Artificial intelligence (AI) is moving deeper into financial services — from analytics and fraud detection to investment research, financial-document processing, risk assessment and increasingly autonomous decision-support.
But the most consequential shift may be taking place where AI begins interacting directly with financial behaviour.
An algorithm that identifies a pattern in a financial statement is one thing. An AI system that detects an investor’s behavioural pattern, issues a real-time warning or influences a portfolio decision raises a different set of questions around trust, explainability, accountability and customer autonomy.
Decimal Point Analytics sits at this intersection of financial-domain expertise, data analytics and AI-led transformation. The India-headquartered company works across financial services and other sectors, with its current technology portfolio including Rakshak, an AI-powered investor protection platform, and Bindu FRS 1.0, a domain-specific GenAI model for financial statement analysis and standardisation.
Gaurav Gupta, Co-founder & Managing Partner of Decimal Point Analytics, brings nearly three decades of experience spanning entrepreneurship, fund management, financial services and AI-powered analytics. His perspective offers an opportunity to examine not only what AI in financial services can do, but also where its boundaries should lie.
In this TechRecast conversation, we explore the evolution of AI in financial services from experimentation to production, the changing relationship between humans and intelligent systems, behavioural nudging, responsible AI, customer trust, accountability and the question of how financial institutions can deploy AI without losing sight of the people whose decisions and money it ultimately affects.
AI in Financial Services: From Experimentation to Consequence
Q1. From your perspective, what has fundamentally changed in AI in financial services over the past three years?
Many organisations have moved from proofs of concept to production deployments. What distinguishes an AI experiment from an AI system that is genuinely delivering measurable business or customer value?
GG: AI has become capable of processing the unstructured information central to investment work: financial statements, earnings calls, research reports and fund documents. The significant development is the ability to connect information extraction with analysis, validation and operational workflows.
However, production value requires more than a successful demonstration. A dependable system must handle difficult cases, integrate with existing processes and deliver measurable benefits after review costs and exceptions are considered. In financial spreading, for example, extracted figures must be correctly classified, reconciled to source documents and mapped into credit templates.
The distinction between experimentation and meaningful deployment is whether AI consistently delivers reliable information into an actual decision process. Better models matter, but data quality, integration and accountability determine their practical value.
Actual Financial Decision
Q2. What does responsible AI mean when the AI system can influence an actual financial decision?
In financial services, an incorrect recommendation can have a very different consequence from an incorrect recommendation in a low-risk consumer application. What principles should govern AI systems operating in such environments?
GG: Responsible AI requires controls proportionate to the financial consequences of error. A research summary, portfolio recommendation and executed trade carry different risks and should have different levels of oversight.
The starting point is the investment mandate: the client’s objectives, permitted exposures, liquidity requirements and risk tolerance. AI must operate within these boundaries. Material conclusions should be traceable to evidence, calculations independently validated and uncertainty clearly communicated.
Systems must also recognise when information is incomplete or conflicting and escalate consequential decisions appropriately. A named institutional owner should be able to explain, challenge and suspend the system. Responsible AI therefore combines investment discipline, dependable engineering and clear accountability. Disclaimers cannot compensate for weaknesses in these foundations.
The Customer and the Machine
Q3. Rakshak design is to identify potentially harmful investor behaviour and provide timely intervention. Where should the boundary lie between helpful intervention and excessive behavioural influence?
How do you ensure that an AI nudge helps customers make better-informed decisions rather than effectively making the decision for them?
GG: The boundary should be defined by the investor’s objectives, informed consent and explicit authorisation. Rakshak’s stated purpose is continuous portfolio protection, including identifying behavioural and portfolio risks. Monitoring, warnings, recommendations and execution should nevertheless remain distinct levels of authority.
A concentration warning, for example, should explain the proposed trade’s effect on exposure, compare it with agreed limits and present the investor’s options. It should acknowledge that an unusual decision may reflect legitimate information or a changed objective.
Behavioural signals should not be treated as definitive evidence of an investor’s emotional state. Interventions should also avoid commercial incentives that conflict with investor interests. Any execution authority should have separately defined limits and escalation arrangements. Effective protection strengthens informed choice while preserving meaningful investor control.
Improving Customer Outcomes
Q4. How should financial institutions measure whether AI is genuinely improving customer outcomes?
Accuracy, speed and cost reduction are relatively easy to measure. But how should organisations measure improvements in financial confidence, decision quality, trust, financial discipline or customer wellbeing?
GG: Institutions should measure improvements in decision quality and reductions in avoidable harm alongside accuracy, speed and cost. Relevant indicators include adherence to agreed risk limits, unintended concentration, unnecessary trading, investor understanding and consistency with long-term objectives.
Confidence should be evaluated together with comprehension: a persuasive system can increase certainty without improving judgement. Complaints, missed warnings, excessive alerts and differences in outcomes across customer groups also deserve attention.
Investment returns require careful interpretation because market conditions can disguise weak processes or temporarily disadvantage prudent decisions. Evaluation should therefore use appropriate benchmarks, risk-adjusted comparisons and meaningful observation periods. Where feasible, controlled evaluations or staged deployments can help distinguish AI’s contribution from other influences. Genuine improvement means better customer outcomes, supported by credible evidence.
Trustworthy Intervention
Q5. If an AI system identifies that an investor is behaving emotionally, what makes its intervention trustworthy?
What level of explainability should a customer receive before acting on an AI-generated warning, recommendation or portfolio-related insight?
GG: Trustworthy intervention starts with observable behaviour and verifiable financial consequences, rather than an assumed psychological diagnosis. Repeated purchases of a declining security may justify a concentration warning, but do not establish that the investor is irrational or distressed.
The explanation should identify the relevant holdings, information sources, assumptions and resulting exposure. Uncertainty or incomplete data should be disclosed. The detail should match the proposed action: rebalancing recommendations require consideration of costs, liquidity and relevant tax implications.
Investors should also be able to challenge the system or correct its understanding of their circumstances. Trust develops when warnings are specific, evidence-based and proportionate, and when the investor can understand both the concern and the available choices.
Accountability and Risk
Q6. Who should ultimately be accountable when an AI system makes or influences a bad financial decision?
Should responsibility rest with the model developer, financial institution, human advisor, platform, customer or some combination of these stakeholders?
GG: The institution offering the service should remain accountable for its deployment of AI, while responsibility for individual failures should reflect each party’s actual control. Developers, platform providers, institutions and advisors have different responsibilities that must be explicitly assigned.
Shared responsibility must not dilute ownership. Every service should have a named accountable owner and a clear process for investigating incidents, supporting affected customers and correcting failures.
A poor investment outcome does not necessarily indicate a defective decision. Assessment should examine information quality, suitability, permissions, disclosure and the decision process itself. Customer choice also matters, but cannot excuse misleading recommendations or inadequate safeguards. As a governance principle, delegating a task to AI should not remove institutional responsibility for how that task is performed.
Increasing Autonomy
Q7. Does increasing autonomy necessarily increase risk?
As agentic AI moves from recommendation to action, what controls should exist around permissions, escalation, human oversight, audit trails and the ability to reverse an AI-generated action?
GG: Greater autonomy increases the potential consequences of mistakes, but can reduce overall risk when tasks are well defined and controls are effective. Automated reconciliation may improve consistency; autonomous trading introduces immediate exposure to capital loss, liquidity constraints and market impact.
Authority should expand gradually from analysis to recommendation and, where appropriate, authorised execution. Relevant safeguards include mandate checks, exposure limits, reliable inputs, duplicate-action prevention, escalation and the ability to suspend further activity.
Audit trails should reconstruct the information used, permissions applied and checks performed. Reversibility must not be assumed: executed trades may be costly or impractical to unwind. The appropriate level of autonomy depends on the institution’s ability to detect failures, contain their consequences and manage remediation.
Biggest Blind Spots
Q8. What are the biggest blind spots organisations currently have when implementing responsible AI in financial services?
Are organisations paying enough attention to data quality, model drift, bias, explainability, privacy and downstream customer impact, or is the focus still primarily on deployment speed and ROI?
GG: A major blind spot is treating AI primarily as a model challenge while overlooking weaknesses in data and workflows. In investment operations, inconsistent identifiers, outdated records and conflicting classifications can produce misleading conclusions even when the model performs as intended.
Decimal Point Analytics alternative-investment reconciliation work illustrates the importance of source validation, transaction matching and explicit exception flags. These foundations become more important as automation scales.
Another weakness is reliance on average accuracy, which can conceal rare but financially material errors. Institutions must also monitor changing document formats, market conditions, model updates and user over-reliance. Privacy, access controls and third-party dependencies require continuing attention. Deployment speed and return on investment are incomplete measures unless accompanied by evidence of reliability and downstream impact.
Domain-Specific AI
Q9. Bindu FRS 1.0 reportedly achieved 98.79% first-pass accuracy in controlled testing. What exactly does that number represent, and why is the definition of accuracy particularly important in financial statement analysis?
What should enterprise buyers examine before translating such a benchmark into assumptions about production performance?
GG: The reported 98.79% first-pass accuracy reflects Decimal Point Analytics effort to make financial statement standardisation more reliable and measurable. Bindu FRS combines domain-specific processing with a feedback mechanism that learns from analyst corrections, supporting continuous improvement.
In financial analysis, accuracy extends beyond extracting a number: it includes assigning the correct accounting category, period, currency and unit. Decimal Point Analytics emphasis on validation is therefore important to the usefulness of the resulting data.
Transparency also requires clarity about how performance is measured. The published figure should be read alongside its test scope, scoring methodology and exception treatment, which the attached materials do not fully detail. Representative client-data testing provides the next step in establishing production suitability. The objective is accurate, reviewable information that professionals can confidently use. The reported percentage is evidence to evaluate, rather than a universal performance guarantee.
Smaller, Specialised Models
Q10. Why does domain-specific AI sometimes make more sense than simply deploying a larger general-purpose LLM?
Where do you see the strongest case for smaller, specialised models, proprietary data and domain-specific architectures in financial services?
GG: Domain-specific AI is often better suited to tasks requiring consistent financial interpretation, controlled outputs and repeatable validation. Financial spreading, for example, requires accounting relationships, template-specific mappings and recognition of ambiguous disclosures. Greater model size alone does not resolve these requirements.
Specialisation can combine financial terminology, relevant examples, analyst corrections and institution-specific rules. It does not necessarily require a model trained from scratch: retrieval, deterministic calculations and validation can complement an appropriate language model.
Decimal Point Analytics Bindu FRS and private-deployment approach reflect this focus on defined financial workflows. Architecture should be selected through testing of reliability, privacy, operating cost and maintainability. The objective is dependable performance on the actual task, rather than maximum general-purpose capability.
Depends on Large Volumes
Q11. Financial AI increasingly depends on large volumes of sensitive and proprietary data. How do organisations balance model performance with privacy, data governance and regulatory requirements?
Where do synthetic data, data minimisation, data lineage and controlled model environments fit into that equation?
GG: The balance begins with identifying the information genuinely required, the rights governing its use and the environments in which it may be processed. Additional data is valuable only when its benefits justify the exposure and governance requirements.
Access restrictions must follow information into AI workflows. Outputs should retain traceability to source documents, dates and processing steps, while retention and model-training permissions require explicit controls.
Private or on-premises deployment can support these objectives, but location alone does not ensure sound governance. Security, access management, evaluation and operational ownership remain essential. Synthetic data can support testing while limiting use of real records, although realism and leakage risks require assessment. Privacy should shape the architecture from the outset.
Enterprise Adoption
Q12. Many financial institutions have accumulated AI pilots without achieving enterprise-scale transformation. What separates organisations that successfully operationalise AI from those that remain stuck in experimentation?
Is the biggest barrier technology, legacy infrastructure, data readiness, governance, talent, organisational culture or something else?
GG: Successful organisations embed AI within an owned business process, supported by clear outcomes and operational responsibilities. A pilot may demonstrate extraction or summarisation; enterprise deployment requires dependable inputs, integration, access controls, exception handling and ongoing support.
The starting point should be a defined bottleneck and a measurable baseline covering time, errors, review effort and cost. Evaluation must then assess the complete workflow rather than an isolated model output.
Data readiness and legacy infrastructure frequently constrain progress, but organisational ownership and adoption are equally important. Teams need appropriate training and changes to working practices. Sustainable transformation emerges through successive operational improvements built on shared data, governance and integration foundations. A pilot establishes feasibility; repeatable operational performance establishes value.
Repetitive Workflows
Q13. What should a financial institution automate first — high-volume repetitive workflows, decision-support processes or customer-facing interactions?
What criteria should CIOs and business leaders use when deciding whether an AI use case is suitable for automation?
GG: Institutions should generally begin with recurring workflows whose outputs are verifiable and whose errors can be contained before affecting financial decisions. Document classification, reconciliation and data validation can provide immediate benefits while strengthening downstream analysis.
High volume alone is insufficient. Selection should consider business value, data readiness, error consequences, reviewability, ownership and exception-handling costs.
Decision support can follow through source-linked research summaries, inconsistency detection and draft credit analysis. Customer-facing applications require additional care because users may interpret confident outputs as personalised advice. Morgan Stanley’s Debrief illustrates an intermediate approach: with client consent, it prepares meeting notes and follow-up drafts for advisor review. The preferred sequence increases authority as reliability and control effectiveness are demonstrated.
Role of Human Professionals
Q14. How is AI changing the role of human professionals in financial services?
Rather than asking whether AI will replace analysts, advisors or fund managers, what work should humans increasingly stop doing, and what higher-value judgement should they retain?
GG: AI is shifting financial professionals’ work from information assembly towards interpretation, challenge and judgement. Analysts can spend less time transferring figures and more time assessing earnings quality, cash-flow credibility and management assumptions. Portfolio managers can examine exposures earlier and devote greater attention to investment trade-offs.
Decimal Point Analytics research-agent materials describe parallel bull and bear analysis. Such capabilities can broaden inquiry, but professionals must still evaluate evidence and determine its significance.
Automation also changes professional development. Junior staff need deliberate opportunities to learn accounting interpretation, source verification and investment reasoning as routine tasks disappear. The enduring human contribution is understanding objectives, weighing uncertainty, recognising omissions and accepting responsibility for consequential decisions. Productivity gains should strengthen these capabilities.
India’s Opportunity
Q15. India has built substantial digital infrastructure across payments, identity, financial data and financial services. Does this give Indian companies a meaningful foundation for developing AI-led financial services?
Where do you see India’s opportunity to move from being an adopter of global AI technologies to becoming a developer of domain-specific financial AI?
GG: India’s digital infrastructure provides a meaningful foundation, but capital-markets value depends on converting accessible information into reliable financial intelligence. Digital disclosures and connected information flows can support research, statement standardisation and portfolio monitoring.
Accessibility alone does not establish consistency, quality or unrestricted usage rights. Public company disclosures and sensitive investor information require different treatment, particularly around permissions, security and reuse.
India’s opportunity lies in combining financial-domain expertise, engineering capability and practical delivery experience to solve specialised problems across reporting practices and languages. Decimal Point Analytics Bindu FRS represents one such focus on financial-information standardisation. Global competitiveness will depend on demonstrable reliability across markets and accounting contexts. Infrastructure creates an advantage; dependable products and trusted execution make that advantage durable.
AI in Financial Services
Q16. Looking ahead three to five years, what will AI in financial services look like when it becomes genuinely mature?
Which capabilities are likely to become mainstream, which current expectations are overhype, and what should financial-services leaders start building today if they want AI to deliver durable customer and business value?
GG: Over the next three to five years, AI is likely to become more deeply embedded in investment workflows. Research preparation, reconciliation and exception monitoring may become continuous, while bounded agents perform selected actions within approved mandates.
Autonomy will develop unevenly. Document processing is easier to constrain than decisions involving suitability, illiquid assets or uncertain market conditions. Human accountability will remain essential wherever consequences are material.
Expectations of a universal agent that consistently selects the right investment are overstated. Institutions should instead build reliable data, connected workflows, rigorous evaluation and explicit permissions, while preserving professional expertise. Maturity will be demonstrated by improvements that endure under difficult conditions, with failures detected and managed. Trust will depend on operational evidence rather than technological ambition.
Consequential Financial Decisions
AI in financial services is moving beyond the question of whether machines can analyse more data than humans. The harder question is whether intelligent systems can participate in consequential financial decisions while preserving trust, transparency and human agency.
For enterprises, that means AI maturity cannot be measured only by the number of models deployed or workflows automated. It must also be measured by how reliably those systems perform, how clearly their decisions can be understood, how effectively risks are governed and, ultimately, whether customers and employees experience better outcomes.
For Decimal Point Analytics, the journey from financial analytics to domain-specific AI and agentic systems reflects a broader transformation underway across the industry.
The next phase of AI in financial services may therefore be defined less by how autonomous the technology becomes and more by how responsibly organisations decide where that autonomy belongs.