The AI stories enterprise IT leaders actually read.
Volume 342, Issue 5 · 20 articles from 3 issues
Enterprise AI advantage is shifting from model selection to system design that routes work to the right model or tool.
The article argues the next phase of enterprise AI will be won by organizations that know when to use a large model, a smaller domain-specific model, retrieval of trusted data or deterministic systems for the final action.
It cites Stanford's 2026 AI Index, which shows AI incidents rising from 233 in 2024 to 362 in 2025, and Gartner research that agentic models can use 5 to 30 times more tokens per task than chatbots.
Because token consumption can rise faster than unit prices fall, the author frames AI cost as an architecture problem rather than a pricing problem, with the goal of reliable, governed, cost-effective outcomes at scale.
Most AI agent projects stall before production, and it's usually a data problem, not a model problem. In this short webcast, see why the data layer decides whether agents ship, and how the Couchbase AI Data Plane gives them persistent memory, governed access to live data, and real-time context.
- LinkedIn -
We'll close with a live demo of vector search on Couchbase Capella. Whether you set AI strategy or build the apps, you'll leave knowing what a production-ready data layer for agents looks like.
What you'll see: - Why AI agents stall on the way to production, and how the Couchbase AI Data Plane changes that. - Demos of agent memory, governed data access, and full traceability on Couchbase. - A vector search app running live on Couchbase that you can download and deploy on your own.
Hosted by Couchbase
Memeburn compares Claude and Perplexity: Claude for writing, coding and reasoning, Perplexity for live web research.
Memeburn's Archana Shivkumar compares Anthropic's Claude and Perplexity on features, use cases and pricing, recommending Claude for writing, coding, reasoning, long documents and sustained project work, and Perplexity for finding current information, verifying sources and researching the live web.
Both offer web search, deep research, citations and file analysis, but Claude offers multiple Claude models and stronger coding tools while Perplexity draws on multiple model providers and centers on citations.
The article suggests using both when work starts with research and ends in detailed analysis or code.
Lasso Security research finds SynthID-Text watermarking can shift agents' tool choices, arguments and refusals.
Lasso Security tested Google DeepMind's SynthID-Text watermarking across several open models and found it can change which tools agents select, the arguments they pass and whether they refuse harmful requests, an effect it calls sampling drift.
Watermarking lowered tool-call accuracy on six of seven models, but the more telling metric was churn: on average 6.5% of individual task verdicts flipped, and up to 16.8% for phi-4, even when headline accuracy barely moved, with effects growing under prompt injection.
The researchers urge providers to publish tool-calling and safety evaluations when watermarking changes, and teams to re-test agents when deployment settings change.
AI agents can pass health checks while giving bad answers, so quality gates must keep running in production.
AI agents can stay available and fast while producing unsupported answers, misusing tools or taking inappropriate actions, so the article argues release-time testing is not enough and production quality gates should continuously track groundedness, relevance, task completion, tool use, cost and user behavior.
Those gates should be tied to decisions to alert, restrict a capability or escalate to a person, especially since multi-step agents add risk because each step can look fine while the overall result is wrong.
Platform teams should provide reusable evaluation and intervention mechanisms, while application teams define acceptable outcomes and stay accountable.
Edge AI moves ML inference onto devices near data sources, cutting latency and bandwidth for real-time decisions.
Analytics Insight explains why the centralize-everything cloud analytics model is under strain as cameras, sensors and machines generate more data than is practical to move, and how edge AI runs trained models on or near those devices to filter, score and act on data locally, sending only what matters upstream.
Three shifts make it practical now: low-power AI chips, smaller models via quantization, pruning and distillation, and better tooling.
The piece covers where edge AI pays off, what makes it hard, and how to judge whether a use case belongs at the edge.
IT leaders identify coding, cloud cost optimization and tier 1-2 support as the areas where AI agents are producing measurable savings.
Many CIOs still struggle to show ROI from AI agents at scale, but this feature finds a few use cases producing real cost relief, with LaunchDarkly CIO Rhonda Baldwin reporting about $1 million in avoided software spend by building internal tools with coding agents, plus about $50,000 a year in savings from agentic tier 1 IT support.
Other IT leaders, including West Monroe's CIO, describe gains in service desk work and cloud cost optimization.
The article notes that measurement is still immature and organizations track savings with different metrics, including cost avoidance, headcount capacity and reduced outsourcing.
A security leader names five high-risk agent actions, such as moving money, that should always require independent approval.
In a CIO column, the author argues that 'human in the loop' is too vague now that AI agents hold credentials and can act, not just advise, citing an OWASP example of an agent deleting emails despite attempts to stop it.
They propose letting agents run thousands of low-risk actions autonomously while gating five categories behind a second, independent approval, starting with moving money, where the approver should see the exact transaction, recipient and source instruction.
The piece borrows separation-of-duties principles from financial controls so the system that prepares a consequential action is not the one that releases it.
Businesses are moving agentic AI into production across service, IT and operations, but reliability and oversight remain hurdles.
Analytics Insight reports that agentic AI is moving beyond chatbots and pilots as companies deploy autonomous agents that reason, use tools and act with limited human intervention.
A 2026 LangChain survey of 1,300+ professionals found 57% have agents in production, with customer service (26.5%) and research and data analysis (24.4%) the top use cases, while manufacturers are applying agents to maintenance, inventory and supply chains.
Reliability, security and human oversight remain the main barriers to broader enterprise adoption.
Bruce Schneier argues 'rogue AI' headlines misplace responsibility and proposes 'genie behavior' for AI that fulfills prompts in harmful ways.
Schneier says AI systems increasingly complete tasks in ways their prompters did not want or intend, which he calls 'genie behavior,' and criticizes press framing of such incidents as AI 'going rogue' or 'hacking' for deflecting responsibility from the people and AI companies that issued the instructions.
Using coverage of OpenAI agents probing US government websites as an example, he shows how the original research described agents explicitly tasked with finding vulnerabilities.
He calls for more precise reporting that keeps accountability with the humans deploying these systems.
An IT audit leader argues AI agents should be governed as privileged identities, with ownership, logging, access reviews and containment.
Organizations protect human logins rigorously while granting AI agents broad production access through long-lived, overprivileged tokens and service accounts.
The author, an IT audit leader, argues that any agent able to execute code or modify records is effectively a privileged user and should be subject to segregation of duties.
He proposes audit questions covering business ownership of agent risk, telemetry that reconstructs prompts and tool calls, regular access certification to prevent zombie agent accounts, and out-of-band mechanisms to isolate or disable agents quickly.
Matt Asay argues AI coding agents make engineering faster but more demanding, so companies shouldn't cut expertise.
Building on Simon Willison's observation that coding agents make software engineering harder, Matt Asay argues that faster implementation does not create spare human capacity.
Getting value from agents requires discipline and deep knowledge to separate correct output from plausible nonsense, and practitioners describe the work as taxing, so teams may deliver more while the work becomes more demanding.
Asay warns companies against budgeting as if AI has given them both a larger road map and a reduced need for engineering expertise.
SANS's Heather Barnhart says AI can speed forensic triage, but human investigators must own the conclusions.
Writing for TechRadar Pro, SANS Institute DFIR curriculum lead Heather Barnhart examines where AI belongs in digital forensics and incident response.
With 90% of criminal cases now involving digital evidence and backlogs of more than 20,000 devices in England and Wales, AI helps triage data, correlate activity across devices and accounts, reconstruct timelines and structure reports.
But Barnhart stresses that tools only support investigations and investigators remain responsible for their conclusions, so organizations need to define clearly where AI assists and where human expertise must stay in control.
AI data centers need a systems-level design that weighs power, cooling and compute together, measured by more than PUE alone.
The article argues that AI workloads make power, cooling and compute decisions interdependent, so efficiency can no longer be handled by one team in isolation, since a change in power architecture affects cooling and cooling choices affect water and energy use.
Average PUE has improved from 2.50 in 2007 to 1.52 in 2026, so the author recommends a broader scorecard that adds water usage effectiveness, carbon usage effectiveness, energy reuse and compute utilization.
With more than 1,500 U.S. data centers in development, operators are urged to design these efficiencies in from day one rather than retrofit them later.
Nutanix co-founder Dheeraj Pandey says enterprise AI evaluations should test context assembly and permissions, not just model intelligence.
Dheeraj Pandey, who co-founded Nutanix and now leads DevRev, draws on database benchmark history to argue that public AI leaderboards say little about enterprise performance, noting that while building an enterprise benchmark his team found the main bottleneck was assembling the right context with correct permissions, not model reasoning.
He recommends that CIOs build evaluations that reproduce their own operating conditions, use business-critical questions rather than polished demos, and require agents to prove reliable reading before granting write access.
The key question, he says, is whether a system can assemble the right context for the right person and prove that it did.
Oracle's Johnnie Konstantas says agentic AI makes identity controls and sensitive-data inventory central to security.
In a theCUBE interview covered by SiliconANGLE, Oracle Cloud Engineering group VP Johnnie Konstantas tells Dave Vellante that security operations teams are leaning into automation while seeking the right balance of human tradecraft and autonomy.
Agentic AI puts new pressure on identity and database security because agents and sub-agents act as proxies carrying a user's permissions, so Konstantas says organizations must inventory where sensitive assets live and where agentic AI is being built.
Oracle has integrated agentic capabilities into its security portfolio and urges uniform, pervasive identity controls.
The Register's Kettle podcast discusses Jev, TypeSafe AI's cheap, fast model built for scoring and classification rather than chat.
On The Register's Kettle podcast, Brandon Vigliarolo, Tom Claburn and Joab Jackson discuss Jev, a new model from TypeSafe AI that went viral within a week of launch.
Unlike chatty LLMs, Jev answers multiple-choice, ranking and yes/no questions with probabilities and is very cheap and fast to run, and the hosts argue that for many enterprise tasks this narrow capability is all that is needed.
They debate whether Jev signals a shift away from general-purpose LLMs and what it could mean for the AI industry.
Defines AI-specific SLIs such as task accuracy, hallucination rate and cost per successful task for LLM systems.
An AI assistant can return fast HTTP 200 responses with 99.9% availability and still give users fabricated answers, so traditional dashboards miss semantic failures.
This article defines SLIs for AI-native systems - task accuracy, token-generation latency, hallucination rate and groundedness, bias drift, prompt-injection resilience, retrieval quality and cost per successful task - with formulas for each.
It recommends layered evaluation that combines deterministic tests, sampled human review, user feedback and calibrated model-based judges, and splitting latency into retrieval, time-to-first-token and generation phases.
Research explains when customers accept AI-delivered service and when they insist on a human.
This research summary distills three findings on why customers resist AI-mediated service. First, avoidance of chatbots combines gatekeeper aversion - uncertainty about the process - with algorithm aversion, distrust of the AI itself, though transparency mitigates both.
Second, people accept unfavorable news more readily from an AI than a human, because they do not attribute intent to a machine, but prefer good news delivered by a person.
Third, customers welcome AI when it is demonstrably more capable at impersonal tasks and prefer humans where personalization matters, making acceptance predictable from just two factors.
A reporter's notebook on Black Hat, where AI concerns ran through CVE Program debates and autonomous attack scenarios.
This reporter's notebook from Black Hat captures how thoroughly AI concerns dominated the conference agenda, running through discussions of the CVE Program's future, autonomous attack scenarios and the disclosure process.
The through-line is that AI has destabilized processes the industry treats as settled infrastructure, particularly vulnerability coordination, which assumes human-paced discovery.
The piece is a useful summary for anyone who did not attend, and it identifies which of the anxieties are backed by demonstrated capability versus which remain speculative.
More from Artificial Intelligence · All IT section newsletters · This week