The operations stories enterprise IT leaders actually read.
Volume 342, Issue 2 · 14 articles from 4 issues
5 Steps for Implementing Predictive Maintenance
Implementing predictive maintenance requires careful planning across organizational teams including IT, operations, and finance.
The article outlines five key steps: assessing organizational readiness, building a strong data foundation, selecting appropriate technology, launching a targeted pilot program, and measuring results before scaling. Leaders must prioritize data quality and integration with existing EAM and ERP systems to avoid false alerts.
A well-executed pilot lasting three to six months helps validate the technology and business case before enterprise-wide deployment, ultimately reducing downtime and maintenance costs when properly executed.
CVSS scores ignore exploitation likelihood and environment, so context-driven prioritization is needed to cut remediation noise.
With roughly 49,000 vulnerabilities disclosed in 2025 and only a small fraction ever exploited in the wild, the article argues that severity-based triage misallocates scarce remediation effort.
CVSS assigns scores from technical characteristics and does not reflect how a vulnerability actually exists in a given environment.
The recommended alternative weighs asset criticality, external exposure, identity permissions, threat intelligence and AI-agent access paths. Because cloud infrastructure and AI automation change constantly, static scores go stale quickly and continuous reassessment needs automation.
One cited case reduced the security noise requiring active attention by 70% after switching to context-driven analysis.
How Businesses Can Future-Proof IT Infrastructure With AI
CIOs are leveraging AI to build resilient, adaptable IT infrastructures that can evolve with changing business needs.
The article identifies six key applications: AI-powered monitoring for anomaly detection and automated remediation, machine learning forecasts for capacity planning, continuous performance modeling to prevent bottlenecks, enhanced security operations through threat detection, AI-generated code and infrastructure-as-code for faster modernization, and improved overall resiliency.
However, successful AI adoption requires significant investment in skilled professionals, appropriate governance, and demonstrable ROI to secure organizational funding, particularly in IT departments where benefits may be less visible than in sales or marketing.
A step-by-step war story on moving a live auth service out of the default namespace with zero downtime.
This practitioner walkthrough covers moving a critical authentication service out of Kubernetes' default namespace into a dedicated one without dropping traffic.
The technique hinges on an ExternalName Service acting as a temporary DNS forwarder, so dependent services keep resolving the old name while the workload relocates.
A deliberate policy exception allows ingress rules to briefly overlap during cutover. The author stresses validating the sequence through dev and staging first, and scaling the old deployment to zero rather than deleting it so rollback stays fast if something breaks.
Why Data Centers Rarely Reuse Cooling Water
Data centers consume massive quantities of water for heat rejection, with a typical 100 MW facility using approximately 2.5 billion liters annually.
Most facilities rely on evaporative cooling systems that release water as vapor into the atmosphere, making recapture impractical. Roughly 20% to 30% of water typically remains in the media and system as residual water that accumulates dissolved minerals, requiring treatment before reuse.
While closed-loop systems, blowdown treatment, and graywater reuse technologies exist, their higher capital costs and operational complexity make them economically unviable for most operators currently.
Leading with the ESM acronym kills adoption; lead with a department's existing pain instead.
Matt Beran argues IT teams undermine enterprise service management by pitching it as a framework and an acronym to other departments.
The better approach is to help a department solve a problem it already has, examining its actual workflow and letting service management practices emerge as the answer. Genuine adoption requires the department to own the outcome rather than IT driving the initiative from outside.
He recommends starting with a single contained, already-painful problem, on the basis that once a department sees the value for itself the vocabulary stops mattering.
AI-speed vulnerability discovery has broken OT patch cycles; triage, containment and surge planning matter more than patching faster.
Models can now surface exploitable flaws in hours rather than the roughly 60 days human researchers needed, but remediation in operational technology still runs at plant speed.
Patches can reboot controllers, changes near safety systems carry their own hazards, OEM validation takes weeks to months, and much legacy equipment has no update path at all.
The author's prescription is three shifts: triage on exploitation evidence, real asset exposure and operational consequence rather than CVSS alone; contain with segmentation, allow-listing and virtual patching where patching is not feasible; and pre-negotiate emergency maintenance windows while rehearsing multi-advisory scenarios before a crisis.
Microsoft shipped 751 fixes, 108 rated critical, including a WinSock driver flaw already exploited.
Microsoft's August 2026 Patch Tuesday delivered 751 security fixes, 108 of them rated critical. One flaw in the Windows WinSock driver is already being exploited in the wild.
Networking services take the brunt of the risk, with 14 CVEs in Windows DHCP Server and six in DNS Server, four of which are critical remote code execution issues. The guidance for IT teams is to patch DHCP and DNS servers first given their exposure, and to concentrate regression testing on printing, fonts, and Remote Desktop.
TechTarget outlines a proactive four-step approach to planning on-premises AI infrastructure upgrades for performance and governance.
TechTarget's Stephen J. Bigelow explains that AI workloads place heavy demands on compute, networking, storage, power and cooling, so reactive upgrades no longer suffice.
Citing McKinsey's projection that inference could reach 40% of data center demand by 2030, he urges regular, budgeted modernization cycles.
The steps include understanding workload demands, setting clear business goals and metrics, and auditing infrastructure, data quality and storage. Upgrades also address scalability, energy-efficient cooling, and security and compliance needs for AI governance.
TechnologyAdvice explains what revenue operations platforms do, their key features and use cases, and how to evaluate one.
A revenue operations platform coordinates data, workflows, routing, reporting, planning and revenue intelligence across the sales, marketing and customer success systems a company already uses.
Unlike a CRM, which is a system of record, a RevOps platform sits between or on top of multiple tools, applying rules, enriching records and triggering actions such as lead routing or deal risk flags.
The category is broad, with products specializing in data operations, routing, forecasting, planning or intelligence. The guide advises buyers to define the problem first and scrutinize integration depth and sync behavior.
AI-generated code that works but breaks architectural boundaries creates Comprehension Debt; executable architecture tests in CI can stop it.
The New Stack warns that AI coding agents can produce functionally correct code that quietly violates domain boundaries, creating "Comprehension Debt" as teams lose their mental model of the system.
Documentation will not constrain agents, and human reviewers cannot catch drift across thousands of PRs. The author advocates Executable Architecture: encoding boundary rules as tests with Python tools like pytest-archon or import-linter that fail CI builds, then feeding failures back to the agent.
Complexity gates like Ruff, Radon or SonarQube and reviewing interfaces rather than every line further help.
Three liquid cooling approaches address AI rack densities that have moved well past air cooling limits.
As AI workloads push rack densities from 5-10 kW to 30-60 kW and beyond, data centers are turning to liquid cooling for its superior heat transfer.
RDHx retrofits existing racks by replacing rear doors with liquid-cooled coils, offering simplicity at moderate efficiency. Direct-to-chip mounts cold plates on CPUs and GPUs, capturing 70-90% of rack heat at the source.
Immersion submerges entire servers in dielectric fluid, enabling extreme density but requiring specialized hardware and operational change. Each approach trades off differently on implementation cost, efficiency, density, and complexity.
KEDA's queue-based autoscaling fixes a blind spot in Kubernetes HPA for bursty AI and agentic workloads.
Author Kishor Patil explains that standard Horizontal Pod Autoscaling watches CPU and memory, metrics that miss the real problem for AI inference: 'a queue can build while existing pods still show low CPU or memory utilization,' so HPA reacts only after latency and backlog worsen.
KEDA (Kubernetes Event-Driven Autoscaling) instead scales workers based on external queue depth from sources like Pub/Sub, Redis, RabbitMQ or SQS, scaling down to zero when idle.
Recommendations include tuning scaling thresholds carefully to avoid thrashing, keeping one warm replica to absorb bursts and avoid 30-second cold starts, and monitoring queue depth alongside pod count, with agentic AI workloads showing even more extreme burstiness than standard model serving.
Operators are deploying AI for cooling, power and fault management, though full autonomy remains distant.
Digital Realty's AI/ML-driven OPDaaS platform helped the firm grow its portfolio 34% while limiting the rise in water use to 3% and saving 17,800 MWh in 2025.
AWS uses AI agents to correlate telemetry and find network root causes in seconds, DPR Construction applies AI to pre-construction sequencing and staffing, and Schneider Electric's DCIM tools cut customer quote processing from 62 steps to 14.
Analyst Roy Illsley of Omdia says AI is still used 'in discrete packages,' with hyperscalers leading and smaller operators lagging, and predicts IT/OT convergence ahead of full autonomy.
More from Operations · All IT section newsletters · This week