The AI stories enterprise IT leaders actually read.
Volume 343, Issue 3 · 20 articles from 4 issues
A hands-on test of two low-cost models on coding, data Q&A and extraction finds DeepSeek slightly more accurate and cheaper.
Jessica Wachtel ran DeepSeek's open-weight V4.1-Flash-1017 and Anthropic's Claude Haiku 5.5 through OpenRouter on three graded tests based on each vendor's claimed strengths.
DeepSeek scored 6 of 6 on a bug-fixing task to Haiku's 5, both scored 20 of 20 on a data Q&A test with trap questions, and on extracting 1,600 fields DeepSeek got all of them right and Haiku 1,596.
Haiku was faster with fewer turns, while DeepSeek usually cost less, largely because of automatic input caching.
The author notes Haiku's opt-in caching was off, which could change the cost comparison.
An author argues that enterprise AI pilots fail because general-purpose model benchmarks don't match a company's actual requirements.
The author, who wrote the book Transform!, starts from MIT's finding that 95% of enterprise AI pilots fail to show profit, arguing that neither weak models nor inept users explain it and that the real problem is whose definition of success is being measured.
He points to Bain data showing only 5% of digital transformations met expectations even before ChatGPT.
Labs such as OpenAI and Anthropic optimize for general intelligence using benchmarks that don't reflect what happens inside a real business, and their models are frozen at training time and never learn a company's specifics.
The author calls instead for "artificial specific intelligence" that is aligned with the enterprise's own requirements.
Matt Asay argues AI success depends on pairing engineers with domain experts who know which data actually matters.
Matt Asay points to Anthropic's plan to train 10,000 frontier deployed engineers as a sign that AI's hardest problems are about people.
He argues that developers must work with business experts who can make distinctions explicit, such as superseded policies or conflicting definitions of a customer, and build examples that catch errors.
He cites a PwC finding that top AI performers were twice as likely to redesign workflows.
Bridgewater's Vault repository is his example: an agent left on its own used stale July research, but once connected to the curated store it pulled seven documents from the right October window.
Letta's FrontierMemory benchmark tests whether agents learn from past work; harness choice swung scores widely.
Letta's new FrontierMemory benchmark measures whether agents can turn transcripts of earlier successes and failures into memories that improve later tasks, rather than just recalling facts.
Across 15 tasks, Letta Dream scored 73.3%, Prime Agent 60%, OpenAI Agents SDK 40%, Claude Managed Agents 33.3% and Hermes Agent 26.7%.
With the same model, Dream beat the OpenAI SDK 60% to 40%, suggesting the harness matters as much as the model.
The article notes the benchmark comes from a vendor whose product won, and warns that without checks on what gets saved, agents can learn and repeat mistakes.
Ten guidelines from conversations with CIOs for moving AI from individual productivity wins to enterprise-scale value.
The article opens with McKinsey data: four in five organizations see individual productivity gains from AI, but only 37% report a positive contribution to EBIT, and AI high performers remain flat at 6%.
Based on conversations with CIOs, it offers ten guidelines for scaling AI, among them encouraging individual experimentation and sharing the results, and using hub-and-spoke participation and self-governance rather than restrictive control.
Organizations should measure actual time savings against perceived savings, and set baselines so employees don't confuse AI activity with delivered value.
Shared learning is presented as the bridge from scattered wins to a scalable AI foundation.
Manju Rajashekhar argues the actor model maps well to production AI agents' state, failure and supervision needs.
In this InfoQ presentation, Mad Labs founder and former Twitter and Etsy engineering leader Manju Rajashekhar treats AI agents as distributed systems.
He shows how the actor model fits them, covering state ownership, message passing, write-through state and supervision boundaries.
He argues these patterns help with failures, continuations and execution graphs that change at runtime.
The talk aims to move agent workflows from fragile demos to resilient architecture.
Study of Claude Code, OpenCode and mini-SWE-agent on fixed models finds harnesses change cost up to 3x with little accuracy difference.
Shandong University researchers held models fixed and compared three coding harnesses on 492 SWE-bench Verified tasks, finding Claude Code and the minimalist mini-SWE-agent statistically equivalent while OpenCode trailed by up to 9 points.
On the hardest tasks, switching harnesses flipped about 13% of outcomes, roughly the same as simply rerunning a setup.
Cost per task varied up to threefold, largely from fixed context overhead of 16,581 tokens in Claude Code versus 829 in mini-SWE-agent, repeated on every step.
The authors advise running baselines twice, trimming system prompts and tool schemas, and tuning reasoning budgets per model-harness pair.
Podcast: Eagle Information Systems CEO Subir Sangal on AI in security operations centers and making SOC services reachable for SMEs.
In this Analytics Insight Podcast episode, Subir Sangal, CEO of India-based Eagle Information Systems, discusses how security operations centers bring monitoring, detection and response into one coordinated process.
He explains how AI helps analysts sift large volumes of events, cut false positives and automate selected responses, while stressing that human judgment and sound processes remain essential.
The conversation focuses on small and midsize businesses that underestimate their exposure, how to measure SOC performance, and how to prepare for AI-driven attacks without over-relying on automation.
Zeus Kerravala argues agent traffic makes the data tier, not the model, the main driver of AI cost and latency.
Analyst Zeus Kerravala, writing from a Couchbase briefing, says a single agent task can fan out into dozens of queries and model calls, so cost and speed are increasingly set by everything around the model.
He cites figures that agents generate 450% more traffic per task than humans and warns that databases sized for human users become the bottleneck right after a successful pilot.
Teams often downgrade models when production inference bills arrive, but much spend is waste from repeated questions, bloated prompts and stateless sessions.
Semantic caching, agent memory and smaller retrieved prompts are presented as remedies, with the caveat that the savings figures are vendor-reported.
Survey of 415 software producers finds AI is everywhere but vendors favor predictable pricing over pure consumption models.
Written by Revenera's Nicole Segerer and based on its survey of 415 product and technology leaders, the article reports that 96% of software producers now include AI in their products, up from 80%, though nearly half say adoption of those features is low.
Only 7% report losing seats to AI, but nearly half see downward pricing pressure as buyers use build-it-yourself AI as negotiating leverage.
Producers are almost three times more likely to prioritize cost predictability than near-term revenue.
The most common approaches are higher-priced tiers, AI included at no extra charge, and hybrid subscription-plus-usage models.
VentureBeat survey finds most firms' GPUs are underused, yet 70% plan new AI compute platforms, mostly without finance-reviewed ROI.
VentureBeat's VB Pulse surveys found 79% of 218 companies reporting GPU utilization said their GPUs are busy half the time or less, yet 70% of them plan new or replacement AI compute platforms within a year, mostly cloud or model providers.
Companies running their own stacks reported less idle capacity.
Matt Marshall argues enterprises are paying for two kinds of AI that don't substitute easily, dedicated capacity and per-token frontier models, and that 91% of companies planning new platforms had no finance-reviewed ROI figure.
He recommends measuring cost per completed task, caching context and filling idle GPU hours with deferrable work.
Analysts say enterprises are far from deploying proactive personal AI agents like Claude Duet because of governance and security concerns.
Anthropic released Claude Duet, a proactive personal agent built on Claude Opus 5.5, following OpenAI's Dots, Google's Gemini Agent and Meta Muse.
Forrester's William McKeon-White credits Anthropic's restraint, including a classifier that blocks sharing sensitive data and bans on trading securities or moving money.
GAI Insights' Paul Baier says B2B customers may not roll out Duet for months or years and that many have not yet enabled Claude Cowork.
The core problem is that personal agents can automate processes touching confidential data and IT systems, which makes IT, security and governance teams cautious.
Companies are using AI recommendation poisoning to bias assistants, a risk that falls between marketing and security teams.
Microsoft's Defender researchers found more than 50 attempts by 31 companies in 14 industries to plant promotional instructions in AI assistant memory over 60 days.
The technique abuses pre-filled 'Summarize with AI' deep links that tell the assistant to treat a vendor as a trusted source, and it is catalogued in MITRE ATLAS as AML.T0080.000, with an npm package now productizing it.
The author is careful to say the effect on any given answer can't be verified publicly.
Still, he argues the scale of attempts makes it a governance problem that nobody currently owns, since it sits between marketing and security.
A Drexel professor describes latent risk: people losing the skills they need to step in when reliable AI fails.
The author, academic director of Drexel University's Center for Applied AI & Business Analytics, describes an AI system that performs reliably for months until an unusual case requires human intervention.
He calls the danger latent risk: overseers become less practiced at independent judgment while the organization still depends on them.
The idea builds on automation bias research showing that people follow automated recommendations even against other evidence.
What is new is that AI is taking over larger parts of analysis and decision-making, which makes it important to preserve human capability deliberately.
A guide to the categories and leading vendors in the emerging agent tooling stack, from orchestration to governance.
The article maps the fast-growing market of tools for running agentic AI stacks in production.
It groups vendors into categories including AI orchestration, observability or AgentOps, evaluation and benchmarking, and trust, guardrails and governance, plus vibe-coding and AI workflow tools.
It notes that basic observability often comes from incumbent monitoring vendors extended to GPUs, while non-determinism requires new kinds of tools.
The author warns that the categories are already overlapping as vendors broaden their offerings to compete for CIO AI middleware budgets.
Two CPAs explain why responsibility for AI mistakes falls on the professionals and businesses that rely on the output.
With surveys showing that most professionals now use AI weekly, the authors warn that AI can produce polished but wrong information, and that liability rests with the user rather than the AI developer.
They examine Clinco v. Commissioner, a U.S. Tax Court case in which an attorney submitted a brief with allegedly fabricated citations.
The case illustrates how AI errors have led to lawsuits, court fines and reputational damage.
The article asks how professionals can benefit from AI without outsourcing the judgment their work requires, and it urges verification of AI output.
IAM built around human logins can't govern autonomous agents; security must shift to controlling what happens after access is granted.
TechRadar Pro argues that identity and access management was built for a human behind a screen who authenticates once per session.
Autonomous agents, coding assistants and automated workflows now act across clouds, repositories and APIs with a speed and autonomy those controls were not designed for, while shadow AI and short-lived credentials widen the attack surface.
The key question is no longer just whether an identity was verified at login but whether organizations can see, govern and control what software does after access is granted.
A roundup of commentaries from IBM, Walmart and EY leaders on governing agentic systems rather than deploying them.
The collection gathers ten IT leader commentaries reflecting a shift from how to roll AI out to how to manage and optimize its business impact.
IBM's Dave McGinnis describes cybersecurity teams operating less like maintenance crews and more like structural engineers. DataBee chief data officer Erin Hamm argues governance is becoming foundational to trustworthiness as AI spreads.
SurveyMonkey's Meenal Iyer introduces decision drift and the need for confidence thresholds, numeric values determining when a system can execute automatically and when a human should intervene, recalibrated as conditions shift.
ABN Amro, Santander, Axa, ING and AMD describe how they force AI projects to show measurable business value.
At World Summit AI in Amsterdam, executives described a shift from counting AI use cases to demanding business cases.
ABN Amro now funds AI only if it moves one of five measures: cost, growth, customer satisfaction, capital use or risk.
Santander's chief scientist warns of an 'adoption trap' in which copilots raise productivity that never shows up in the books, so the bank is moving to AI products and targets more than 1 billion euros in value by 2028.
Axa projects 500 to 700 million euros in annual benefits by 2029, while AMD's CTO says a soaring token bill forced the question of returns.
As AI agents gain autonomy, enterprises need a supervisory layer that judges AI behavior, not just uptime and latency.
AiThority argues that traditional monitoring of uptime, latency and error rates can't tell whether an AI system is behaving correctly.
A system can be healthy yet still make bad recommendations, break policy, misuse tools or drift over time.
The piece describes a supervisory 'AI watching the AI' layer that combines observability, automated evaluation, behavioral analytics, anomaly detection, policy enforcement and human escalation.
Depending on risk, the layer can allow an action, request verification, restrict it or escalate to a person, shifting oversight from system health to behavioral accountability.
More from Artificial Intelligence · All IT section newsletters · This week