Tech Talks Daily

What happens when an AI agent makes a business decision using information that's already out of date? It's a question I wanted to put to Stephen Deasy, CTO of Confluent, when we met at Current 2026 in San Francisco. After attending more than 20 technology conferences this year, I've heard countless promises about what AI agents will eventually achieve. But there's a growing problem behind those ambitions. Businesses are still struggling to get accurate, current information out of the systems they already own.
Confluent has introduced Streamhouse, a data architecture designed to address that problem. Stephen explains why the company believes existing approaches can leave applications working with delayed or incomplete information, and how the new architecture brings streaming, processing, governance, and querying together. We also discuss Lightning Tables, one of Confluent's major announcements at the event, and what direct access to current Kafka data could mean for developers and their businesses.
I was particularly interested in what this means for companies that have already invested heavily in data warehouses and lakehouses. Stephen makes the distinction between analyzing past performance and running business operations using current information. He explains why Confluent sees these approaches working together, and why existing technology investments can provide the foundation for introducing Streamhouse.
We also discuss the experiences of Instacart and Citizens Bank. Stephen recalls how Instacart faced a sudden increase in demand during the pandemic, when accurate information about orders, inventory, and deliveries became especially important. He also describes how Citizens Bank improved loan fulfillment processes through real-time data streaming. Both examples help explain why the speed at which information moves between business systems can affect customers directly.
The conversation also covers IBM's relationship with Confluent, open technologies including Kafka and Flink, and the announcement that Tableflow is now generally available on Google Cloud. Stephen shares his thoughts on model evaluation, human oversight, and how companies should measure returns from their technology spending.
One of my final questions concerns the misconceptions surrounding Streamhouse. Stephen explains that the architecture brings together approaches many organizations have already been working toward, and that its value lies partly in making those capabilities easier to use together.
With so much money being invested in AI, are businesses paying enough attention to the information those systems depend on? Listen to our conversation and share your thoughts.

What should a business consider before moving AI inference onto infrastructure it owns?
In this episode of Tech Talks Daily, I speak with Paul Speciale, Chief Marketing Officer at Scality, about the decisions behind enterprise AI infrastructure. Following the launch of Scality AI Inference Factory, we discuss why organizations are assessing on-premises inference alongside cloud services, with attention to costs, model control and the location of sensitive data. Paul expects a hybrid approach, with different workloads running in the environments that suit their requirements.
The conversation begins with the work a GPU repeats when an AI session needs to process familiar context. Paul explains how saving and retrieving that information can reduce recomputation, and why storage and networking deserve attention alongside compute. We discuss where the economics change as usage grows and why owning the infrastructure also means taking responsibility for utilization, performance and the operating bill.
We then consider data placement across storage tiers. Paul argues that there is no single best location for every workload, with performance, cost and sovereignty requirements shaping the decision. I ask where automation should stop and human approval should begin. He describes automating health checks, forecasts and recommendations while retaining approval for consequential changes such as deleting data or changing policies.
Cyber resilience is another part of the discussion. Paul explains why an immutable copy needs to be supported by tested recovery, and why sovereignty rules should be checked through attempts to perform actions the system is meant to refuse. We close with the responsibilities infrastructure teams take on as they manage inference directly, from tracking time to first token and context reuse to deciding when a model should change. For leaders assessing their next AI deployment, the conversation offers practical questions about control and the evidence behind the business case. Which workloads would you consider running locally, and what would you want to prove before making that decision? Please share your thoughts with me.

How do you introduce electric vehicles and AI-assisted safety technology into a collection fleet that customers depend on every day?
On Tech Talks Daily, I'm joined by Brett Rogers, Vice President of Operations Technology at Republic Services, to discuss the engineering, infrastructure, and human decisions behind modern waste and recycling operations. Brett's background includes work at Toyota, General Motors, and Cruise. He describes arriving from the automotive sector and discovering how much complexity sits behind a service most people expect to happen without thinking about it.
We give particular attention to electric refuse trucks. Range is part of the calculation, but the vehicle also powers equipment that lifts containers and handles material. Brett explains how Republic Services models energy use across the chassis and collection body, taking account of terrain, climate, speed, and the work required on a route. The goal is to identify where an electric vehicle can replace a conventional truck while maintaining service.
Route selection can produce some surprising answers. A residential truck making frequent stops may benefit from regenerative braking, while a vehicle with fewer container lifts can spend much of its day covering distance. Brett describes why his team began with residential applications, then used operational data to inform expansion into other vehicle types. The number of lifts alone does not tell the whole story.
Charging has its own planning cycle. Brett says utility discussions, design, and construction can require decisions roughly two years before trucks are needed. We discuss why depot readiness has to lead vehicle delivery, alongside the arrangements that keep service running if power is interrupted. He describes spare vehicles and equipment that can accept emergency power, while on-site storage and renewable energy form part of the options being considered. These approaches have different levels of maturity, and the conversation keeps that distinction clear.
Technology-enabled safety is the other major focus. Brett describes adding a bird's-eye camera view to help drivers see around large vehicles, then working with vendors on systems that distinguish people from objects. Spoken alerts such as "pedestrian rear" are intended to tell a driver what is happening and where, without requiring them to keep checking another screen.
The discussion goes beyond installing cameras. Brett says in-cab alerts were associated with an initial reduction of around 20% to 30% in what Republic Services calls unfocused behaviors. He reports a reduction of 95% after coaching was added to the approach. These are company-reported behavioral measures, not a claim that collisions fell by 95%. No evaluation period or underlying methodology is provided in the interview. The useful operational lesson is the connection between technology, explanation, and the opportunity for drivers to correct a recurring behavior.
We also discuss employee adoption. Brett recalls drivers who were initially reluctant to move from diesel trucks and describes how time in the electric vehicles changed their view. He points to lower noise, vibration, and heat, alongside cab design. His account is a reminder that the experience of the person doing the job can determine whether a promising deployment becomes a useful everyday tool.
We finish with routing visibility for municipal customers and plans to use AI and other tools to improve operations. For fleet leaders, the conversation offers practical considerations around duty cycles, charging resilience, driver support, and measuring results. For everyone else, it reveals the work behind a truck arriving on the right street at the right time.
I'd welcome your thoughts. What would give you the most confidence in an electric fleet rollout, and how should organizations demonstrate that their safety technology is helping drivers and communities?

What would make you comfortable giving an AI agent the authority to act inside a live business process?
In this episode of Tech Talks Daily, I speak with Valerie Henderson, CEO of Caylent, about what has to happen between a successful demonstration and a system a business can depend on. The conversation centers on production readiness, including access controls, spending limits, traceability, workflow design, and the people responsible for the outcome.
Valerie describes why introducing agents requires changes to how work gets done. If an organization adds new technology without reconsidering the handoffs and decisions around it, the hoped-for productivity gains can remain difficult to achieve. Her advice is to begin with a bounded problem where automating a few steps could provide a clear benefit, then build on the evidence rather than starting with an ambitious network of agents.
We discuss the difference between read-only workloads and agents that can alter live systems. Valerie says many customers are already comfortable with agents retrieving information, while autonomous operations such as self-healing infrastructure are concentrated among the most advanced customers she sees. Those observations reflect her experience with Caylent engagements. They do not establish a universal level of enterprise adoption.
For me, one of the most useful questions is what meaningful human oversight looks like. Valerie is clear that people remain accountable for the instructions and access they give agents. She emphasizes deliberate testing, appropriate governance, and the ability to understand what a system has done. We also discuss how agents can repeat work without producing anything useful, creating a bill that needs to be considered alongside accuracy and reliability.
The conversation moves into Caylent's own operating model. Valerie describes moving away from assumptions built around human time and materials, making existing technical knowledge easier to reuse, and changing delivery incentives to reward speed and customer satisfaction. It is a practical reminder that organizations adopting AI may need to reconsider what they reward as well as which tools they buy.
She also shares a customer example involving Smarsh. According to Valerie, redesigning its software delivery process with agentic AI freed engineering capacity that could be reinvested in the business. She describes that capacity in financial terms, but the interview does not establish a measurement period or separate capacity value from realized cash savings. The useful lesson is the connection between a defined objective and a decision about how to use the resulting capacity.
We touch on Caylent's agentic AI readiness research, while keeping the broader discussion focused on practical experience. The useful question for leaders is what controls and evidence should come before greater autonomy, and how they will reassess that decision as the system changes.
Finally, Valerie explains how she thinks about success for employees as well as customers. She wants technologists to have time for difficult problems and opportunities to develop their skills. She also argues that customers should be able to explain the value of the work they have undertaken. Model performance is part of that assessment, alongside the environment, permissions, costs, and business process in which the agent operates.
I would love to hear how your team is making these decisions, so please share your thoughts. Which responsibilities are you ready to delegate to an AI agent, and what evidence would you need before going further?

Useful Links
Caylent website: https://caylent.com/

Caylent's Agentic AI Readiness survey report: https://caylent.com/blog/agentic-engineering-survey-report

What makes an AI fall alert useful to the person who needs help and the team responsible for responding?
In this episode of Tech Talks Daily, I speak with Mike Link, chief operating officer at KamiVision, about deploying computer vision in residential care. We discuss the connection between the sensor, the people receiving its notifications, and the decisions that follow. The conversation also looks at privacy, connectivity, false alarms, and why a system that works in a care facility needs further work before it can serve a wide range of homes.
Mike describes KamiVision's path from home-security cameras to its KamiCare platform. His account of fall detection involves three stages. A device identifies a possible event, a cloud model filters notifications, and a person reviews a blurred, encrypted clip before an alert reaches the user. The human review is part of the service he describes, so it belongs in any discussion of how the system performs.
I ask about the company's 99.5% accuracy claim and what buyers should understand before relying on a number like that. Mike explains the filtering process, but the interview does not establish a test population, a missed-event rate, or a complete statistical definition. His answer is a starting point for further questions about performance across the whole service, rather than proof that a device detects 99.5% of all falls by itself.
One of the most useful examples concerns a resident found on the floor. Mike describes how reviewing an event can help staff understand whether someone fell or deliberately lowered themselves because they needed help with a task. That information may lead to a change in a routine or room layout. It is an example of how visual context can inform care, rather than an independently measured claim that the technology prevents injuries.
We also discuss the difference between a managed care setting and a home. A residential facility can have relatively consistent room layouts and lighting. Homes introduce furniture, pets, several rooms, and areas outside a sensor's view. Mike talks about investigating other sensing technologies to address those gaps. He presents those ideas as work under consideration, not a set of generally available capabilities.
Connectivity is another practical issue. Mike says the version discussed uses Wi-Fi and that a cellular option has been considered. He also explains that the consumer version was being tested with families and was not yet available to purchase at the time of recording. That distinction between an established care-facility offering and a product still being developed matters when assessing where the technology can be used today.
Privacy and consent are part of deployment. Mike says facilities obtain consent from residents or an appropriate representative, and he describes the service as focused on event clips rather than live viewing into a room. Those explanations help frame the conversation a buyer should have about access, retention, review, and the information people are agreeing to share.
Finally, we discuss the business case. Mike connects the value of event information with staff response, communication with families, and adjustments to care. He also describes possible reporting and risk-identification capabilities that had not been deployed. The interview offers a view of the operational questions involved, without establishing a controlled clinical result or a fully quantified return on investment.
I would love to hear your thoughts, so please share your experiences. How should care providers assess whether an AI alert improves the work people do, while respecting the privacy and choices of those receiving care?