Inside the Agentic SOC Where Humans and AI Defend at Machine Speed With Barracuda

What does a security operations center need when attacks are arriving at a speed and volume that human analysts cannot manage alone? I recorded this episode with Adam Khan, VP of Global Security Operations and AI Security at Barracuda, during the 20th anniversary of the Barracuda Tech Summit in Alpbach, Austria. Adam has spent over 25 years in technology and security. When we last spoke at the event, he used Home Alone and soccer to make complex security topics easier to understand. This year, expectations were high, and he arrived with Formula One. The comparison begins with what spectators see. Attention naturally falls on the car and driver. Behind them sits a much larger operation involving engineers, strategists, mechanics, simulations, telemetry, and rapid decisions. Adam believes modern security works in a similar way. Customers want to run their businesses, while a largely unseen combination of analysts, threat intelligence, automation, and AI works behind them. That operating model has developed into what Barracuda calls the Agentic SOC. AI agents follow the same playbooks analysts use when examining endpoints, malware connections, artifacts, threat intelligence, identity behavior, and other signals. They can complete repeatable investigative work quickly and consistently across volumes that have grown from hundreds of alerts to thousands or millions. Adam says Barracuda now has hundreds of agents with hundreds of individual skills. These agents can support the process from triage and intelligence gathering through correlation and response. When the system has high confidence that an ordinary user account has been compromised, it may disable that account. If the incident involves an administrative account capable of locking down an entire customer environment, a person must confirm the action. This matters because an AI system can misclassify an event or reach a conclusion that requires additional context. Adam describes feedback mechanisms through which people review decisions, identify mistakes, and feed those findings back into the system. Barracuda also records an audit trail of the actions and queries performed by its agents. One of the most surprising details concerns employment. While headlines frequently associate AI with reducing headcount, Adam says his team has doubled since adopting it. Analysts previously occupied with repetitive investigation have moved into threat hunting, model development, prompt engineering, and attack and defense exercises. The team is also attacking its own systems so that its agents can learn from emerging techniques before a genuine incident occurs. The audience saw this operating model during Adam's keynote. Attendees used their phones to launch controlled business email compromise, QR-code phishing, and ransomware scenarios against Barracuda's attack and defense environment. The platform then analyzed and blocked the activity while the audience watched. Adam says approximately 395 attacks were initiated during the demonstration and all were successfully blocked. That result comes from a controlled Barracuda demonstration rather than an independent test, but it gave attendees a rare view of the speed required during an active incident. We also discuss how Barracuda Managed XDR uses behavior and telemetry across email, endpoints, cloud services, identities, networks, and other technology. An employee traveling with a familiar laptop and phone should not create the same response as an unknown device attempting an unusual login. Historical patterns, device identifiers, signatures, and location data can help reduce unnecessary alerts while highlighting activity that deserves attention. For Adam, the purpose of AI is to increase the speed and reach of security experts rather than remove them. People determine strategy, examine high-consequence decisions, test systems, and remain accountable for customer outcomes. Could the Agentic SOC give security teams the speed they need without surrendering the judgment and accountability customers expect? Listen to the episode and share your thoughts.

How should security leaders respond when AI-powered attacks compress detection and response windows from minutes to seconds? At Barracuda TechSummit 26 in Alpbach, Austria, I spoke with Rohit Ghai, Chief Executive Officer at Barracuda. The conversation took place exactly one year after Rohit joined the company, giving us an opportunity to discuss what brought him to Barracuda, what he inherited, and how his first year has influenced his plans for the business. Rohit explains that Barracuda's focus on smaller and resource-constrained organizations was an important reason he accepted the role. A cyber incident can threaten the survival of a smaller company, particularly when it has a lean IT team and limited access to specialist security knowledge. For these businesses, Rohit argues that AI-supported and increasingly autonomous security is a practical requirement. We discuss Barracuda's platform strategy and why genuine integration must extend beyond a shared interface. Rohit compares loosely connected product portfolios to supermarkets. Customers may find it easier to purchase several products from one supplier, but that commercial convenience does not mean the products share data or produce a coordinated response. Traditional tool sprawl forced analysts to interpret information across several screens. Agent sprawl could introduce systems that act independently, disagree with one another, or take conflicting actions. Rohit believes security platforms must connect information across email, identity, applications, data, and infrastructure so they can reason across the complete attack sequence. Identity is another major part of the discussion. Barracuda's acquisition of Evo Security addresses privileged access for managed service providers and smaller businesses. Rohit expects machine and non-human identities to greatly outnumber human users, raising questions about excessive privileges and how organizations grant temporary access to autonomous agents. We also discuss the economics behind AI security. Rohit explains how Barracuda is adapting its value reports to account for token consumption as well as staffing, software, and security outcomes. Barracuda absorbs the direct token costs associated with its own AI capabilities and selects different models according to the task, an approach intended to keep its products affordable for smaller customers and MSPs. Rohit is cautious about calls for the AI industry to slow development. Coordinating a worldwide slowdown between companies and countries would be extremely difficult. He also argues that cyber defenders cannot pause while attackers continue using widely available models to improve their campaigns. The conversation ends with a wider leadership question. Rohit believes intelligence will become widely available, making empathy and trust more valuable. AI can generate an answer, but customers, partners, and security professionals must decide whether they trust the organization acting upon it. Will connected AI security platforms reduce complexity, or could autonomous agents introduce a new form of operational risk?  Share your thoughts.

What happens when enterprises spend trillions of dollars on AI but the systems underneath it still cannot provide the context those models need to make reliable decisions? In this episode of Tech Talks Daily, I reconnect with Christian Monberg, CTO at Zeta Global, to examine what separates AI experimentation from production systems that organizations can actually trust. Our previous conversation focused on how businesses could use AI to scale marketing without losing the human connection with customers. This time, we move deeper into the technology underneath those experiences. Christian explains why disconnected tools and fragmented data remain barriers to AI adoption, and why Zeta rebuilt its data architecture using Palantir Foundry. We discuss the role of context graphs in connecting customer identity, business objectives, previous decisions, campaign history and outcomes so AI systems can understand more than isolated pieces of information. We also examine one of the biggest questions surrounding agentic AI: when should businesses allow an AI agent to take action? Christian shares what enterprises need around explainability, permissions, observability and learning loops before AI systems can safely move from recommendation to execution. With global AI spending expected to reach $2.59 trillion in 2026, the conversation ultimately comes back to a simple question: how can technology leaders prove that their AI investments are producing measurable business value?

Why is employee AI use rising so quickly while measurable business value remains difficult for many organizations to find? In this episode of Tech Talks Daily, I speak with David Martin, Senior Partner and Global Leader of People and Organization at BCG, about the firm's fourth annual AI and workforce report. The research surveyed 11,749 employees across 14 countries and points to a growing divide between companies that distribute AI tools and companies that give people a clear plan for changing how work gets done. BCG reports that 74 percent of frontline and nonmanagerial employees now use AI regularly, an increase of 23 percentage points from the previous year. Adoption, however, is only part of the story. The report says 71 percent of employees receive little or no guidance about what to do with the time AI frees, while over half are not redirecting that capacity into strategic work. David describes one of the research findings that best captures the problem. Companies where employees understand the strategic direction but rate their AI tools poorly can realize greater value than organizations with strong tools and limited strategic clarity. Better technology helps, but its impact remains small when employees do not understand which business problem they are solving or how the operating model should change. The report connects clearer strategy with a roughly 25 percentage point increase in measurable business impact when companies redesign workflows from end to end or create new business models. BCG says strong tools without that clarity produce an improvement of roughly five percentage points. Companies that redesign workflows also outperform tool-only adopters by 23 percentage points on measurable business impact, 22 points on time saved, and 20 points on job satisfaction. David explains what redesign looks like in practice. Giving software engineers stronger coding tools may improve part of a development task, but keeping the overall product lifecycle unchanged limits the result. A deeper redesign considers how research, product management, engineering, and decision-making operate together, then changes roles and processes around the capability of AI. The objective is a better business outcome rather than a faster version of the same work. This distinction also explains why promising pilots fail when companies attempt to expand them. A pilot can prove that a model works inside a controlled environment. Wider deployment tests whether the organization surrounding that model works. David cites BCG research indicating that 70 percent of the factors determining whether AI scales with a return relate to people, organization, and process. Talent, operating models, cross-functional teamwork, incentives, learning, and leadership all become part of the result. Measurement must also move beyond adoption. David argues that the final metrics remain familiar business outcomes such as conversion, competitive win rate, price realization, cycle time, and inventory performance. A pilot can use controlled comparison to test whether AI changes one of those outcomes. The missing management step is often accountability. If several executives share ownership but nobody is responsible for the return, the investment can continue without a clear test of success. There is a case for broad experimentation because it can build familiarity and surface ideas. David warns that hundreds of isolated use cases can also fragment investment, increase risk, and save small amounts of individual time without producing company-level value. His preferred balance combines focused governance with structured opportunities such as hackathons, where employees contribute ideas but the organization selects which ones receive investment. The workforce findings add an important human dimension. BCG says 67 percent of regular AI users report higher job satisfaction, while 41 percent also report higher cognitive load. David connects that tension with the effort required to assign work to agents, evaluate quality, and keep those agents operating. He refers to separate BCG research called AI Brain Fry, which found productivity rising as employees managed additional agents until a limiting point. In that research, productivity fell when workers moved beyond managing three agents. Training remains another stubborn problem. The report says 72 percent of employees believe AI has changed skill expectations, and nearly half say their role is moving toward directing and managing AI. Only 36 percent feel they have received enough training, a figure David says has not improved despite new learning programs. His recommendation is in-context training that brings AI into daily work, followed by peer discussion about what worked, what failed, and how behavior should change. The purpose of recovered time may be the most revealing management question of all. David says employees report saving an average of around eight hours a week, but many use that time to perform additional versions of the same tasks. That can become demoralizing if greater output benefits the company without giving employees room for learning, infrastructure improvement, experimentation, or new product work. Leaders need to explain where the capacity should go and why. Clear communication also reduces fear. Automation targets introduced without an explanation of strategy can leave employees assuming that efficiency is a code word for job loss. When leaders explain whether AI is intended to improve customer experience, create growth, reduce cost, or change the business model, employees have a better basis for understanding what is expected of them. If strategic clarity is producing greater value than better tools, should the next AI investment begin with another platform or with a decision about how the work itself must change? Listen to the episode and share your thoughts.   Useful Links AI at Work: Strategy Matters More Than Tools When Using AI Leads to "Brain Fry" When Everyone Uses AI, Companies Risk Losing Critical Skills LinkedIn – BCG on the CHRO Agenda    

What will customers remember about a business when the voice answering their questions becomes the main expression of its identity? In this episode of Tech Talks Daily, I speak with Ruth Zive, who leads marketing at Voices, about the business and human questions surrounding AI voice. Voices is an enterprise marketplace and platform where companies can find professional voice actors, license AI voices with consent, and source custom voice data for training models. The company says its global talent network includes millions of performers and has served brands including Microsoft, Shopify, and Cisco. For years, much of the AI voice debate focused on whether synthetic speech could sound convincingly human. Ruth believes the commercial conversation has moved toward provenance, brand integrity, permission, and usage rights. A voice can sound polished while exposing a company to reputational damage if the performer did not understand the use, the license is unclear, or the same generic voice appears in a competitor's customer experience. Voices research cited during the interview found that 79 percent of business leaders believe inauthentic AI voices could damage brand perception. Ruth also says almost half of enterprise decision makers regard tone and emotional expression as the most important vocal factor in authenticity. That matters when a customer is frustrated, confused, or asking for help. A voice that sounds human but responds without suitable emotion can weaken trust at the exact moment a company needs to earn it. Ruth describes responsible licensing through three ideas: compensation, control, and consent. The performer should understand how the voice will be used, retain an agreed degree of control, and receive payment that reflects the commercial use. Those decisions need to appear in the contracting, licensing, and entitlements before a model is trained or placed in front of customers. We also consider the economic effect on professional performers. AI voice can change existing work, but Ruth argues that it can also create additional assignments when licenses are written carefully. An actor might provide the voice for an in-car assistant while continuing to record commercials in unrelated categories. Other opportunities include contact center experiences and the creation of specialized voice data used to train models. The positive case depends on clear boundaries and fair commercial terms rather than unlimited reuse. The brand question may become even larger as customers move from websites and typed interfaces toward spoken conversations. Ruth points to BMW's careful selection of voices based on customer profile, tone, language, accent, and how each performer sounded inside the vehicle cabin. Her advice is to treat a voice as a long-term brand asset, test it in the setting where customers will hear it, and confirm that the company has the required rights before deployment. Voice AI offers companies a more natural customer experience and gives performers access to new forms of paid work. It also raises difficult questions about disclosure, ownership, exclusivity, and trust. Should every company now have a formal policy for choosing, licensing, and governing the voice that speaks on its behalf? Listen to the conversation and share your thoughts with me.

Can ideas developed thousands of years ago help leaders make better decisions about AI, data and digital transformation today? In this episode of Tech Talks Daily, I speak with Alfonso Asensio, author of Digital Wisdom: Leading Transformation With the Sophia Factor and head of data measurement for global clients at Google in Tokyo. Alfonso has spent his career working across data, digital business and global client leadership, but his latest work looks beyond technical capability. He asks what changes when organizations bring sound judgment, ethical reasoning and human purpose into the decisions that shape digital change. We begin with Sophia, the classical Greek idea of wisdom. Alfonso argues that modern business often treats wisdom as another word for knowledge, even though the older idea also included practical intelligence and judgment. Technical expertise can tell a company whether a system can be built. Wisdom asks why it should be built, who benefits and what consequences may follow. That distinction matters when businesses feel pressure to adopt AI because competitors are doing the same. Alfonso shares the story of a large company with the resources, talent and urgency to pursue an ambitious digital program. When he asked what the business was trying to achieve, the answer became a list of fashionable technologies. AI-driven customer activity, blockchain supply chains and data optimization had become substitutes for a clear objective. His conclusion was that the company was attempting to build its future on buzzwords. Sometimes the better decision is to remain analog in a particular process if that choice serves customers and employees better. We also consider Socratic thinking in organizations where boards and investors expect certainty. For Socrates, confusion was a stage in learning rather than a failure of leadership. Alfonso believes leaders can use probing questions to expose assumptions and contradictions before a technology plan becomes expensive. Admitting uncertainty can be difficult, but false certainty can send a business confidently in the wrong direction. Heraclitus offers another useful comparison. Technology resembles a river that never stops moving, while employees need stability and meaning. Alfonso argues that leaders should create stable banks around that flow through a consistent capacity for improvement, adaptation and long-term thinking. The tools will continue to change, but organizations can reduce exhaustion when people understand the purpose behind that change and have a reliable way to respond. One of the most memorable parts of our conversation compares large language models with the Oracle of Delphi. Ancient leaders sought answers from an institution whose workings they could not fully see. Modern users can receive equally confident guidance from AI systems without knowing which data, assumptions or commercial interests influenced the response. Alfonso's point is not that machines are mystical. It is that people need discernment, source awareness and judgment when an answer arrives with authority. We then turn to Epicurus and the idea of ataraxia, or freedom from anxiety. In a business setting, Alfonso connects this with reducing unnecessary friction, decision fatigue and overload. Systems should be reliable, suited to the organization and valuable to employees as well as customers. Governance sometimes requires deliberately adding friction before investment, so teams can pressure test assumptions and ask whether people will be served by a tool or forced to serve it. Alfonso closes with two questions for any leader considering a major AI decision. What is our identity as an organization, and are we acting ethically? He uses Blockbuster as an example of a company that defined itself through videotape rental rather than entertainment. A clear identity can help a business choose technology that supports its purpose. The ethical check then asks whether transparency, consent and accountability are present, or whether data and algorithms are being used to manipulate people. Are organizations giving themselves enough time to ask why an AI system should exist before asking how quickly it can be deployed? Listen to the conversation, then share your thoughts with me.

Why is launching a consumer product easier than ever while turning it into a profitable global brand remains so difficult? In this episode of Tech Talks Daily, I speak with Omer Kaplan, co-founder and CEO of ZyG, about the operational gap between creating a product and building a durable ecommerce business around it. Omer previously helped build ironSource into an $11 billion public company before its acquisition by Unity. He explains how recognizing the move from desktop to mobile helped shape that company's growth and why the ability to adapt quickly matters even more when AI capabilities are changing every week. ZyG is building what it describes as an operating system for ecommerce scale. It combines AI agents with experienced human specialists to manage the work surrounding a consumer product, including the online store, creative production, advertising, retention, customer support, analytics, and other commercial operations. The brand retains its product, identity, and intellectual property, while ZyG operates the connected scale engine and is assessed by the resulting performance. Omer argues that existing routes solve only part of the problem. A marketplace can provide distribution, but a young brand may disappear among thousands of competitors. A commerce platform can make it easy to open a store, but the store alone does not create demand, coordinate marketing, or build customer loyalty. Agencies and software products can fill individual gaps, yet their data, incentives, and messages often remain separated. We discuss ZyG's approach to what Omer calls scale market fit. Its team creates the store, campaigns, and brand assets with agentic systems, then tests them with real paid traffic and real customer behavior. Omer says each test includes about $10,000 in media spending and that ZyG has completed over 100 tests. Cost of acquisition, predicted customer value, category benchmarks, and expected performance at higher volumes are combined into a score intended to show whether a brand can grow in the US market. He says the full process can be completed in about a week, compared with a far longer manual exercise before current AI capabilities. The conversation also examines the move from software as a product toward outcomes as a service. ZyG's consumption-based model takes a percentage of the revenue it manages. Omer is careful to distinguish accountability from assuming every commercial risk. His point is that one party should own the end-to-end result, removing the familiar cycle in which creative, advertising, and retention providers blame one another when growth stalls. Omer also shares why he returned to startup life after ironSource. Music, travel, and family offered appealing alternatives, but he saw the current technology cycle as a rare period for creating enduring companies. His advice to founders is to pursue large, complicated problems that general-purpose AI cannot easily reduce to a single feature. ZyG recently announced a $60 million Series A led by Accel, following a $58 million seed round two months earlier. Can its combination of AI agents, human expertise, real-world testing, and commercial accountability provide the missing infrastructure for the next generation of consumer brands? Listen to the conversation and share your thoughts with me.

Why do pet owners so often learn the limits of their insurance when an animal is already ill and the veterinary bill is growing? That trust problem sits at the center of my conversation with Hedda Båverud Olsson, co-founder and CEO of Lassie, a pet insurer that combines coverage with preventative health guidance, rewards, activity tracking, and AI-assisted claims. Hedda's reason for starting Lassie is personal. Her mother is a veterinarian, and Hedda grew up around healthy pets without fully appreciating how much fear and financial pressure many owners experience. After working at McKinsey and EQT, she became absorbed by an idea she describes as putting her mother in every owner's pocket. The aim was to help people understand risks earlier and make better daily choices, rather than waiting until an animal needed treatment. The claims process shows where AI can offer immediate value. Lassie has developed a system called Bark Office that scans an invoice, reads each line, identifies whether the treatment relates to illness or an accident, checks the policy, and decides whether enough information is available. Hedda says that when the system is confident, the money can reach the customer in around six minutes. She reports that approximately 65 percent of claims in Germany follow that route. Automation has limits, especially when a blurry receipt, missing diagnosis code, incomplete journal, unusually expensive treatment, or uncertain policy detail prevents a reliable decision. Those cases can prompt a request for further information or move to a human reviewer. Hedda says customers receive a line-by-line explanation of what was and was not covered, with the option to dispute a result and request another review. She reports an error rate below 2 percent for automated claims and compares it with what she describes as a 5 percent human error rate across insurance. Those are Lassie's figures, but the operating principle is useful across many regulated services: automate clear cases, explain the result, and give uncertain or sensitive cases to a person. We also consider why an insurer should have a role when nothing has gone wrong. Hedda says over 90 percent of Lassie customers use its app and roughly a quarter use it daily. Owners can watch health videos, complete quizzes, follow life-stage guidance, record activity, and earn rewards that can reduce their insurance price. Advice changes according to breed, age, and season, covering subjects such as weight, joint health, toxic foods, nail trimming, and ticks. Lassie also works with Tractive, allowing customers to connect a tracker and bring activity data into the app. Hedda explains that Lassie customers can receive a tracker while paying the Tractive subscription, and existing Tractive users can connect their current device. Owners who do not want a tracker can record activity manually. The feature gives the company another regular point of contact while helping customers follow their pet's routine. That daily relationship has commercial consequences. Hedda says regular app use supports customer loyalty, reduces churn, and raises lifetime value. Preventative actions may also support lower prices for owners. The opportunity is to make insurance useful before a claim, although firms must avoid turning care advice and rewards into confusing conditions or allowing gamification to distract from clear coverage. The conversation moves to the UK, where the supplied briefing estimates that around 20 million pets remain uninsured. Hedda believes culture and distrust may outweigh price alone, comparing the UK with Sweden, where she says approximately 90 percent of dogs and 50 to 60 percent of cats are insured despite higher prices. She also argues that established insurers have been slowed by old systems and disconnected technology, making simple onboarding, mobile service, and automated claims harder to deliver. For Lassie, the test is knowing where automation improves the experience and where it would make a difficult moment worse. Customers may welcome an administrative claim completed in minutes, but few want to speak with a bot when a pet is seriously ill or dying. That distinction between speed and empathy may be the most useful lesson for any business automating emotionally sensitive work. Can insurance become something customers value every day without losing the clarity and human care they need during a crisis? Listen to the episode and share your thoughts with me.  

How much of construction's cost, delay, and waste begins with information that fails to survive the journey from design to delivery? That question runs through my conversation with Julian Geiger, Chief AI Officer at Nemetschek Group, as we look at the practical role of AI across architecture, engineering, construction, and operations. Julian describes what he calls the industry's 90, 40, 20 problem. According to the figures he shares, 90 percent of projects are over budget or over time, the built world accounts for roughly 40 percent of global carbon emissions, and around 20 percent of material is wasted. His argument is that many poor outcomes start as information and decision problems. Each project phase may work reasonably well on its own, but handovers can strip away context. A building information model becomes a PDF, a PDF becomes an email, and a decision may never be recorded against the object it changed. We discuss how AI, building information modeling, and digital twins can identify missing information, scope gaps, clashes, and design choices before they become expensive construction site problems. Julian shares Nemetschek's work bringing Firmus AI into Bluebeam to review two dimensional drawings, then explains the broader goal of feeding lessons from construction back into design and engineering tools. The commercial promise is easy to understand. Finding a mistake while a wall exists only in software costs far less than finding it after workers and materials are waiting on site. The conversation also moves beyond the assumption that every task needs the largest available model. Julian sets out a four tier approach. Deterministic calculations such as structural math should remain deterministic. Stable, high volume checks may be handled by conventional rules. Smaller domain models can classify objects, retrieve data, and interpret geometry close to the source. Frontier models earn their place when the work involves ambiguity, reasoning across documents, or several dependent steps. His test is refreshingly practical: use the least expensive method that is reliably right and fast enough for the person waiting on the answer. That discipline matters when finance teams ask for proof. Time saved on drawing reviews or tender preparation can be measured quickly, while reductions in rework or missed issues require a longer data series. Julian also notes a familiar problem for enterprise AI programs. If a firm never established a baseline, it becomes difficult to show what improved. Usage can indicate that people find a tool useful, but adoption alone does not settle the return on investment question. Data sovereignty adds another layer. Construction files can include valuable designs, commercial information, and details tied to national infrastructure. We discuss where the data is stored, who processes it, which jurisdiction applies, and whether customer material is used for model training. Julian argues for separating genuine intellectual property from routine usage data, then matching controls to the sensitivity of each project rather than treating every data set as identical. Finally, we consider people. In an industry facing a skills shortage, removing junior roles creates a future shortage of experienced professionals. Julian sees AI as a way to shorten the apprenticeship period and reduce repetitive documentation, while preserving a clear line of accountability: AI proposes and a qualified human decides. Could that model help construction professionals spend more of their time on judgment, design, and better buildings, and where should the industry draw the line? Listen to the episode and share your thoughts with me

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.