Vishnu Ramakrishnan

I consider three key questions while designing a governance framework for AI. 1. What are we doing and why? Team Charter: A public, plain-language document outlining the organization's commitment to ethical AI. This includes principles like Fairness, Accountability, and Transparency, as well as Security and Human Oversight. Use Case Registry: A central, living inventory of all AI models in development or production. For each model, it should clearly state its purpose (the business process it automates and the specific decision it influences), data (what data it was trained on and what data it uses to make decisions) and the owners (business leaders accountable for its performance and impact) 2. Who is responsible? Roles that play both technical as well as business accountability. Model Owner (Business): A leader from the business unit who is ultimately accountable for the AI's business outcomes and risks. Model Owner (Technical): The data scientist or MLOps engineer responsible for the technical health, monitoring, and maintenance of the model. 3. How do we get better? Performance Dashboards: There should be a continuous mechanism to monitor the performance of AI's real-world business impact and operational health. Feedback Loops: A clear, simple process for end-users to flag unexpected or seemingly incorrect AI behavior. Audit Trail : For every decision made by AI, the system must be able to log why it made that decision in a way that a human auditor can understand. This isn't just for compliance; it's essential for troubleshooting and building trust. 4. Who Should Be Involved? A multi-disciplinary approach ensures all angles are covered. Steering Committee (Strategic Level): A cross-functional group of senior leaders that sets the overall AI strategy and principles. (COO, CTO, CDO, Head of Legal/Compliance, and key business unit leaders) AI Review Board (Operational Level): A hands-on group responsible for reviewing Impact Assessments and risk audits Business & Technical Teams (Execution Level): People on the ground. (Model Owners & Stewards responsible for the day-to-day success and health of a specific AI model).

With AI chatbots predominantly replacing human customer support agents in handling the first level of customer interactions, the existing customer center Customer Satisfaction Score (CSAT) may not be reflecting the customer's overall experience or their loyalty to the brand. A customer might be "satisfied" with a quick answer but not truly engaged or emotionally connected. Customers are increasingly concerned about the need to interact or speak directly with humans. As AI begins to automate decisions, personalize experiences, and reshape workflows, new metrics should emerge that measure the efficiency and effectiveness of the AI models used. Say, for example, AI agents are used to predict the customer lifetime value for a telecom company—there should be right metrics factored in to continuously monitor the success of predictions and a feedback loop to train the models accordingly.

I see AI as a tool, and the role of an AI solution architect as an enabler to improve processes. The role of MBB/process consultant is still valid as they possess in-depth understanding of the process nuances. The AI tools can make the MBB role less prominent or diluted, as many statistical analyses/methodologies can be driven by AI. In that case, the MBB role needs to evolve from a hardcore process improvement expert to that of a value facilitator/integrator who ensures AI solutions are adopted across the right problems in the workflow/processes. MBB's in-depth knowledge of business processes still provides an upper edge to decide if the AI initiatives are aligned with process excellence, customer value, and organizational priorities. The unique value of an MBB in an AI project lies in their ability to bridge the gap between business strategy and the technical intricacies of AI implementation. Their role shifts from a traditional process mapping and optimization to become one of a 'value orchestrator.' Practical strategy to collaborate with AI teams 1. Define & Design Phase—While MBBs can lead the identification of high-impact business problems (using DMAIC methodologies), AI architects can focus on the right AI solutions/models and architectural POVs/solutions. 2. Develop & Test Phase—AI solution architects can lead the solutioning and train the right AI models while MBBs identify critical data sources and elements required for the AI model based on process knowledge and ensure data quality and integrity from a business process standpoint. 3. Implement & Change Management—MBBs lead the change management, value realization, and stakeholder management. The AI team can lead the deployment and oversee the technical performance and stability of the solution.

To get to the heart of the issue, it’s essential to go beyond surface-level symptoms. Here are a few practical ways to separate the noise from the real root causes:Its really critical to understand the root cause analysis and following are the ways to separate noise from real causes. 1.Start probing more with the “5 Whys” approach. This classic technique wont stop until getting to the core problem 2. Look at from multiple data points and right range - Look at the issue from multiple data points: Operational data (e.g., delivery logs), Customer data (e.g., complaints, churn rates), Market data (e.g., competitor moves, seasonal trends). 3. Create a visual represenation using Current reality tree - Tools like CRT, SIPOC diagrams or value stream mapping help visualize where handoffs or delays occur. In order to prevent chasing the wrong cause, ensure you are following a structured problem solving framework like DMAIC, reviewing KPIs to catch early signals of issues and validating the hypothesis with right data points.

In the Measure phase it’s easy to grab whatever data is easiest to collect, like timestamps from a system. But that doesn’t always tell the full story. To really understand what’s going on, we may need to focus on metrics that matter, not just ones that are convenient. For example, instead of just tracking how long patients wait from check-in to being seen, we might also want to measure how long they wait after being seen, or how long it takes to get test results. These might require a bit more effort to track, but they paint a much clearer picture of the patient experience. As for catching bad data before it messes everything up—one way is to always do a quick sanity check. If a number looks way off (like a patient supposedly waited 0 minutes or >5 hours), flag it. Also, look for patterns: if one nurse’s entries always seem faster than everyone else’s, maybe there’s a data entry issue. Visualizing the data—charts and graphs can make outliers and inconsistencies jump out in a way raw numbers don’t.

Recently, I worked on the billing and invoicing process for a Business Unit in my company. At first glance, it appeared to be a very efficient process, especially considering the small number of people handling a large volume of invoice, billing, and collection transactions. A benchmark analysis indicated that the team is on par with the top 90th percentile of performers. Despite this, there were some inherent system issues, but the team identified RPA (Robotic Process Automation) solutions to address these inefficiencies. Upon deeper investigation, I realized that the processes were not as efficient as they seemed from the outside. There were many patchwork processes and RPAs used to bridge gaps. Before simplifying or transforming a process, it is essential to ask, "Do we even need that step/process/workflow to begin with?" If we can eliminate processes and still achieve the desired business outcome, that would be the ideal scenario. For example, the team accepted physical cheques and used manual processes to scan and deposit them. In today's business environment, this is an outdated practice. As a team, we decided to make the right business decisions to accept only wire transfers or credit card payments. Another area we examined was exception handling. All standard processes were automated with RPAs, while exceptions were routed through human intervention. We are now exploring the use of AI agents to handle both automations and exceptions with minimal human oversight. With the advent of AI agents, the future team structure will be a hybrid model, with bots to automate tasks, AI agents to manage exceptions and close workflows, and humans providing oversight.

In our company, we have an internal AI platform called Alex. It is a text generation model, provide information and assist users with various language based tasks. I would like to answer the question based on my experience with the tool - whats working good from a customer experience and what could be better. 1. Tone of Responses and Personalization: Alex adopts a friendly, helpful, and empathetic tone. It avoids robotic or overly formal language.But at the same time, its flexible. Based on the user prompt the answers can vary from very formal, direct to friendly tone. Also there is sufficient level of personalization bulit into the tool. For example, based on the login, the tool addresses users by name. Users can also set preferences such as the desired creativity level of responses. 2. Response Time: Since its a company wide AI tool, I suspect if its backed by sufficient tokens and processing power. As a result response time can be high, sometimes frustrating. I would prefer having the option to choose which LLM to use. Some models are lightweight and optimized for quick responses (e.g., Gemini 2.0 Flash Lite), while others offer advanced reasoning capabilities for complex problem-solving (e.g., Gemini 2.5 Pro). Users should be able to select the LLM based on their context. 3. Feedback Style: The tool provides clear, concise feedback throughout the interaction. Whenever it takes time to process the request, it acknowledges and says " Thanks for the patience, taking time to ensure we have the right info for you". In addition to answering the user’s request, it also asks whether additional tasks related to the response are needed. 4. Error Handling: I’ve noticed that as usage of the tool increases, its error handling has improved. There appears to be a system in place to track and analyze errors, continuously enhancing the AI’s ability to manage them. Example: If the AI doesn’t understand a request, it might respond with, “I'm sorry, I didn't quite get that. Could you please rephrase your question?”

I am assuming the summary of issues/observations/escalations from different departments – HR, Finance, Tech Support are all related to some common processes that cut across these departments. Prompt and flow based AI, could help us in spotting those hidden connections. Data inputs - AI works better when the input data is structured. But we can also use AI to prompt and add more structure to the existing summaries/observations. For e.g. Summary of Issues / Observations/escalations Short descriptions of the issues, escalations, whatever being reported. Department - Where it came from (HR, Finance, etc.) Use AI prompt to categories into Issue categories Severity - How bad it was (Low, Medium, High... or whatever scale you use) Occurrence - time stamps/which system/what triggered etc Additional context - what processes, team interdependencies Once the data is properly structured and tagged to its knowledge base, we can start prompting to see any hidden patterns. Example Common issue categories across departments Correlation between issues across departments Historical trends of the problem RCA of the recurring problems The AI would then go through the data, look for patterns, and generate insights. Once AI has figured out the hidden patterns, its a matter of right visualization tools to get those insights to the leadership team. This can be achieved through interactive self serve dashboards with easy to understand summaries and reccommendations. The agents can even show alerts if the KPIs are not within the threshold. Basically, we're trying to use AI to connect the dots, identify the root causes of issues, and provide clear, actionable information

I envision extensive use of AI across multiple domains within the telecom industry. Over the next five years, I anticipate a shift in team dynamics from a fully human model to a combination of AI Agents and humans. AI agents will form the foundation of the pyramid, managing data processing, volume transactions, routine customer service, and more, while humans will focus on the top - strategy, oversight, and decision-making roles. AI agents will drastically transform the following areas within the telecom domain: Customer Experience: The current mode of customer support channels, such as call centers/CCAS, will transition to a leaner, AI-led customer support system. This shift will replace hundreds of human agents with AI agents capable of handling superior call volumes and cases. With the ability to self-learn, AI agents will continuously improve the knowledge base and enhance key metrics like FTR and NPS. They will provide personalized troubleshooting and service recommendations, significantly improving customer experience. Field Services: Field services constitute a major part of the workforce, where teams are responsible for maintaining network infrastructure and servicing customer complaints. Currently, a central team manages the scheduling and dispatching of field service agents. AI agents can be utilized to triage customer complaints, manage periodic maintenance requirements, and optimize the scheduling and dispatching process for field service agents, thereby increasing efficiency and reducing response times. On an average the telecom companies (in Canada) employ more than 1/4th of employees under its field services team. There are indsutry predictions which say this number will be cut by half in the next 5 years with proper adoption of AI agents. Network Maintenance and Automation: This area currently involves significant complexity and human intervention. Although automation is advancing, there is still considerable room for improvement. AI agents are well-suited to perform complex scenario analyses and predict failures before they occur. They can autonomously monitor, optimize, and repair telecom networks in real-time, ensuring seamless operations and reducing downtime. Overall, AI agents will take on routine tasks and optimize resources, allowing human agents to shift from reactive management to roles that involve strategic decision-making, oversight, and control. This transformation will lead to a more efficient and effective telecom industry, driven by the synergy between AI and human capabilities.

I can think of following scenarios where AI may need to escalate to a human. Incomplete Knowledge Base: When the query and resolution are not present in the knowledge base, AI may not be able to address the query effectively. Complex Queries: If the query is unclear and AI cannot route it to the appropriate problem category, human intervention may be required. Security Concerns: Queries originating from suspicious or stolen identities (e.g., flagged in the database for fraud or bad debt) may necessitate escalation for security reasons. Sensitive Customer Data: If handling the query involves some sort of sensitive customer data, it may be required to escalate to a human to ensure proper management and compliance. I can describe an example that falls under the third category - Security Concerns. In the telecom industry, one common type of fraud is Subscription Fraud. This occurs when fraudsters use stolen identities for KYC verification, sign up for services, and then fail to pay for them. Often, these fraudsters approach call centers or online sales channels, which are front-ended by AI chatbots, to request new or additional services. To combat this, additional AI/ML models are integrated into the AI chatbot workflows to screen and weed out suspicious calls and requests. If the model flags a query as a potential fraud case, the AI chatbot escalates it to human operators for further scrutiny and verification.