to 55% of customer queries in the first interaction, in under 30 seconds, running continuously inside the banking app and through WhatsApp.
Santander and Mastercard executed Europe's first live end-to-end payment by an AI agent in March 2026, completing a regulated pilot using Mastercard Agent Pay.
Citizens Bank is piloting agentic AI in its retail call centre, targeting 25% of calls handled by AI agents by end of 2026, with a longer-term goal of 50%-plus.
Why are technology companies acquiring conversational AI companies in 2026?
The acquisitions reflect a race to control what Forrester analyst Aurelie L'Hostis calls the "agent runtime" - the execution layer that sits between a customer's expressed intent and the bank's systems. The common thesis: the interface layer - where intent becomes action - is the next strategic battleground, and the companies that control it will define how customers interact with their banks for the next decade.
Salesforce acquired Fin for $3.6 billion in June 2026 to bring agentic customer service into its Agentforce ecosystem. ServiceNow completed its $2.85 billion acquisition of Moveworks in December 2025 to control the enterprise workflow orchestration layer. Backbase acquired Kasisto in June 2026 to bring over ten years of banking-specific conversational and agentic AI infrastructure into the Banking OS.
What do banks need to build to make conversational AI genuinely work?
A conversational interface that executes banking work safely requires four things to be in place underneath it.
First, a unified customer state - a single, continuously updated view of who the customer is, what they hold, and what they are entitled to do - that the conversational layer can read from in real time. Second, a policy engine that governs what actions the AI can take on a customer's behalf, under what conditions, and with what limits. Third, a full audit trail for every action taken - every execution needs to be explainable and traceable for regulatory purposes. Fourth, orchestration infrastructure that connects the customer's intent to the bank's systems and completes the work end to end.
Banks that have these four things in place can deploy conversational AI that acts. Banks that do not will find that their conversational interface produces recommendations that still require a human to execute - which is a better chatbot, not a new operating model.
What happens to banks when LLM models commoditize?
When frontier LLM quality becomes effectively equal across open-source and proprietary models - which is already happening and is likely to accelerate within twelve months - the model itself stops being the differentiator. Every bank will have access to roughly equivalent intelligence at declining cost.
What will not commoditize is the execution layer underneath the model: the unified customer state, the policy engine, the audit infrastructure, and the orchestration layer that turns intent into completed work inside a regulated bank. The banks that have built that foundation will be able to deploy any model on top of it and compound the advantage. The banks that have been focused on model selection without building the foundation will find the question was always the wrong one.
Is conversational AI in banking a feature or an operating model decision?
It is an operating model decision, and banks that treat it as a feature will find it underdelivers. A conversational interface bolted onto a fragmented architecture - where the conversation does not share the same customer state as the advisor workspace, does not apply the same policies as other channels, and does not produce a full audit trail - is a better chatbot. It does not change how the bank operates.
A conversational interface built as part of a unified operating model - where customers, employees, and AI agents all work from the same customer state and the same policy engine - changes the fundamental economics of how the bank serves its customers.
The question is not which conversational AI vendor to choose. It is whether the bank's architecture can support an operating model where intent flows into governed execution across every surface.
.png)
