When Every Bank Has the Same AI, What Still Matters?

A model advantage has a short half-life.

One bank gets access to a new capability. For a few months, its demos look better, its internal assistant feels smarter, and its executives can claim a lead. Then the same model appears in every major cloud platform, every enterprise software suite, and every competitor's procurement pipeline.

The gap closes.

This is already happening. Financial institutions are increasingly buying models from the same small group of providers. The Bank of England and FCA's survey of UK financial services found that the top three providers represented 44% of reported third-party model supply. A third of all AI use cases relied on third-party implementations.

The Financial Stability Board has reached the same conclusion from a systemic perspective. Its 2025 work on AI monitoring highlights the industry's dependence on a small number of providers across specialized hardware, cloud infrastructure, and pretrained models.

Concentration creates risk. It also creates strategic clarity.

If every bank can call the same intelligence, intelligence itself is not the moat.

What matters is everything around it.

Model parity will arrive before operating parity

Banks often discuss AI strategy as if model selection were the decisive choice. It is not irrelevant. Accuracy, latency, cost, data residency, and contractual protections all matter. But a stronger model cannot compensate for a weak operating system around it.

Give the same model to two banks and the outcomes can still be radically different.

At the first bank, the assistant has access to current policies, understands which documents each employee may see, works inside the system of record, cites its sources, routes uncertainty to the right person, and measures whether the task was completed faster or better.

At the second, the assistant lives in a separate window. Employees copy and paste sanitized fragments into it, receive polished text without reliable provenance, and then re-enter the result into another application. Usage may look respectable. Economics will not.

The model is identical. The institution is not.

This is why the model is the cheapest part of enterprise AI. The scarce capabilities are integration, permissions, process design, evaluation, adoption, and accountability.

The first moat is proprietary context

Banks possess an extraordinary amount of information. That does not mean their AI can use it.

Useful context is not a data lake. It is current, governed information connected to a specific job. It includes product rules, client permissions, transaction history, risk policy, market research, previous decisions, and the institutional logic that explains why work is done in a particular way.

DBS offers a useful example. The bank says its internal DBS-GPT provides role-based access to more than four million policies and content items. The phrase role-based matters as much as the number. A system that retrieves everything is not more intelligent. In a regulated institution, it is dangerous.

DBS also reported more than 2,000 models across over 430 use cases and attributed approximately SGD 1 billion of economic value to data analytics and AI/ML initiatives in 2025. Those figures are company-reported, but its operating approach is more instructive than the headline number: controlled access to institutional knowledge combined with redesigned workflows and measurement. DBS's 2025 CEO reflections describe nine operating-model transformations built around human-AI collaboration.

Models can be purchased. An institution's permissioned context, accumulated decisions, and process history cannot.

The second moat is workflow

AI creates value when it changes a unit of work.

The relevant unit may be a transaction alert, a client review, a service request, a credit file, a policy question, or a line of code. The institution needs to know how that work enters the process, which systems it touches, where judgment is required, what evidence must be retained, and what successful completion means.

JPMorganChase reported that AI in transaction screening allowed its corporate and investment bank to review more than twice the volume while halving manual operator checks. That is a company-reported result, but it is the right kind of metric. It describes throughput and human effort in a real workflow, not access to a model. Its 2025 annual report also makes the strategic dependency explicit: AI is being applied to end-to-end client journeys and supported by an organized data estate.

This is also why many of the first valuable deployments will remain out of sight. As I argued in The First Banking AI Winners Will Be Invisible, customers may never see the system that prepared an advisor, reduced a false positive, or routed an exception correctly. They will feel the result.

The winning institution will not have the most AI interfaces. It will have the fewest unnecessary handoffs.

The third moat is trust by design

Trust in financial services is not a branding layer applied after deployment. It is an operating capability.

A trusted AI system knows what it may read, what it may recommend, what it may change, and when it must stop. It produces evidence. It records which version of a policy or model was used. It distinguishes a reversible drafting task from a material action affecting a client. It gives a qualified person enough context and authority to intervene.

These controls can look like friction when a team is optimizing for launch speed. At scale, they are an accelerator. A reusable evaluation process, access pattern, audit standard, and escalation design lets the next use case move faster. Without them, every deployment becomes a new negotiation among technology, risk, legal, compliance, and the business.

That is an organizational question, not a model question. As I argue in “Human in the Loop” Is Not a Governance Model, responsible oversight depends on the design of authority and intervention. Institutions that turn governance into shared infrastructure will ship more, not less.

The fourth moat is the learning loop

Every bank can buy a capable model. Few banks can learn from production quickly without weakening control.

The learning loop begins with observable outcomes. Did handling time fall? Did first-time resolution improve? Did false positives decline? Were more cases completed without rework? Did the recommendation remain appropriate across customer segments? When employees overrode the system, why?

Those signals should change the workflow. Repeated escalation may mean the system has too little context. Low adoption may indicate that it adds steps. Excellent offline accuracy with poor business results may reveal that the team optimized the wrong objective.

The institution that captures these lessons can improve prompts, retrieval, interfaces, permissions, policies, and training. The one that measures only logins and token consumption cannot.

This produces a compounding advantage. Better workflows generate better feedback. Better feedback creates safer expansion. Safer expansion creates more usage and more evidence. The model may remain the same while the operating performance diverges.

A better competitive scorecard

When evaluating an AI program, I would stop asking how many models or copilots the bank has and ask five different questions:

Context: Can the system retrieve current, entitled, decision-relevant information?
Workflow: Is it embedded where the work already happens, with fewer handoffs than before?
Control: Are authority, evidence, escalation, and accountability explicit?
Economics: Is there a baseline and a measurable change in time, cost, quality, risk, or revenue?
Learning: Does production evidence improve the system and determine whether its scope expands or contracts?

That scorecard will reveal more than any benchmark leaderboard.

What still matters

The model market will continue to move quickly. Costs will fall. Capabilities will converge. Banks will switch providers, use several models, and route tasks to different systems depending on risk and economics.

This does not make strategy less important. It moves strategy up a level.

The durable advantage will sit in permissioned context, redesigned workflows, customer and employee trust, distribution, and the ability to learn from real outcomes. It will sit in the unglamorous machinery that turns a probabilistic output into reliable work.

Soon, every bank will be able to generate an impressive answer.

The winners will be the banks that know what should happen next.