Apoorv

Indicators of AI readiness: Clear business objectives First of all the organisation needs to have well defined goals for AI adoption ( eg, reducing manual work, enhance customer experience, drive innovation etc.) Rather than follow the trend, AI initiatives must be tied to measurable business outcomes Data readiness Data available is of high-quality, structured and relevant Data governance policies are well established (privacy, compliance, security) There is ability to integrate data from multiple sources for AI models Leadership Commitment Senior Leadership is actively advocating AI initiatives with champion mindset There is strong commitment from the top level to invest in AI strategy, resources and building talent Actively driving a culture that supports innovation and experimentation with the evolving AI technology Skilled Workforce There is presence of AI talent (data scientist, ML engineers) or strong partnerships with industry experts to fill gaps if any The L&D team is equipped to train and understand workforce on leveraging AI tools Clear training programs Cross-functional collaboration between business and technical teams Technology Infrastructure There are tools available for model development, deployment, and monitoring Strong controls and cybersecurity measures to protect AI systems and data Scalable cloud or on-prem infrastructure to support AI workloads Ethical & Responsible AI Framework Control and governance framework with strong compliance and adherence with regulatory and legal standards ( GDPR etc) Policies for fairness, transparency and accountability in AI There is mechanism to detect and mitigate bias in AI models Change management & Culture Employees are comfortable and open to adopt AI-driven processes Organisation is embracing data-driven decision making Clear communication from top down about the impact on the organisation, roles and workflows Pilot project & Proof of concept Smaller implementations before scaling at a larger level Learnings from pilots are implemented in future implementations informing broader strategy ROI & Sustainability There is a defined framework to measure AI impact (cost savings, revenue growth, efficiency) Along with this there is clear sight & long term map available for scaling the use of AI beyond initial use cases Warning Signs Lack of ownership No single team or leader accountable for AI outcomes Ambiguity about who manages AI models and decisions Unstructured or Poor quality data Data is siloed, inconsistent or incomplete There is a heavy reliance on manual data entry Insufficient Budget or resources AI projects are under funded or treated as side experiments No allocation for ongoing maintenance and monitoring Immature Processes Frequent exceptions and workarounds Lack of standard operating procedures Over reliance on AI Leadership sees AI as magic wand without understanding its limitations No fallback plan if AI output fails or are inaccurate Cultural Resistance Employees are uncomfortable with the introduction of AI Distrust for AI recommendations

A Retail Banking Example of When Quicker Credit Decisions May Not Be Better Credit card and personal loan approvals in retail banking are one procedure where AI can significantly shorten decision times. These judgements, which involved manual underwriting, bureau reviews, and rule checks, used to take hours or days. Applications can now be approved or rejected in a matter of seconds using AI-based credit rating models. At first glance, this speed clearly improves outcomes. Customers get instant responses, drop-off rates reduce, sales productivity improves, and banks can process much higher volumes without adding staff. In competitive markets, faster approvals also translate directly into higher acquisition. However, if not managed appropriately, I have observed that this same speed can subtly introduce additional risks and blind spots. Where speed enhances results In high-volume, low-risk segments, speed is most beneficial. For paid clients with a spotless bureau background, straight-through processing eliminates needless human labour and delays that don't really provide value. Decisions made more quickly help lessen operational weariness because teams aren't scrambling to finish backlogs at the end of the month. In these cases, faster is genuinely better. Where speed creates blind spots The risk emerges when speed replaces judgement rather than supporting it. Historical data is used to train AI models. The model will duplicate past assumptions, out-of-date economic situations, or biassed approval patterns in that data, but it will do so much more quickly. Instant choices can conceal information that a human underwriter would typically look into for borderline profiles, such as self-employed applicants or new credit clients. Another risk is false confidence.When approvals happen instantly, business teams may stop questioning outcomes altogether. Over time, weak signals — early delinquency, concentration risk in certain customer segments, or over-approval during economic upswings — may go unnoticed because the system “seems to be working.” In short, AI can make bad decisions faster and at scale if assumptions are wrong. What checks are needed to ensure speed does not hurt quality Based on experience, I would put three safeguards in place. First, I would clearly define where AI is allowed to decide autonomously and where it must pause. For example, low-risk profiles can be fully automated, but edge cases should still trigger human review — even if that slows them slightly. Second, I would implement outcome-based monitoring, not just model accuracy metrics. Early delinquency trends, approval-to-default ratios by segment, and sudden shifts in portfolio quality are stronger signals than model confidence scores alone. Third, I would ensure periodic human challenge sessions, where risk, operations, and business teams review AI decisions together.The goal is not to audit the model, but to ask uncomfortable questions: “Would we still approve this customer if a human were deciding?” That dialogue often surfaces blind spots early. Closing thought AI undeniably improves decision speed in credit approval processes, and in many cases that speed is beneficial. But speed also reduces the natural friction where humans question, reflect, and challenge assumptions. The real risk is not fast decisions — it is unquestioned fast decisions. The best outcomes come when AI accelerates routine judgement, while humans remain responsible for deciding where speed should slow down.

In my experience, AI was implemented in a retail banking credit card approval process to address what was widely perceived as the primary obstacle: inconsistent and sluggish credit decision-making. Particularly during periods of high application volume, manual rule checks, bureau analysis, and policy interpretation were taking hours or even days. A machine learning–based credit scoring model was introduced to automate approvals and rejections in near real time. Initially, the impact looked impressive. Approval turnaround time dropped sharply, queues disappeared at the credit desk, and business teams declared the constraint “removed”. However, within a few weeks, it became clear that the constraint had not disappeared — it had simply moved and transformed. How the constraint shifted: Once AI accelerated decisioning, two new constraints started emerging. The first was data dependency. The AI model’s output was only as good as the input data it received. Upstream issues that were previously tolerable — delayed bureau pulls, inconsistent income fields, mismatched customer identifiers — suddenly became critical. Earlier, a human underwriter would pause, interpret, or correct such issues. Now, the system either failed silently or routed cases to manual review, creating a new queue that hadn’t existed before. The second, and more subtle, constraint was human trust and governance. Risk and compliance teams were uncomfortable with fully automated approvals, especially for edge cases. As a result, new approval checkpoints were added: model explainability reviews, threshold validations, and periodic overrides. Ironically, while the model made decisions in seconds, files started waiting for sign-offs on whether the model should be trusted in those scenarios. So while AI removed the visible credit processing bottleneck, it introduced constraints related to data readiness and organizational confidence. Signals that the new constraint had emerged The signals were not obvious at first, but they became clear over time: A rising percentage of applications being diverted to “manual exception” queues Increased complaints from operations teams about data issues rather than workload Risk teams requesting frequent recalibration or tighter thresholds despite stable outcomes Overall end-to-end TAT starting to creep up again, even though credit decision time remained low More time being spent in governance reviews than in actual credit evaluation These were strong indicators that the constraint had shifted from decision speed to data quality and trust in automation. Reflection What this experience reinforced for me is that AI rarely eliminates constraints — it reshapes them. In this case, the process moved from being people-constrained to being data- and governance-constrained. Without addressing upstream data discipline and downstream decision ownership, AI simply accelerated the process into a new bottleneck. For process leaders, the key lesson is to watch not just the speed gains from AI, but also where work, waiting, and anxiety reappear elsewhere in the system. That is often where the new constraint is quietly forming.

Finding the True Limit in the Approval of a Home Loan I have personally witnessed this difficulty in the retail banking house loan approval process, where turnaround time is a persistent source of frustration. The credit underwriting staff is nearly always brought up when discussing delays. Underwriting appears to be the bottleneck because files pile up and clients follow up. Over time, though, I discovered that the total turnaround time seldom improved, even with the addition of more underwriters or increased team effort. At that point, it became evident that the bottleneck that was most obvious wasn't always the actual restriction. The fact that delays were dispersed across the process made it extremely difficult to pinpoint the restriction. Due to insufficient or inconsistent paperwork, files were frequently moved back and forth. In order to reach goals, sales teams would submit applications fast, but with incomplete income documentation or ambiguous customer information. Additional loops that were not immediately apparent in dashboards were formed by policy exclusions for self-employed or borderline clients. These delays seemed "normal" to people and were not questioned. This is where I believe an AI-based approach could add real value. AI can reconstruct what actually occurs to a loan file instead of what we think happens by using process mining on real system logs. Such an analysis makes it clear that frequent rework between sales, operations, and customers accounts for a significant amount of the delay rather than underwriting capacity. Additionally, AI is able to identify patterns that are extremely difficult for humans to recognise. For instance, even though no single team seems overworked, it can demonstrate that specific client profiles, channels, or policy exceptions routinely take much longer. When thousands of applications are examined collectively, it is evident that the true constraint is a combination of document variability and policy complexity. Simulation is another effective application of AI. Teams can explore scenarios like adding underwriters, streamlining document checklists, or implementing early risk screening instead of arguing points of view. Such simulations, in my experience, frequently demonstrate that increasing the number of people downstream has little effect unless upstream quality issues are resolved first. Nevertheless, AI is not a panacea. It finds it difficult to comprehend why some policies are in place or how human behaviour influences the procedure. It won't automatically detect, for instance, that sales teams purposefully postpone uploads in order to prevent rejection. It still takes human judgement, expertise, and change management to interpret these insights and take appropriate action. In conclusion, the real obstacle in the home loan process is frequently not the length of the wait but rather the places where rework and variability are subtly introduced. AI aids in the objective discovery of these hidden limitations, but humans are still required to contextualise the results and promote useful advancement. The combination of both is where the true power is found.