Moving Enterprise AI From Hype to Accountable Results With Freshworks

What happens when an AI agent follows your documented process perfectly, but that process bears little resemblance to how decisions are actually made? In this episode, I speak with Bill Wilson, Executive Head of Data and AI Solutions at NTT DATA UK&I. Bill oversees AI globally for NTT DATA's public sector work, giving him a close view of how governments are using AI while trying to manage risk, accountability, public confidence, and constrained resources. Bill offers a refreshingly practical test for any proposed AI system: is it competent, and what is the worst thing that could go wrong? He describes this potential consequence as the system's "blast radius." An AI assistant helping somebody understand a grant application presents a very different level of risk from an agent making decisions that affect employment, justice, taxation, or access to public services. We also discuss why companies can make a mistake before deploying their first agent. Automating an inefficient process simply allows the organization to perform the wrong work faster. Bill argues that teams should examine complete workflows, identify where several AI capabilities could produce a measurable result, and remain prepared to redesign the process as they learn. Another major problem is tacit knowledge. Employees frequently make decisions using experience that was never written down. An agent trained solely on formal documentation may therefore understand the official process while missing how the work gets done in practice. Bill explains how targeted questions, behavioral traces, feedback, and supervised learning could capture some of that reasoning. Public sector AI provides several useful examples. Bill discusses systems that process volumes of information beyond human capacity, emergency response work in Tennessee, and case management applications that gather information before a human reviews it. In these situations, AI can reduce administrative work and waiting times while leaving consequential decisions with people. But human approval alone provides no guarantee. If employees lose direct experience of the work, they may eventually approve whatever the system recommends. Bill compares this with airline pilots maintaining manual flying skills and describes how known test cases can reveal when reviewers are becoming overly trusting. For CIOs deciding which AI pilots should reach production, the advice is equally direct: choose work with measurable returns, group related use cases where their combined effect can be seen, learn from a varied set of deployments, and avoid building something a software provider is about to include in an existing product. As AI agents gain access to external information, internal data, and operational tools, how should your organization decide what they may do alone and when a person must intervene? Listen to the conversation and share your thoughts with me.

Who carries responsibility when an AI agent begins reviewing contracts, applying regulatory rules or making commercial decisions on behalf of an organization? In this episode of Tech Talks Daily, I speak with John Nay, founder and CEO of Norm Ai, about Agentic Law and the attempt to redesign legal work around AI agents, legal engineers and experienced attorneys. John has worked across AI, law and public policy for approximately 14 years. His early research adapted neural network methods to legal and government text before large language models became a commercial force. The company information supplied for this episode states that Norm Ai recently raised $120 million in Series C funding at a $1.2 billion valuation, bringing total funding to over $260 million. Norm also says organizations representing over $30 trillion in assets under management use its technology for legal and compliance work. Those figures provide useful context for the scale of interest, while our conversation concentrates on how the model works and where responsibility remains human. John describes Norm Ai as automating the first pass of legal and compliance tasks. One example involves an in house team using an agent to review communications against relevant rules before a person finalizes the decision. Another involves Norm Law receiving transaction documents, assigning the first analysis to AI agents and then presenting the output to an experienced attorney. The attorney decides what happens next, communicates with the client and supervises anything leaving the firm. That division of labor matters because legal reasoning contains several layers. Some work can be handled through deterministic rules. Frontier models can then apply guidance and precedent to new situations. Human judgment remains responsible for high stakes advice and the review of agent output. John also stresses that the model is not making an isolated request to a generic system. Legal judgment is embedded in the way agents are designed, tested and called before live matters enter the workflow. We discuss legal engineering as the bridge between software and professional practice. Norm's legal engineers are trained attorneys who spend much of their time building, testing and validating agents. They work with practicing lawyers, clients and AI engineers to translate preferences, policies and matter specific requirements into systems that can operate within real workflows. Pricing is another part of the model. Norm Law prices selected matters around outcomes rather than hours. John acknowledges the limits. Some complex work can be scoped with enough confidence for a fixed price, while highly unpredictable litigation is much harder to price upfront. The opportunity is to give AI the incentive to examine additional documents and identify inconsistencies without increasing a client's bill for every human hour. The conversation then moves to the proposed Delaware AI Company initiative. John describes it as a regulatory sandbox for a legal entity managed by an AI agent while humans remain involved in its creation and supervision. His argument is that autonomous agents will take increasingly consequential economic actions, so policymakers must decide whether those activities happen within a recognized legal order or outside it. The proposal is designed to test questions around liability, disclosure, capitalization and government oversight before any broader model is adopted. John also believes companies deploying agents today should consider supervisory AI. If an operational agent can act faster and at greater volume than a person, a human team may be unable to inspect every decision. A second agent can evaluate the first against laws, regulations and company policies, with people retaining authority over exceptions and consequential outcomes. Does adding a supervisory agent create stronger accountability, or does it introduce another system whose reasoning must also be tested and questioned? Listen to the episode and share your thoughts with me.

Has enterprise AI finally reached the point where impressive demonstrations are no longer enough? In this episode, I speak with Murali Swaminathan, CTO at Freshworks, about the growing pressure on AI investments to deliver measurable business value. Murali has over 30 years of enterprise software experience, including roles at ServiceNow and CA, and now leads engineering and architecture teams at Freshworks. Murali believes the AI hype cycle is being replaced by an accountability cycle. Buyers want to understand reliability, governance, total cost of ownership, traceability, and the return generated by every deployment. They also want the ability to audit decisions, override outcomes, and use feedback to improve performance. Productivity alone provides an incomplete measure. Within service operations, companies can examine time to resolution, the volume of repetitive work automated, the number of issues completed without human intervention, and the quality of the employee's experience. Murali describes the difference between service-level agreements and experience-level agreements. Resolving a ticket within two minutes means very little if the employee's problem remains. The better question is whether AI completed the workflow and restored the person's ability to work. We also discuss why mid-market and agile enterprises provide a demanding test for AI. These companies have complex requirements but cannot absorb lengthy implementation programs, unclear pricing, or failed experiments. Murali recommends beginning with a limited process, measuring the result, establishing whether it can be repeated, and expanding only after it has proved reliable. Architecture plays an important role. Murali argues that ease of use begins beneath the interface. Configuration-led platforms can be upgraded as new capabilities arrive, while heavily customized systems can leave companies trapped on older releases. Autonomous service operations do not require removing people from every process. Murali uses the example of a printer incident. AI can read the ticket, classify the problem, route it to IT or facilities, and apply an automated fix when a trusted process exists. People retain responsibility for unusual, uncertain, or higher-risk decisions. Scaling this model requires cloud infrastructure that respects regional data residency, privacy, encryption, routing, and audit requirements. AI requests and diagnostic logs must remain within the correct geographic and regulatory boundaries. The conversation concludes with engineering skills. AI coding tools can generate software quickly, but engineers must understand architecture, usability, testing, and customer requirements. Companies also need rules determining which code can be reviewed by AI and which changes require human approval. Is your company measuring whether AI genuinely improves service operations, or is it counting deployments and calling that progress? Listen to the episode and share your thoughts with me.

What happens when an organization writes careful AI governance policies but its infrastructure cannot enforce any of them? In this episode of Tech Talks Daily, I speak with Sabina Anja, Chief Technologist at Broadcom within the VMware Cloud Foundation division, about the infrastructure controls required as AI agents move from generating answers to accessing data, calling APIs, modifying systems, and triggering work. Sabina brings experience from both sides of enterprise technology. She remembers cabling networks, dealing with unstable infrastructure, and receiving those weekend calls when downtime had already upset the business. That background informs her belief that ambitious AI programs cannot succeed without stable, observable, and enforceable infrastructure beneath them. Many organizations are repeating a familiar pattern. Business teams adopt AI services before IT has established visibility, ownership, or control. The terminology may have changed from shadow IT to shadow AI, but the management problem remains. Sabina argues that CIOs first need an inventory of agents, nonhuman identities, data access, processes, and accountable owners. The risk becomes greater because agents behave differently from people. They operate across multiple systems at machine speed and can perform repeated actions without appreciating the wider business outcome. An agent does not need malicious intent to cause disruption. Excessive permissions, flat networks, inconsistent access rules, and years of deferred infrastructure work can give it plenty of opportunities. Sabina recommends brokered access rather than direct access, alongside dedicated virtual machines or namespaces, microsegmentation, lateral security, east-west policy controls, and tamper-evident logging. Organizations also need to define which data an agent can view, modify, or move, especially when sovereignty and regulatory requirements apply. One of Sabina's most memorable ideas is to treat an AI agent like a superhuman contractor. It should have a defined purpose, a named manager, a clear access specification, an activity record, and an end date. Additional permissions should be earned through evidence of reliable behavior rather than granted on the first day. She also warns about agent debt. AI systems are developing rapidly, so an agent created today may become outdated within months. Sabina recommends assuming that many agents will expire after six to nine months rather than allowing forgotten systems and permissions to accumulate indefinitely. For CIOs wanting an immediate test, her advice is straightforward. Create an inventory of nonhuman identities with production access. Then select one agent and examine every part of the infrastructure it attempted to reach. The question is not simply whether the application produced the expected result. Leaders should ask whether the agent entered systems, networks, or data stores that nobody expected it to access. We also challenge the familiar claim that AI agents will take everybody's jobs. Sabina sees an opportunity to remove repetitive tasks and give technology professionals new skills, although she warns that agents may behave like teenagers armed with infrastructure permissions. They may not take your job, but they could become remarkably good at testing your patience. I'd love to hear your thoughts. Does your organization know how many AI agents have production access and who is accountable for each one?  

What does a rising cloud bill actually tell you about the value your business is creating? Eight years after our first conversation, I welcome Kunal, co-founder and CEO of Unravel Data, back to Tech Talks Daily. We compare the data infrastructure he was optimizing during the Hadoop era with today's enterprise stacks built around Databricks, Snowflake, BigQuery, AI pipelines, and autonomous agents. Kunal says Unravel Data has analyzed over 10 billion workloads across hundreds of enterprises. From that work, he argues that data platforms and infrastructure can account for up to 60% of cloud spending at some global businesses, while 30% to 40% of data platform spending may produce no business value. These are company claims, but they frame a problem many technology and finance leaders will recognize. The cloud bill arrives after thousands of individual engineering decisions have already been made. We discuss where cloud waste hides, including oversized clusters, hot storage holding cold data, abandoned pipelines, inefficient queries, duplicate datasets, and development jobs consuming production-level resources. The people creating those workloads seldom see the price attached to their decisions, leaving technology leaders with an aggregated bill that explains what was purchased but not why it was needed. AI adds another complication. Humans create workloads at human speed, while agents can generate queries, launch infrastructure, and consume tokens around the clock. An agent is designed to complete its task, not worry about whether a single query costs $5 or $5,000. Kunal argues that machine-speed consumption cannot be governed through monthly human reviews. We also discuss the difference between cost cutting and cost optimization, why aggressive reductions can damage performance and reliability, and how FinOps must connect cost with business outcomes. Kunal explains why leaders should measure cost per pipeline, model, agent, successful run, customer report, and business result. Finally, we consider the benefits and risks of autonomous data platform optimization. Kunal describes autonomy as a dial, with bounded, reversible, and validated actions earning wider authority as trust develops. Does your cloud bill show healthy growth, or is expensive waste hiding behind the headline number? Share your thoughts with me.

How can an independent safety evaluation remain meaningful when the AI inside a product may change after its next update? In this episode of Tech Talks Daily, I speak with Dr. Robert Slone, Senior Vice President, Chief Scientist, and Innovation Officer at UL Solutions. Robert has spent almost 30 years leading science, research, product development, and innovation teams. He now helps guide UL Solutions' scientific work across safety, security, and sustainability. Many listeners will recognize the UL Mark without knowing what happens behind it. Robert explains how UL Solutions tests products to their limits, which can involve setting them on fire, finding their breaking points, inspecting manufacturing facilities, and determining whether they meet defined safety requirements. That work began over 130 years ago when electricity was introducing unfamiliar risks. Today, the same broad question applies to artificial intelligence: how can society benefit from a new technology while understanding and managing the harm it could cause? The need is becoming increasingly visible as AI moves into healthcare, transportation, manufacturing, financial services, infrastructure, and consumer products. Robert recalls being approached about evaluating an AI-enabled teddy bear capable of talking with children. It is a memorable example of how decisions made inside an AI model can reach directly into everyday life. Robert organizes AI product safety around three pillars. The technical pillar considers robustness, risk management, functional safety, and whether the system performs its intended purpose. The ethical pillar includes fairness, bias, privacy, transparency, and explainability. Governance covers data management, product updates, accountability, and the complete operating life of the system. We also discuss one of the hardest problems in AI certification. Traditional products and software can be evaluated against a defined version, but AI systems may be updated, retrained, or affected by changing data. Robert explains why meaningful safety assurance requires version-specific testing, annual reviews, disclosure of significant changes, and eventually telemetry capable of identifying problems much closer to real time. For business leaders buying AI, the conversation provides a practical vendor checklist. Where did the training data come from? How was performance measured? What are the system's known limitations? How were privacy and bias assessed? Who takes responsibility if its behavior changes? Independent testing cannot promise that an evolving product will remain safe forever. It can provide evidence about the version evaluated, expose gaps, establish accountability, and create a process for monitoring future changes. What proof would you demand before allowing an AI product to influence an employee, patient, customer, or child? Listen to the episode and share your thoughts with me.    

Could allowing patients to choose their own appointment times help reduce missed visits and shorten NHS waiting lists? In this episode of Tech Talks Daily, I speak with Alison MacDonald, European Lead and Senior Vice President at Nordic Global. Alison brings an unusual combination of clinical and technology experience as a registered nurse who moved into digital health over 15 years ago. Her career began in community nursing, where she was asked to lead an electronic health record project because colleagues thought she was good with computers. What initially appeared to be a simple exercise in converting paper forms into digital records encouraged her to question whether healthcare could redesign the process rather than copy it onto a screen. We discuss Nordic's work with Cambridge University Hospitals NHS Foundation Trust on patient self-scheduling and automated earlier appointment offers. Before the program, missed outpatient appointments were removing valuable clinical capacity while administrative teams spent time calling patients and rearranging bookings. Cambridge introduced self-scheduling through Epic MyChart, allowing patients to select appointment times through the patient portal. Alison says the DNA rate fell from 5% to 2.2% during the program. According to the supplied results, over 20,000 patients successfully scheduled their own appointments and over 3,000 accepted offers to attend earlier when cancellations created availability. Patients moved appointments forward by an average of 16 days. Forty percent of accepted earlier appointments occurred within seven days of the offer, while 7% took place on the same or following day. Administrative teams also saved an estimated ten minutes for every self-booked appointment. Alison explains why patient control can improve attendance. People can choose times that work around employment, caring responsibilities, travel, and family life instead of receiving a fixed appointment through a letter or telephone call. Patients can also cancel or reschedule without waiting for somebody to answer the phone. The operational lesson goes beyond appointment booking. Healthcare systems may be able to recover existing capacity by examining missed appointments, theater scheduling, waiting list processes, pre-visit questionnaires, and patient communications before concluding that every problem requires additional staff or facilities. We also discuss where AI is producing practical results in healthcare. Alison points to medical imaging, emergency department triage, waiting list management, clinical documentation, and workforce deployment. She warns against discussing AI as one generic solution because each application requires a defined use case, suitable data, workable processes, governance, and staff adoption. Ambient clinical documentation offers one example. An AI scribe can record a consultation, prepare a structured note, and pass it to the clinician for review and correction. Alison cites an NHS evaluation reporting a 23.5% increase in direct patient interaction time and an 8.2% reduction in appointment length. Interoperability remains another major challenge. Healthcare journeys cross hospitals, primary care, community services, and specialist providers that may use different systems or a mixture of electronic and paper records. Even basic differences, such as one organization measuring pain on a five-point scale and another using ten points, can prevent reliable comparison. Alison recommends agreeing on common data standards, defining the minimum patient information required during care transitions, including interoperability requirements in procurement, and avoiding bespoke integrations that make future information sharing harder. Digital access also requires balance. Online services can improve convenience, but healthcare providers must retain appropriate alternatives for patients who lack digital skills, connectivity, confidence, or access. Could your healthcare organization improve patient access and staff capacity by redesigning one familiar process before purchasing another large technology platform? Listen to the episode and share your thoughts with me.

What would change if a customer could return three days later through a different channel and continue the same conversation without repeating a single detail? In this episode of Tech Talks Daily, I speak with Paul Adams, Chief Product Officer at Fin, the company previously known as Intercom. Paul has spent almost 13 years with the business and provides a candid account of how it abandoned its previous roadmap, placed a company-wide bet on AI, and rebuilt its products and working practices around AI agents. Our conversation begins with Fin's move from a customer service agent toward what Paul calls a single customer agent. The idea is that customers do not care whether their request belongs to sales, service, or customer success. They want the company to understand their situation and help them complete the task. Paul explains how AI agents can bring customer history, company knowledge, operational data, and business goals into the same conversation. This could allow an agent to resolve an issue, support a purchase, recognize a valuable customer, or transfer the conversation to a human without losing the context already provided. We also examine the economics behind poor customer service. Many companies are not ignoring customers through a lack of concern. They are receiving volumes of requests that cannot economically be handled by adding people alone. Paul says some Fin customers are resolving between 70 and 90 percent of customer queries through AI. Rather than seeing entire teams disappear, he is observing employees move into customer success, knowledge management, AI supervision, and higher-touch services. The episode also provides an unusually candid account of what it took to rebuild an established SaaS company around AI. Paul describes the process as brutal. Strategies were discarded, familiar processes were removed, and some people decided the new direction was not for them. His advice is to prioritize speed, place smaller experiments in front of real customers, and learn from evidence rather than waiting for every internal condition to become perfect. Paul also recalls working on early versions of mobile YouTube and Gmail when colleagues questioned whether anyone would watch video or answer email on a phone. Those stories provide a timely warning about judging new technology by its early limitations. Could AI agents finally give customers continuity across sales, service, and support, or will internal company structures remain the greater obstacle? Listen to the conversation and share your thoughts with me.

Can organizations trust an AI recommendation when they cannot understand the evidence, relationships, and previous decisions behind it? In this episode of Tech Talks Daily, I welcome back Jim Webber, Chief Scientist at Neo4j, to discuss the company's acquisition of GraphAware and its move from graph database provider to graph intelligence platform. GraphAware has worked with Neo4j for many years and developed Hume, an intelligence analysis platform used to connect and examine complex information. Bringing the two companies together gives Neo4j a direct role in applications serving police forces, governments, intelligence agencies, and other organizations handling connected data. Jim explains why context has become one of the biggest requirements for dependable AI. Enterprises already possess enormous volumes of data, but facts alone provide endpoints rather than the complete path leading to a decision. An agent needs to understand the knowledge available, the conversation taking place, and the record of previous decisions. It also needs to know which actions produced good outcomes and which produced poor ones. Jim compares these information layers to SimCity. Each can be viewed separately, but their greater value appears when they are combined. Knowledge, conversations, and decision traces can then help an agent understand why something happened and learn from the result. This introduces an interesting lesson from scientific research. Positive outcomes are frequently published, while failed experiments receive less attention. An AI agent needs both. Recording the breadcrumbs behind good and bad decisions provides the material required to improve its future behavior. We also discuss why large language models cannot understand every organization by themselves. Jim describes a model as a lossy compression of the internet. It can generate impressive natural language, but it does not automatically understand a company's policies, customers, history, evidence, or operating environment. Retrieval augmented generation can introduce relevant organizational information into the process. Graph RAG adds relationships between facts, helping the system understand how people, events, products, accounts, and other entities connect. According to research Jim references from the National Innovation Centre for Data, Graph RAG can improve accuracy while reducing costs by using fewer, higher-quality tokens. Explainability becomes especially important when AI supports decisions across policing, cyber defense, taxation, intelligence, banking, and government. A fluent answer may sound authoritative while containing a serious technical mistake. Jim shares an example from his own work where an agent confidently warned him about a "committed minority" inside a fault-tolerant computing protocol. The statement sounded plausible, but only a majority could commit within that protocol. Someone without Jim's technical knowledge might have accepted the recommendation and removed working code. This leads us to human oversight. Jim argues that the correct level depends on the consequences of the action. Automating a routine banking process with monitoring and safeguards may improve the customer experience. Ordering someone's arrest based solely on an agent's conclusion demands human involvement. We also consider digital sovereignty and why control over data has become a strategic concern for governments and large enterprises. Geopolitical instability, overseas technology dependencies, privacy requirements, and changing national policies are forcing leaders to ask where their data resides and whether they can retrieve or move it. Jim explains how Neo4j intends to offer organizations flexibility over where their information is stored and how it is deployed. The discussion also examines the opportunity for Neo4j and Hume to provide an alternative within a market where Palantir has held a powerful position. Looking ahead, Jim imagines intelligence analysts directing swarms of digital agents. Those agents could search data, connect evidence, identify relevant patterns, and present findings while humans retain responsibility for consequential decisions. If AI can connect information at machine speed, how do we ensure the person making the final decision can inspect the evidence and challenge the conclusion? Listen to the episode and share your thoughts with me.

What does an AI agent need to understand about your business before you allow it to make decisions and take action without waiting for human approval? In this episode of Tech Talks Daily, I speak with Kash Mehdi, Field CTO at Reltio, about the move from analytical AI that supports decisions to agentic AI that can execute them. Kash argues that leaders should begin treating AI agents as a workforce rather than another collection of software tools. A digital workforce needs training, boundaries, oversight, trusted information, and clear permissions before it can act safely. He uses the analogy of raising a puppy. When the puppy misbehaves, the problem may be inadequate training or poorly defined boundaries. AI agents present a similar leadership challenge. Organizations must ask what the agent has learned about the business and what authority it has been given. We discuss why model selection may be receiving too much executive attention. Kash describes four components of an agentic system: the model, tools, data, and context. Models are improving rapidly and tools are increasingly available, but business context remains incomplete across many enterprises. Data tells an agent a fact. Context helps it understand what the fact means within a particular customer relationship, geography, policy, or business process. Kash illustrates the difference with a pizza order. The data may confirm that someone is logged in, the model can interpret the request, and a tool can place the order. Context tells the system that it is Friday night, the customer is watching television, and they usually order pineapple and cheese pizza. The same principle becomes far more serious when an agent is dealing with medical equipment, supply chains, financial customers, or regulated information. It must understand which entities exist, how they relate, what information it may access, and which actions it has authority to complete. Kash identifies three requirements for safer autonomy: a governed source of truth, a live feedback loop, and enforceable permission boundaries. Trust must be built into the data and operating rules before the agent acts because the familiar human review step may no longer exist. We also discuss how governance changes when AI can execute decisions at machine speed. A poor decision made by one employee can usually be reviewed and corrected. A poor decision repeated automatically across thousands or millions of transactions can become a business incident before anyone intervenes. Kash shares examples involving restaurant menu launches, medical equipment deliveries, and call center offers. Each depends on current information and the relationships connecting customers, products, suppliers, locations, and previous interactions. For CIOs preparing today, Kash recommends building context around reusable entities rather than constructing an isolated data project for every AI use case. He points to Schneider Electric as an example where one unified foundation supported sales, shipping, operations, and marketing use cases. The conversation ends with a warning about slow data. Autonomous agents need current context because information that arrives after a decision has been made may no longer carry much business value. Kash predicts that the half-life of enterprise data will become a board-level measure. If a smarter agent can make a poor decision faster and with greater confidence, is your organization investing enough in the context, governance, and feedback needed to keep it on course? Listen to the conversation and share your thoughts with me.     Useful Links https://www.reltio.com/ https://www.reltio.com/datadriven/