Why Enterprise AI Agents Fail Beyond the Model With Databricks

Why do businesses replace the AI model when the failure may have started somewhere else entirely? In this episode of Tech Talks Daily, I speak with Richard Shaw, Technology General Manager for Databricks in the UK and Ireland. Richard leads the field engineering organization that works closely with customers on data and AI problems, giving him a practical view of what happens when promising agentic AI projects meet production workloads. Richard argues that the model often receives the blame because it is the most visible part of the system. The actual fault may come from stale data, missing business context, inconsistent permissions, an unsuccessful tool call, or another point in the workflow. Replacing the model before tracing the request from start to finish can recreate the same problem in a new place. This is why lineage, end-to-end tracing, and continuous evaluation matter once an agent moves beyond a controlled pilot. We discuss what a production-readiness rehearsal should include. Richard recommends realistic data, realistic user volumes, unauthorized requests, ambiguous questions, failed tool calls, and tests of what the agent should refuse to do. Teams also need agreed standards for quality, security, cost, and auditability, along with a clear decision about which actions an agent can complete independently and where a person must review or approve the result. The conversation also looks at model choice and infrastructure cost. Richard believes the strongest test is performance on the organization's actual work rather than a benchmark leaderboard. A frontier model may suit complex reasoning, while a smaller or open-weight model may perform routine extraction or classification at a lower cost. Access policies, observability, and spend controls need to remain consistent as those model choices change. Conversational analytics creates another governance challenge. Databricks customers such as Virgin Atlantic and Repsol are using natural-language tools to make company data easier for employees to question. Richard says wider access should preserve existing permissions, ownership, definitions, and lineage. An answer becomes far more useful when the user can see where it came from and which team owns the information behind it. We also cover the boundary between historical analytical data and fast operational workloads. Richard describes how Databricks positions the lakehouse for broad enterprise context and Lakebase for immediate reads and writes, such as updating an account, placing an order, or storing agent memory, while keeping both connected to a common data and governance base. Are companies ready to trace and test the whole AI workflow, or are too many treating the model as both the hero and the culprit? Listen to the episode and share your thoughts with me.  

How can businesses turn growing investment in AI infrastructure, cloud capacity and devices into outcomes that employees, customers and finance teams can actually measure? In this episode of Tech Talks Daily, I speak with Neil Sawyer, who manages HP's business across Europe, the Middle East and Africa. Neil works with companies across one of HP's largest global regions as they move from AI experimentation into wider deployment, making him well placed to discuss what happens when early enthusiasm encounters cost, security, governance and the realities of the workforce. We begin with the gap between building AI capacity and applying it to a business problem. Data centers, models and powerful devices provide options, but the investment only becomes useful when a company identifies the workflow it wants to improve. Neil argues that leaders should begin with the outcome, understand where AI can remove friction and decide how they will measure productivity, employee experience and business performance before buying another layer of technology. That raises a difficult question about productivity. If AI helps someone complete a task faster, does the organization use that saved time to improve the work, develop new ideas and give employees room to think, or does it simply add another task to the queue? I share my own experience as a business of one, where every efficiency gain has a habit of becoming extra output rather than a Wednesday afternoon at the cinema. Neil compares the current moment with earlier periods of industrial change and makes the case that automation should release people from repetitive administration so they can contribute creativity, judgment and higher value work. We also discuss why AI costs are becoming a boardroom issue. Token based services and agentic systems can produce growing and unpredictable bills as adoption spreads across a company. Neil explains why every workload does not need the same model or environment. Large language queries may benefit from cloud capacity, while sensitive data, company specific information and some recurring tasks may be better suited to local or on device processing. The decision affects cost, responsiveness, privacy, security, data sovereignty and environmental impact. Neil describes HP's view of hybrid AI, including devices with neural processing units and Z by HP Boost, which can connect available workstation GPU resources. He also explains why device refresh decisions should reflect workforce personas. A data scientist, account manager and office administrator may work for the same company, yet their computing needs can be very different. Mapping technology to the employee's role can help a business spend with greater discipline while giving people the performance they need. The conversation also covers governance and measurement. Informal use of public AI services can be difficult to see, assess or manage. Neil recommends giving employees an approved AI toolkit, using enterprise services that provide telemetry and examining how technology availability and performance affect the employee experience. Adoption figures can show that a tool is being used, but they do not prove that it is improving an outcome. We finish with a practical checklist for leaders. Define the outcome, identify the workflow, determine where each workload should run, calculate the cost, agree the measures of success and put clear controls around data, privacy and cybersecurity. That approach gives cloud and device based AI distinct jobs within the same business strategy. How is your organization deciding which AI workloads belong in the cloud, which should run closer to the employee, and whether the investment is producing measurable value? Share your thoughts with me.

What could your team achieve if a document that previously took two hours could be created in eight minutes? In this episode of Tech Talks Daily, I speak with Oskar Konstantyner, Chief Product Officer at Templafy, about the rapid adoption of AI agents across enterprise document workflows and what those productivity gains mean for knowledge workers. According to Templafy's proprietary usage data, AI agent adoption among its enterprise users grew from virtually zero in October 2025 to 53% by June 2026. Its analysis found that documents created without agents had a median completion time of two hours and an average of 5.6 hours across 16,000 sessions. With AI agents, the median fell to eight minutes and the average to 27 minutes across 14,000 sessions. The most common documents included pitch decks, company communications, sales materials, and product roadmaps. However, Oskar cautions against treating speed as the final measure of AI productivity. We discuss an accounting firm that could not respond to thousands of tenders because it lacked the capacity to create enough proposals. Faster document production could allow that business to participate in additional opportunities while applying its knowledge about what makes a winning submission. The benefit comes from increased commercial capacity and stronger results, rather than counting recovered hours alone. Oskar also explains what happens during those eight minutes. AI can locate relevant information, find approved content, recommend a presentation structure, apply previous lessons, and complete much of the production work. Humans remain responsible for original thinking, client judgment, factual accuracy, and final approval. In many cases, the agent may produce 60% to 90% of the document, but the beginning and end of the process remain human-led. The conversation also considers the growing volume of generic AI documents. A business already has approved slides, company descriptions, brand assets, legal statements, and sales messages. Regenerating all that material wastes tokens and risks inconsistency. Oskar argues that agents should determine when existing content should be reused, when rules should be applied, and when something genuinely new needs to be created. We also discuss why AI adoption improves when agents appear inside PowerPoint, Claude, OpenAI, and Copilot. Most employees are unlikely to abandon familiar workflows every time another AI application arrives. Is your company measuring AI success through minutes saved, or through the additional business those minutes make possible? Listen to the conversation and share your thoughts with me.

What happens when an AI system gives a customer the wrong price, misrepresents a product, or recommends a competitor using outdated information? In this episode of Tech Talks Daily, I speak with Alex Sherman, co-founder and CEO of Bluefish AI, about the growing influence of AI-generated answers on brand reputation, product discovery, and purchasing decisions. Search once gave companies a reasonably visible path between a customer's question and the websites informing the answer. AI changes that relationship. Systems such as ChatGPT, Gemini, Rufus, and other assistants combine information from many sources into a single response. Customers may never visit the original pages, read the supporting evidence, or know which source carried the greatest influence. Alex explains that inaccurate AI answers are not always extraordinary hallucinations. Models learn from an internet filled with conflicting product descriptions, outdated specifications, opinionated reviews, creator videos, and content generated by other AI systems. Bluefish says its monitoring has found inaccuracies or misleading portrayals in roughly 10% to 20% of the AI responses it tracks. A company may publish a concise product page containing a few hundred words, while a customer writes a lengthy Reddit post describing why they love or hate the same product. The longer and more detailed source may give an AI model material it can use across many customer questions, even when that source presents an extreme or unbalanced view. Bluefish also reports finding small YouTube creators carrying considerable influence over how models describe certain brand attributes. This creates a new responsibility for marketing teams. Visibility alone is no longer enough. Businesses need to understand whether they appear in AI answers, how positively they are presented, whether the facts are accurate, which sources influence the response, and whether that exposure contributes to a sale. Alex describes how Bluefish's AI Accuracy tool monitors model responses, flags potential errors, identifies the cited source, examines the content behind it, and helps a company determine what action could correct the result. The source may be an external article, but the problem could also come from the company's own website. A model might confuse this year's device with last year's version because the distinction between their specifications was unclear. Correction is only part of the process. Brands must measure whether new content changes the AI response and whether the revised answer cites the information they supplied. This turns AI accuracy into an ongoing measurement discipline rather than an occasional reputation exercise. The stakes rise further with agentic commerce. Alex believes companies will increasingly serve two audiences: the person buying the product and the AI agent researching or acting on that person's behalf. Marketing teams will need to understand what agents read, how they evaluate choices, and why a recommendation resulted in a purchase or a lost customer. Our conversation ends with a wider concern about convenience and choice. Alex compares AI discovery with opening Netflix and accepting the options placed on the first screen. AI can make research faster and easier, but relying on synthesized answers may weaken our willingness to search beyond what an algorithm selects for us. I'd love to hear your thoughts, so how much influence should AI have over what consumers discover, compare, and ultimately buy?

What happens when an AI assistant becomes the first place a potential customer learns about your company, but the answer it provides is inaccurate, outdated, or influenced by a competitor? In this episode of Tech Talks Daily, I speak with Justin Seibert, founder and president of DOM, also known as Direct Online Marketing, about generative engine optimization, AI search visibility, and the brand narratives being assembled by ChatGPT, Claude, Google AI, and other answer engines. Justin has worked in digital marketing since 2001 and founded DOM in 2006. He has watched search marketing develop from the early days of measurable clicks and online leads into a system where an AI assistant may answer the customer's question before that person visits a company website. We discuss why appearing in AI search results tells only part of the story. Companies must also understand what the system says about them, whether the information is accurate, and which sources influenced the answer. Reviews, Reddit discussions, press coverage, social media, newsletters, company websites, and third-party platforms can all contribute to the narrative. Justin recommends separating prompts into four groups: branded searches, competitor searches, top-of-funnel informational questions, and bottom-of-funnel transactional questions. Each category requires different measurements. Branded prompts reveal sentiment and accuracy, while transactional prompts show whether the company reaches the shortlist presented to a motivated buyer. He also explains why companies should define the niche in which they want to become the preferred recommendation. A broad insurance company may struggle to dominate every AI conversation, for example, but it could establish authority among a specific age group, product category, or geographic market. According to Justin, visitors arriving through AI referrals can convert at rates two to four times higher than traditional search visitors. He believes these buyers often arrive better informed, with a shorter list of options and a stronger intention to make a decision. That makes exclusion from an AI-generated shortlist a serious commercial risk. We also consider what happens when paid placements become common inside AI experiences. Justin argues that companies building organic authority today may retain an advantage as AI advertising becomes increasingly crowded and expensive. Finally, Justin offers a practical audit any leader can perform. Log out, use a private browser, check from relevant countries, compare the company with its competitors, ask the AI why it produced its answer, and inspect the sources it cites. Have you checked what AI assistants say about your company, and would a potential customer trust the story they find? Listen to the conversation and share your thoughts with me.    

Can electricity grids built for an earlier era support AI data centers, expanding manufacturing, electric vehicles, severe weather, and rising customer expectations at the same time? In this episode of Tech Talks Daily, I speak with Mark Hollis, utility executive advisor at SAP, about the pressures reshaping the utility industry and the practical choices available to leaders today. Mark spent over 15 years at Duke Energy before moving to SAP, where his work gives him visibility into utility organizations across North America. Mark describes a combination of load growth, disruption, long construction timelines, and regulation. Data centers are receiving much of the attention because of the electricity required by AI, but he argues that they are only part of the story. Manufacturing is returning to parts of North America, transport and heating are becoming increasingly electrified, and utilities must prepare for wildfires, hurricanes, winter storms, and other events that affect generation and delivery. The obvious response is to produce additional electricity, but every option comes with physical and commercial limits. Wind and solar contribute to the generation mix, although output depends on conditions. Small modular nuclear reactors could support future demand, but commercial deployment takes time. Batteries can store electricity and release it later, but they do not generate the power they hold. Customer programs can also reduce pressure at busy periods, including arrangements that allow a utility to adjust connected thermostats by a few degrees. This makes modernization a portfolio of decisions rather than a single bet. Utilities must decide how to divide capital among generation, transmission, resilience, customer systems, and new technology. The people who understand existing processes are often the same employees needed to design and implement replacements. At the same time, information technology and operational technology are becoming increasingly connected, forcing companies to reconsider how business functions share information and how technology decisions support outcomes across the enterprise. AI creates another tension because it contributes to electricity demand while also offering utilities new ways to work. Mark says most utilities he meets are cautiously optimistic. Their questions include where to begin, whether the value is proven, how long adoption will take, and whether poor data must be fixed before useful work can start. He warns against choosing the hardest problem first or judging a business process by the forgiving standards people accept from consumer AI tools. His advice begins with the business problem. Automating an inefficient process can make it more expensive and harder to correct. Utilities should define what they need to improve and why, establish connected data with the right business context, set clear quality requirements, and retain human review where errors could affect customers, safety, finance, or regulatory obligations. Mark brings this to life with several utility AI use cases. AI could summarize customer interactions across field service and contact center systems, allowing the next employee to understand what happened previously. It could review billing exceptions during unusually hot or cold periods and support earlier customer communications when consumption is likely to produce a much higher bill. He also discusses using AI to summarize lengthy rate-case rulings before approved changes enter billing systems. During outages, an AI system could help dispatch crews by considering skills, equipment, parts, certifications, location, safety, and customers with medical needs. A human dispatcher could then review and approve the recommendation rather than building the complete schedule manually. The opportunity is real, but so are the limits. Utilities operate regulated infrastructure where reliability, safety, auditability, and public trust cannot be treated as optional features. Where should the industry begin, and which use case offers the right combination of low effort, meaningful impact, and manageable risk? Listen to the conversation and share your thoughts with me.  

What happens when an AI system moves beyond recommending the next sales action and begins running a connected revenue workflow? In this episode of Tech Talks Daily, I speak with Abhijit Mitra, CEO of Outreach, about the operational work required to turn agentic AI into measurable revenue outcomes. Abhijit argues that adding another AI tool can create extra complexity when customer data remains fragmented and applications cannot share context. The starting point is the business process: what problem is being solved, which data supports it, what agents may do, and where human judgment remains necessary. We discuss the difference between a recommendation and an autonomous action. Revenue teams may begin with supervised spot checks while an agent researches accounts, identifies prospects, drafts messages, and runs targeted campaigns. Once the data and results earn confidence, parts of that process can operate continuously. Multi-step work adds another requirement because the output of one agent must become useful input for the next. Research, outreach, coaching, forecasting, and expansion cannot deliver their full value as isolated tasks. Context runs through the entire conversation. Abhijit describes the customer history, product usage, prior interactions, industry signals, buyer priorities, and organizational memory that can turn a generic model response into a commercially useful action. He says access to a frontier model alone does not create a revenue platform because each business still needs its own context layer and controls. We also discuss Outreach's MCP Server and Client, which allow agents to receive information from surrounding systems and return their work to platforms such as Salesforce Agentforce, OpenAI, Anthropic, and Microsoft. That interoperability creates a governance question. Abhijit's advice is to give an agent the same roles, permissions, and data access as the person or team it supports. If the platform cannot control access at that level, he advises companies to wait. The conversation then turns to the changing software model. Abhijit describes a move from assigning SaaS licenses to employees toward deploying agents with particular skills and a defined capacity. That makes workflow design and measurement increasingly important. He recommends establishing a baseline before rollout and tracking revenue against cost through indicators such as win rate, deal size, quota attainment, pipeline movement, forecast accuracy, and seller productivity. Can revenue teams use agents to remove administrative work while protecting customer trust and keeping relationship building human? Listen to the episode and share your thoughts with me.

What would happen if specialty insurance underwriters and risk capital providers could work from the same timely, detailed information? In this episode of Tech Talks Daily, I speak with Jeff Radke, CEO and cofounder of Accelerant, about the infrastructure behind specialty insurance and why his team chose to rebuild it around a data-driven risk exchange. Jeff has spent decades in reinsurance broking, reinsurance underwriting, and specialty insurance across New York, Bermuda, and London. That experience gave him a direct view of a process he describes as expensive, slow, and supported by weak data flows. Jeff explains that Accelerant backs independent underwriting specialists who focus on narrow areas of risk, from pickleball courts to New York brownstones. These teams need the regulatory ability to issue policies and the capital required to support them. Accelerant connects those needs through a shared platform that routes risks to insurance companies and distributes them across a group of capital providers. The economic argument is striking. Jeff says the traditional chain can consume about 40 cents of each premium dollar in expenses and overhead. Accelerant instead seeks portfolio-level solutions across a diverse book of business, reducing repeated negotiations and transfers between intermediaries. He also addresses the tradeoff created by concentrating information in one exchange, including the need to protect data and cash flows when participants depend on a shared platform. Data quality sits at the center of the conversation. Jeff says older policy administration systems often retain only eight to twelve exposure characteristics for each policy, while Accelerant captures over 60 on average. That fuller record gives underwriters and capital providers more information when selecting risk, reviewing performance, or investigating a problem. We also discuss why smaller underwriting organizations may hold an advantage over established insurers. New teams can begin with data at the center of their operating model, while larger companies must change processes built around older systems. Jeff argues that the biggest barrier is often mindset rather than budget. AI has a defined role in that model. Accelerant uses agentic AI to organize varied incoming data, identify products whose performance needs attention, support portfolio construction, and improve internal operations. Jeff draws a firm boundary around responsibility: underwriters remain accountable for underwriting outcomes. His father's advice captures the principle neatly: do not blame the wheelbarrow; responsibility belongs to the person driving it. Could shared data and lower operating expense return more value to policyholders while preserving human judgment? Listen to the conversation and share your thoughts with me.  

Why can an AI pilot produce an impressive result and still fail to create measurable value for the business? In this episode of Tech Talks Daily, I speak with Jitendra "Jit" Putcha, chief operating officer at Tredence, about what the company calls the last mile of AI. This is the gap between generating an insight and making sure it reaches the person, process, and decision where it can produce a useful result. Jit argues that many companies are facing an execution problem rather than a shortage of technology. Models are widely available, and teams can build demonstrations at remarkable speed. The harder task is redesigning a complete workflow so that employees can use AI without leaving one system, checking another, and manually carrying information between the two. Trust, explainability, governance, and continuous evaluation also become harder once a pilot moves from a small group into everyday enterprise operations. We discuss why resistance from employees should not be dismissed as stubbornness. People are trying to understand what AI means for their role, judgment, and future. Jit recommends translating the program into a practical question: how will this make somebody's Monday morning better? He describes human and AI agent teams, along with workshop-based learning that allows employees to solve real problems, test the tools, and understand where human judgment remains necessary. The conversation then turns to measurement. Jit challenges technology teams to move away from vanity measures such as the number of models built or code interactions recorded. Instead, he recommends examining margin improvement, loss reduction, cycle time, conversion, customer satisfaction, and other measures already understood by the business. Jit supports the argument with several customer examples. He says one retail workflow reduced analyst effort by 70 percent, while a manufacturing supply chain platform reportedly produced $10 million in first-year savings. He also describes a supermarket forecasting program that reportedly produced close to $200 million in value and replenishment match rates above 90 percent, along with another modernization program associated with a reported $100 million loss reduction. These are Tredence customer examples shared by Jit during the interview and should be presented as attributed company claims. We also discuss an AI-native operating model built around three layers: foundation, intelligence, and experience. Data infrastructure and governance support the foundation, intelligence turns data into decisions, and the experience layer brings those decisions into human workflows. Jit adds five supporting elements covering human and agent teams, execution rhythm, business measures, the technology ecosystem, and company culture. His final advice is refreshingly practical. Escape the demo trap, prepare the whole organization for deployment and ongoing operation, consider an internal marketplace for reusable agents, and give the supposedly boring work a larger role. Data hygiene, evaluations, governance, change management, and runbooks help AI continue producing value after the launch presentation has ended. Is your company measuring the number of AI projects it has created, or the business outcomes those projects have changed? Listen to the conversation and share your thoughts with me.

What if the vehicles already traveling through our towns and cities could report road damage before a pothole becomes dangerous and expensive? In this episode of Tech Talks Daily, I speak with Jonathan Selbie, CEO of Stockholm-based Univrses, about using computer vision and vehicle sensor data to give road authorities a much clearer picture of the infrastructure they manage. Jonathan's career has taken him from Formula One engineering at Red Bull Racing to unmanned aircraft and autonomous navigation, before bringing those lessons into automotive AI and road monitoring. Univrses can work with cameras installed by vehicle manufacturers or retrofit cameras and processors to vehicles already operating around a city. Waste collection trucks and taxis can continue their normal routes while gathering information about surface damage, obscured traffic signs, roadworks and deteriorating road markings. The video is processed on the vehicle, and authorities receive mapped findings and recommended actions rather than hours of footage. Jonathan explains that some Swedish cities moved from road condition surveys every five years to updates every two weeks. According to the figures discussed in our conversation, one council reduced its pothole count from around 3,000 to 900 in six months. He also says repairing damage at an early stage can cost up to 15 times less than waiting for it to become a major pothole. That changes road maintenance from an expensive reaction into a regular process based on current evidence. We also discuss whether road infrastructure is ready for autonomous vehicles. Waymo uses a broad mix of cameras, radar and lidar alongside detailed maps, while Wayve is pursuing an approach designed to adapt to changing roads without relying on the same level of pre-mapping. Jonathan explains why faded lane markings can reduce the performance of driver-assistance systems, creating a useful feedback loop in which vehicles rely on roads and also provide data to maintain them. The conversation also covers Pirelli's 30 percent investment in Univrses and the combination of connected tire data with forward-facing cameras. A tire can feel the road surface while a camera sees what lies ahead, giving vehicles and road operators different views of the same conditions. Jonathan also addresses privacy, explaining that Univrses detects and blurs faces and license plates before deleting the original imagery. This is a practical example of AI producing value through existing fleets, frequent data and earlier decisions rather than another expensive technology project searching for a problem. Could the cars, taxis and service vehicles already using our roads become part of the infrastructure maintenance system, and would you be comfortable with that if privacy protections were clear? Please share your thoughts.