Ravi Nemalikanti is Chief Product and Technology Officer at Abrigo, leading AI, data, and product strategy for financial institutions.
In banking software, the startup advantage often ends at the first vendor-risk questionnaire.
AI startups can move quickly and operate without legacy code. And yes, the demos are very compelling. But in regulated industries, a compelling demo is only the beginning. It still has to pass an information-security review, satisfy model-risk and compliance teams, integrate into the bank’s environment and earn the confidence of a buyer who has to live with the consequences if it fails.
That creates a different competitive dynamic. The vertical SaaS incumbents already embedded in consequential banking workflows have an advantage that startups cannot simply raise capital to replicate. But it is an advantage with an expiration date. Incumbents must convert it into better products and faster learning loops before AI-native competitors narrow the gap.
Workflow Context, Not Just Data
The advantage is not a secret corpus waiting to be poured into a model. It is a permissioned workflow context: the evidence reviewers use, the exceptions they resolve, the steps that lead to an approval or escalation and the feedback needed to measure whether an AI system improved the work.
Foundation models know a great deal about the public corpus of data. They do not know how a particular bank’s credit committee reaches a decision, which details an investigator weighs in an AML alert or why an examiner challenged a policy interpretation. Incumbents operating inside those workflows have years of real decisions with real consequences behind them.
The opportunity is not simply to train larger models. It is to use that context, within customer policies and strong data controls, to make AI outputs more accurate, explainable and useful in the moment of work.
Domain Judgment About Where AI Can Be Wrong
In regulated markets, the hard part is rarely choosing a model. It is deciding where the model is allowed to be wrong. A missed fraud alert can become a loss event. A false positive may create more customer friction. An AML recommendation that cannot be explained can become a regulatory problem. A credit decision that cannot be reconstructed can undermine a bank’s ability to defend its process.
Incumbents have accumulated that judgment through examiner conversations, audit cycles, implementation failures and customer escalations. They know where AI can safely accelerate work, where a human must remain accountable and where automation should not be introduced at all.
Startups can learn this. But they often learn it one difficult implementation at a time.
Distribution That Starts Inside The Customer Relationship
The startup with a better model still has to win a bank’s procurement process, pass vendor risk management, survive an information-security review and convince a board that it will be a viable partner years into the contract.
An incumbent starts from a different position. It is already integrated into the bank’s environment, known to security and risk teams, and embedded in workflows customers rely on every day.
Adoption is not automatic. Material AI capabilities may still require new approvals and controls. But the conversation begins inside an existing commercial, security and integration relationship.
Why Incumbents Still Lose
None of this is self-executing. There are multiple failure modes for incumbents.
First, they treat AI as a feature roadmap rather than an operating model. Adding a chatbot to an existing product checks a box but rarely changes the underlying work. The winners will redesign workflows, beginning with their own product and engineering organizations, then extending that learning into customer operations.
Second, they confuse caution with trust. Banks do not trust a vendor because it moves slowly. They trust a vendor because it can explain what its systems do, identify their limits, preserve accountability and prevent errors from becoming material problems.
Third, they let their workflow advantage sit idle. Data becomes a moat only when it feeds a governed learning loop. Every reviewed alert, resolved exception and human correction should improve the product, where customer permissions, tenancy protections and data governance allow it. Without that loop, the incumbent is sitting on an archive, not building a sustainable advantage.
The Speed-Trust Equation
Speed and trust are not opposites. They are a sequencing problem. Move quickly where errors are inexpensive and reversible: internal productivity, research, drafting, summarization and, in some cases, code generation. Move deliberately where errors are consequential and regulated: decisioning, customer-facing outputs and any workflow an examiner may later scrutinize.
The mistake is applying one speed to everything. Uniform caution wastes the incumbent’s advantage. Uniform speed, on the other hand, destroys it.
The practical discipline is straightforward: Classify AI use cases by the consequence of error; create fast governance paths for low-consequence work; apply rigorous controls to high-consequence work; and maintain a regular cadence of governed production deployments. Organizations learn far more from a carefully instrumented capability in use than from another quarter of planning.
What This Means For Product And Technology Leaders
So if you are a product leader at an incumbent, your primary competition is not the startup demo. It is your own clock speed. The startup must build data access, workflow credibility, customer trust and distribution from zero. The incumbent must build urgency from zero. One is harder to acquire. The other is easier to defer, because the roadmap is always overflowing, and without deliberate urgency, you are the next relic.
Start with work where AI can create visible value without taking consequential decisions: preparing a loan-review package, drafting an AML alert narrative or retrieving the relevant policy and supporting evidence for a human reviewer. Then measure the result, show the controls and use the credibility earned there to take on harder workflows.
The winner in banking AI will not be the vendor that started first. It will be the one that earns the right to sit inside consequential workflows, and then gets smarter with every governed interaction while competitors are still stuck at the demo.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?