The difference between AI that works and AI that sells

From fraud detection to Generative BI: a practical guide to which AI applications in financial services deliver results and which ones just make decks look good.

When you strip away the marketing language, which specific AI applications in financial services are delivering measurable results?

The AI applications delivering measurable returns are the ones connected to clear workflows, clean data, and accountable business outcomes.

Fraud detection is the most mature example. Companies like Stripe Radar and Feedzai use machine learning to detect suspicious transaction patterns, reduce false positives, and lower fraud losses. Measurable means fewer losses, faster reviews, lower operational costs, and better customer experience.

Another fast-growing area is Generative Business Intelligence. Many institutions already have data, but leadership cannot access insight quickly enough. INSART is building systems that let executives or advisors ask questions in plain English, generate charts, prepare narrative summaries, and understand performance drivers without waiting for manual reports.

In one anonymised INSART use case, a reporting platform moved from static reports toward an AI assistant that supports natural-language questions, dynamic SQL generation, chart creation, saved reports, and follow-up conversations. It's use cases like this that underpin McKinsey's estimate of USD 200–340 billion in potential annual value for banking.

What are the signs that a financial institution is using AI as a valuation narrative rather than a functional tool?

The clearest sign is when AI is described as a capability, but not attached to a business process. ‘We use AI’ is not a business outcome. ‘We reduced manual review time by 30%’ is.

There's a reason those questions matter: Gartner predicted 30% of generative AI projects would be abandoned after proof of concept, citing poor data quality, unclear business value, and escalating costs; exactly the gaps those three questions are designed to surface

I usually ask three questions: what data does the system use? What decision does it improve? What KPI changes after deployment? If those answers are unclear, AI is probably being used as a valuation narrative.

A company like Morgan Stanley is a useful positive because its AI assistant for wealth advisors is attached to a real knowledge workflow. It helps advisors retrieve and use information more efficiently.

At INSART, we see the same distinction in reporting. A dashboard with an AI label is not a transformation. But a Generative BI system that helps a CFO ask ‘why did margin decline this month?’ and receive a structured answer from live business data is a functional tool. 

Credit scoring with AI raises tension between predictive power and explainability. How are institutions navigating that?

Responsible institutions are not treating AI credit scoring as full automation. They are treating it as AI-augmented underwriting. 

Companies like Upstart and Zest AI show where the market is going: broader data, stronger predictive models, and greater pressure to explain decisions. The model cannot simply be accurate. It must be explainable, auditable, and monitored for bias.

In INSART projects, the same architectural principles apply even outside lending: AI systems must include access controls, data lineage, validation, and human-readable outputs. In our Generative BI work, for example, we focus not only on generative answers, but also on validating SQL queries, enforcing role-based access controls, and preventing unauthorised data exposure.

Regulators are not the real excuse. Weak architecture is. If explainability, audit trails, and governance are built in from the beginning, AI can improve decision-making without removing accountability. The opportunity is real: 45 million credit-invisible adults represent an underserved market. But capturing it responsibly requires the architecture to match the ambition.

When a bank decides to build proprietary AI versus buying a vendor solution, what tends to go wrong?

Build and buy both fail when institutions underestimate integration. 

When banks build internally, they often underestimate the complexity of production. A model is not a product. It must connect to legacy systems, reporting workflows, security rules, compliance controls, and user behaviour.

When they buy, the risk is vendor overselling. Some products look impressive in demos, but fail when connected to messy data, real permissions, edge cases, and regulated workflows. Companies like DataRobot represent the broader category of AI platforms, but success still depends on whether the institution can operationalise the models. 

At INSART, our anonymised Generative BI work shows the practical difficulty. The challenge was not just generating SQL. It was intent interpretation, query validation, chart configuration, access control, cost management, user feedback, and future scalability. 

That pattern shows up in the data: BCG found that only 26% of companies have developed the capabilities needed to move beyond pilots and generate value at scale. The right decision is not build or buy. It is: where do we have proprietary data, proprietary workflow, and long-term control requirements?

Financial institutions sit on historically biased lending and risk data. How often is that a genuine obstacle?

It is a genuine obstacle, especially in lending, insurance, and risk scoring. Historical data reflects historical decisions. If those decisions included bias, a model can automate the past while presenting it as innovation. At the concentration levels the industry operates at, FICO scores are used by 90% of top US lenders, so the risk isn't marginal. It compounds.

Companies like FICO, Upstart, and Zest AI operate in an environment where predictive performance must be balanced with fairness, explainability, and compliance. The institutions that handle this well do not treat bias as a legal add-on. They address it in data selection, feature engineering, model monitoring, exception handling, and governance. 

At INSART, the same principle applies to any AI decision systems. In our reporting and Generative BI work, we emphasise controlled data access, validated outputs, and business context interpretation. A system should not only answer quickly; it should answer within the correct data, permission, and governance boundaries.

The future of AI in finance will not be black-box automation. It will be governed by intelligence.

Which AI applications will look obviously valuable in five years, and which will look underwhelming?

The obviously valuable applications will be embedded into financial infrastructure: fraud detection, AML, KYC, credit risk support, operational automation, document intelligence, and Generative Business Intelligence.

I am especially bullish on Generative BI because it moves AI from a separate tool into the decision layer of the organisation. Traditional BI answers, ‘what happened?’ Generative BI helps answer, ‘why did it happen, what changed, and what should we examine next?’

Companies like Bloomberg, Morgan Stanley, and AlphaSense show the broader direction: AI is becoming part of the professional search, interpretation, and action on information.

At INSART, our mission is to help institutions move from reporting chaos to decision clarity: classic dashboards, client or investor reporting portals, consolidated data layers, and conversational intelligence built on trusted data.

The underwhelming category will be generic chatbots, AI-labelled dashboards, and thin wrappers with no proprietary data or workflow integration. The strongest AI in finance will be almost invisible: trusted infrastructure that improves decision speed, risk control, and business judgement.

McKinsey’s USD 200-340 billion annual banking value estimate supports the idea that the largest gains will come from embedded, workflow-level AI rather than isolated experiments.

The future of AI in finance belongs not to institutions with the loudest AI story, but to those with the strongest data architecture, clearest governance, and fastest decision loops.

About Vasyl Soloshchuk

 

Vasyl Soloshchuk is the CEO and Founder of INSART, a B2B Tech & Fintech Business Studio. With 20+ years in B2B Tech, fintech, software engineering, and business system design, he has helped 30+ fintech and B2B technology companies launch products, accelerate product roadmaps, modernise legacy platforms, and build scalable engineering capabilities. Vasyl is also a co-founder of the Kharkiv IT Cluster, which develops the business ecosystem of the Global Ukrainian Tech Nation, founder of the Fintech CTO Club, and a contributor to the fintech industry publications, including The PayTech Book and The InsurTech Book.

 

About INSART

INSART is a B2B Tech & Fintech Business Studio on the mission to build digital freedom through B2B technology that pays back by helping founders, executives, and innovators to turn data into decisions, decisions into dollars, and software into real business value. INSART builds B2B tech-enabled businesses, not just B2B software — from idea to traction, from traction to scale, and from scale to valuation. Headquartered in Boca Raton, Florida, with clients concentrated in the New York Metropolitan Area, North Carolina, Virginia and other US East Coast financial & tech hubs, with engineering hubs in Portugal, Spain, Poland, and Ukraine, INSART operates at the intersection of an AI-enabled Software Engineering Powerhouse, Startup Studio, and an Innovation Lab – co-building new ventures, accelerating existing fintech products, and driving digital transformation across the financial ecosystem. Its expertise spans B2B Tech, AI, BI, Reporting, WealthTech, Capital Markets, Investment Analytics, Corporate Finance, Non-profits, and other financial services domains. Our clients include AdvisorEngine (a subsidiary of Franklin Templeton), iCapital (through the Mirador acquisition), Accel KKR portfolio companies like Enmark and Salsa Labs, and MPI Markov Processes International, serving global leaders such as BlackRock, Fidelity, and John Hancock.