Jayanthi Mani

I support View B — Preserve flexibility. My clear position is that AI should standardise the case-handling process while still allowing local teams to make context-based decisions. To support this view, I will first explain the role of flexibility in service operations, then analyse a specific industry example, and finally discuss how an integrated approach can balance AI-driven accordance with human decision-making. In service operations, the right answer is not always the most standardised answer. The right solution is the one that solves the customer issue while remaining within policy and risk controls. A particular operational example is the customer complaint and service recovery process in the hotel industry. The Ritz-Carlton illustrates the main lesson that it is possible to maintain high service standards while also granting frontline employees decision-making authority. As a highly standardised luxury hospitality brand, it authorises employees to spend up to $2,000 to resolve guest issues without manager approval, ensuring timely, customised solutions tailored to specific customer needs. According to Forbes, this policy demonstrates how the company successfully integrates strict service standards with flexibility for local judgment, highlighting the value of empowering staff to manage unique situations effectively. This is important because hotel service recovery cannot be managed by a single central rule. For example, if a guest arrives and their room is not ready, a standard AI recommendation may be: apologise, offer a drink voucher, and prioritise room cleaning. That may be fine for a routine case. But if the guest has travelled internationally, is attending a wedding in one hour, has already had a previous bad experience, or is a high-value loyalty customer, the local hotel team may need to make a different decision. They may need to arrange another room type, arrange transport, offer a meal, involve the duty manager, or take an immediate stronger recovery action. This is the process lesson: Ritz-Carlton standardises the service expectation, but it does not fully standardise the human response. Employees are trained on brand standards, but they are still trusted to act based on the guest situation. That is the same operating model I would apply to AI. In an AI-led service organisation, the AI should first review the case details, customer history, SLA, policy rules, risk level, and past complaints. In most cases, the team ought to follow the AI's recommendation. However, when a case involves unique or complex circumstances, local teams should be permitted to override the recommendation by providing a mandatory reason code, such as repeat failure, regulatory risk, high-value customer, vulnerable customer, or local operational constraint. Critics could argue that allowing such overrides undermines the purpose of standardisation by opening the door to inconsistent service delivery, subjective decision-making, or misuse of override authority. While these risks are credible, they can be successfully managed. Overrides should not be hidden; instead, they should be subject to monthly review to assess whether the employee made an appropriate decision, whether additional coaching is required, or if the AI rule itself requires refinement. This approach tackles concerns regarding inconsistency while preserving the necessary flexibility for context-driven judgment. This gives the organisation both coherence and judgment. AI reduces random variation in routine cases, but local teams still have room to manage real-world exceptions. Leadership can measure this by tracking AI recommendation acceptance rate, override rate, customer satisfaction after override, repeat complaint rate, and SLA performance. This is why I challenge Bex’s position. Bex correctly argues that standardisation improves quality, training, and extensibility, which are necessary for preserving consistent service levels across large organisations. However, Bex’s perspective overlooks circumstances in which strict adherence to standard procedures may yield less-than-ideal outcomes, particularly in complex or exceptional cases. Assuming that all variation is undesirable fails to recognize that certain variations stem from necessary human decisions and local expertise, rather than from a lack of discipline. Therefore, while eliminating unwarranted inconsistency is important, it is equally critical to preserve space for context-driven decisions that can fulfil unique customer needs successfully. Therefore, local flexibility should not be reduced. It should be controlled, measured, and used to improve the AI model. AI should guide decisions, but local teams should retain the authority to make justified exceptions when customer context and operational reality demand it.

My View is B. Organisations must keep humans in control of AI-driven process changes, regardless of the AI's confidence. To support this argument, I will first outline AI’s limitations in managing organisational complexities. I will then discuss essential regulatory requirements mandating human oversight, and finally analyse the risks associated with removing human review. Practical oversight mechanisms, such as requiring human approval workflows, implementing audit trails, and conducting regular peer reviews of AI-driven decisions, are essential. These methods ensure that leaders can maintain meaningful human involvement and oversight throughout every stage of the process. Bex says we should trust proven AI to make changes on its own for greater speed and efficiency, citing Siemens’ manufacturing results. This view downplays the risks. In real operations, even accurate AI can make costly, dangerous, or unethical errors if not checked. Quality of Reasoning and Argument: This section will demonstrate the necessity of human oversight by first outlining AI’s limitations in handling organisational complexity, then discussing key regulatory requirements, and concluding with an analysis of the risks posed by removing human review. AI cannot predict every ripple effect, ethical issue, or rare edge case in dynamic organisations. Human oversight is essential. It is the last check for accountability, legal compliance, and ethics. Regulations such as GDPR in Europe, the FDA's requirements for medical devices, and financial sector standards like SOX all mandate some form of human review and accountability when using automated systems. Removing human review increases risks, erodes trust, may put organisations in violation of industry standards, and exposes them to harm. Relevant Examples: Knight Capital (2012): An unchecked algorithm lost $440 million in under an hour, demonstrating the risks of lacking a final human review. Amazon pricing bots: Feedback loops set a book price at $23 million, showing even proven algorithms need human oversight in live use. Healthcare: An AI-recommended drug dose was caught by a nurse, proving human context is vital in crucial scenarios. Tesla Autopilot: An update caused cars to brake unexpectedly, triggering investigations and showing AI risks without oversight. Going Beyond Bex’s Analysis: Bex relies on the idea of 'proven reliability,' asserting that consistent performance under tested conditions justifies granting AI greater autonomy. However, this perspective underestimates significant limitations inherent to AI systems. Evidence from real-world operational failures demonstrates that even highly reliable AI can perform unpredictably when exposed to unfamiliar environments, novel data, or complex interactions with other technologies. Siemens’ example, while illustrative within a controlled manufacturing context, does not address the diversity or unpredictability present in many organisational settings. Furthermore, this argument overlooks the fact that decision-making in sectors such as finance, healthcare, retail, and transport often confronts ambiguous scenarios and rapid change. In these domains, success frequently depends on the capacity to exercise ethical judgment, interpret nuanced information, and adapt to evolving regulations—areas where AI, irrespective of prior reliability, remains fundamentally limited. By focusing solely on technical reliability, Bex’s view fails to account for broader organisational and ethical complexities. It is therefore imperative for leaders to recognise that even well-trained AI, when operating without adequate human oversight, can generate failures with significant, and at times irreversible, consequences across diverse industries. AI can recommend and prepare changes, but humans must have the final say. This prevents major errors and keeps organisations ethical and trustworthy. True operational excellence comes from combining AI’s power with human judgment.

I support View B: Wait until the organisation is ready. Even the best AI-recommended process change will fail if the organisation is unprepared. Success depends not just on data, but on the people who implement changes daily. Without adequate buy-in, training, or communication, even data-driven innovations can stumble, breed resistance, and diminish trust in both the process and AI. Real-World Examples Demonstrating the Need for Readiness 1. Electronic Health Record (EHR) Implementation in Hospitals Hospitals that skipped preparation and staff training suffered workflow problems, lower productivity, and staff resistance. Hospitals that focused on readiness and gradual rollouts saw better results. MD Anderson Cancer Centre rushed its EHR rollout, resulting in $62 million in losses, workflow confusion and staff frustration. 2. British Airways’ Terminal 5 Opening British Airways opened Heathrow Terminal 5 with new automated baggage systems and workflows, but staff were not adequately trained to use them. The result: thousands of lost bags, cancelled flights, and widespread criticism. Problems improved only after thorough retraining and process adjustment. Why This Matters for AI-Driven Change AI recommendations, although powerful, are only as effective as the organisation’s ability to implement them. If a team is sceptical of AI or feels the change is being “forced” on them by an algorithm rather than internal expertise, resistance will be high. Training and communication are not just “nice to have”—they are essential for any successful change management strategy. Conclusion Invest in readiness. Explain why the change matters, using data and context. Involve team members in the transition process. Provide training and support. Address scepticism directly by combining AI insights with human expertise. Rushing may boost short-term performance, but risks long-term setbacks and distrust in AI. Change works only when people are ready. Supporting View B is about making progress last, not about resisting change.

I support View A: Roll back immediately. When a new feature is rolled out to a user base, it is expected to enhance the overall product experience while maintaining reliability. A product earns and maintains trust by consistently delivering a reliable experience to all users. Recent industry research, such as the 2023 Forrester report on software adoption, shows that even minor disruptions can significantly affect user confidence and lead to negative perceptions of brand reliability. If a feature causes significant issues, even for a minority, it reveals a quality or compatibility gap that can erode confidence, particularly among users who may already feel marginalised, such as those on older devices or with atypical usage. Rolling back immediately shows a commitment to user trust and product stability, which are essential for long-term adoption and brand reputation. Example: In 2018, Microsoft released a Windows 10 update with new features and performance improvements. Shortly after, a small subset of users reported critical data loss. Although most users were unaffected, Microsoft paused and rolled back the update for everyone. This proactive decision aligns with the argument that immediate rollback is necessary to preserve user trust and product stability. By only resuming the rollout after resolving the root cause, Microsoft not only regained customer confidence but also prevented broader reputational damage, illustrating the importance of prioritising reliability even when issues affect only a minority of users. Reasoning: Trust is difficult to restore once lost, especially when a notable percentage of users encounter errors. If 8 to 10 per cent of users experience such issues, the consequences may extend beyond individual dissatisfaction, as these users are likely to churn, complain publicly, or discourage others from engaging with the product. This risk highlights the broader impact on user trust and retention, demonstrating how even problems affecting a minority can undermine the product's overall perceived reliability and reputation. Long-tail risk: Small affected segments are particularly important because they can include influential customers whose opinions shape broader perceptions of the product, as well as edge cases that may expose underlying, systemic issues not immediately apparent in mainstream usage. Furthermore, compliance-sensitive users, such as those who rely on accessibility features or operate in regulated environments, may experience disproportionate negative impacts. Failing to address problems encountered by these groups not only risks alienating key stakeholders but can also signal a lack of commitment to inclusivity and regulatory compliance, potentially resulting in legal challenges or reputational damage that extend far beyond the initial user subset. Operational efficiency: Debugging and selectively fixing issues in production while a feature remains live increases complexity, risks further instability, and diverts resources. Culture of accountability: Rolling back signals to all stakeholders that quality and user experience are non-negotiable. Conclusion: Rolling back is the responsible choice. Some may contend that maintaining the new feature could foster short-term engagement, expedite user feedback, or accelerate innovation, particularly if most users are not directly affected by the observed issues. Proponents of this perspective argue that continuous feature delivery and rapid iteration are essential in fast-paced markets, suggesting that prompt remediation or targeted fixes could mitigate adverse effects without significantly disrupting the broader user base. They argue that this approach enables organisations to remain agile, learn from real-world use, and address defects with minimal interruption to ongoing development. However, this line of reasoning underestimates several critical risks. Even targeted fixes may not resolve underlying systemic issues, and the visibility of persisting problems can amplify user dissatisfaction, especially among those who feel neglected or marginalised. Additionally, the perception that only the majority’s experience is prioritised may erode inclusivity and long-term loyalty. The potential for negative word-of-mouth, slow but cumulative attrition, and reputational damage outweighs the incremental gains in engagement or the speed of feedback. Ultimately, while continued rollout and rapid iteration may appeal for their perceived efficiencies, reliability for all users must be championed, regardless of segment size, because reputation and user confidence remain the true drivers of long-term product success.