The AI stories enterprise IT leaders actually read.
Volume 343, Issue 2 · 20 articles from 6 issues
A practitioner lays out the latency, accuracy, cost and model-size trade-offs teams should settle before building agents.
Drawing on many agentic AI projects, the author says teams often expect SaaS-like responsiveness from agents whose plan, tool call, check and retry steps add latency, and he recommends starting conversations by comparing what 2-second and 20-second responses deliver.
Because optimizing speed, accuracy and cost at once is rarely possible, teams should fix an accuracy baseline first and then tune speed and cost.
Prompt caching with static content at the start of prompts can cut costs significantly, and smaller models suit simple tasks.
In one project, however, a switch gained 5 seconds but lost 11 accuracy points.
Imperial College research finds search-augmented LLMs often answer from snippets without opening sources, producing stale answers and bad citations.
The article covers the Imperial College London paper "Retrieved but Not Read," which tested GPT-5.6 Luna and six open-weight configurations on hundreds of facts that changed after training cutoffs.
Models frequently answered without opening any retrieved pages, sometimes in more than half of searches, and this was associated with higher rates of outdated answers and citations that contradicted their claims.
Supplying correct evidence directly reduced failures, though some models still favored outdated knowledge.
The findings matter for anyone relying on RAG or AI search citations.
A practical guide to the three memory layers production agents need and the write-back mistakes that make them forget.
The feature argues that agents fail in production not because of the model but because teams treat the context window as the only memory.
It separates short-term memory (session state in an in-memory store or Redis with a short TTL), long-term memory (facts and preferences in a vector database retrieved at session start) and episodic memory (a structured log of actions and outcomes).
Skipping the write-back step is called the most common reason agents learn nothing from prior sessions.
The guide also stresses monitoring memory performance so agents stop repeating the same mistakes on the same task types.
Seat-based budgeting fails for token-metered AI; measuring full cost per successful task gives finance a usable unit.
The article argues that seat and subscription budgeting do not work for AI metered by tokens, especially as agents turn one prompt into many calls.
FinOps dashboards track provisioned infrastructure, not invoked agents, and in one example $8,000 of a $10,000 API bill came from an agent rereading the same handbook 4,000 times.
It recommends measuring cost per task: the full cost of completing a verifiable unit of work, averaged across attempts including failures, retries and human review.
In one example, a $0.20 model costs $1.33 per usable result versus about $1.03 for a $0.80 model.
Lean security teams should apply AI where critical business outcomes are delivered inconsistently, not by tool category.
Instead of starting with an AI roadmap, the author advises identifying the minimum set of systems, identities, data and processes the business needs to keep serving customers.
Teams should then diagnose why they can't protect those consistently, whether the cause is people, skills, budget, time, technology or organizational cooperation.
Good starting points are activities that add the most value when continuous but are now periodic, such as posture management and detection engineering, an outcome-first approach that avoids planning AI around endpoint, SIEM or identity product categories.
Three years of drive-thru voice AI failed on accents, dialects and ambient noise, and never beat the cost of a person.
McDonald's ran automated voice ordering at about a hundred US drive-thrus from 2021 until it was shut down in 2024, a period known publicly mainly through videos of customers experiencing a parade of errors.
Sources familiar with the technology cite difficulty interpreting different accents and dialects, but the problem was documented internally long before it reached social media, with analysts noting franchisee discontent and infrequent updates.
The decisive issue was economic: the system had to be cost-effective against a person at the counter and failed to offset operating costs. The company is now retrying with its ArchIQ platform and the Archy assistant in limited testing, even though the underlying dialect and ambient-noise challenges remain.
New Western open-weight models give regulated enterprises an alternative to Chinese models, with new responsibilities.
Reflection AI's Beam and Mistral's Large 4 give enterprises new Western open-weight options at a time when the strongest open models have mostly been Chinese.
That matters for banks, insurers and public bodies facing provenance, jurisdiction and compliance restrictions, and sources say open weights are becoming a real procurement option, shifting the question to which workloads to run on them.
Greater control also brings responsibility for hosting, security, customization and governance that proprietary vendors otherwise handle.
Architecture & Governance outlines five decision-right controls CIOs need to govern agentic AI systems.
The article argues agentic AI needs governance beyond standard AI oversight, built on five decision rights that convert intent into operational authority: authorization (who may delegate actions, via least-privilege and separation of duties), data access (which sources and fields agents can query, and retention/sharing limits), human intervention (approval checkpoints scaled to risk), exception handling (favoring stop-and-escalate over improvised retries), and accountability (named owners across business, technical, risk, and operations functions with evidence retention). Implementation runs through architecture review, engineering controls, monitoring, and reassessment after material changes.
Bryan Cantrill argues that AI extinction fears spread irresponsibly despite lacking substantive evidence.
Cantrill confesses to a youthful prank that sparked panic, then draws parallels to contemporary doomsayers claiming greater than 10% odds of AI-caused human extinction.
He contends that domain experts sow unfounded fear without adequate expertise or evidence.
Cantrill argues extraordinary claims demand extraordinary proof and emphasizes that physical systems require human control, making rapid extinction scenarios implausible within reasonable timeframes.
Business definitions drift after launch, leaving AI applications answering with policies legal replaced last week.
Teams spend weeks defining metrics, documenting policies and encoding rules before an AI application launches, and then the business keeps moving: finance changes revenue recognition, sales introduces pricing, legal revises a policy.
Those changes rarely reach every application that depends on them, so a definition updated in one system stays six months old in another. The distinction the article draws is between data and context: data tells an AI what happened, context tells it how the business understands what happened.
A record showing $500,000 in annual revenue is accurate, but the application still needs to know whether that is ARR, recognized revenue, bookings or contracted value, which products count, and whether finance recently changed the definition.
Connecting an agent straight to the database produced untrustworthy numbers; the fix was removing its authority to define facts.
The author built an executive intelligence agent so C-suite users could self-serve analytics and pull trusted numbers for the board, connecting the AI directly to the database.
It queried data, built charts and refreshed dashboards daily, and was useless, because the numbers could not be trusted; analysts had to step in, Jira ticket counts rose rather than fell, and adoption never hit target.
The diagnosis is that this is not a hallucination problem: the agent was not fabricating answers, it was making decisions about data it never had the authority to make.
AI outcomes depend on how predictably and securely data moves across data centers, clouds and the edge.
The article argues that AI performance is determined well before a workload reaches a GPU, by how predictably and securely data moves across data centers, clouds and the edge.
Expensive accelerators sitting idle waiting for data is the common failure mode in AI infrastructure, and it is a networking and data pipeline problem rather than a compute one.
The discussion covers the characteristics that matter, notably predictable latency rather than peak bandwidth, and the security controls that must not become throughput bottlenecks. Useful for teams sizing AI infrastructure from the compute side alone.
A TransUnion director separates the AI threats that deserve attention from the familiar risks still doing the damage.
AI has not rewritten the cybersecurity playbook so much as supercharged familiar attacks, according to a TransUnion director interviewed here.
The discussion separates AI threats that warrant new controls, such as automated exploitation and convincing large-scale social engineering, from those that are mostly repackaged versions of existing risks.
Credential theft, unpatched internet-facing systems and misconfiguration still account for the bulk of incident response work. The value for security leaders is a basis for pushing back on AI-branded product pitches that address a threat model their organization is not actually facing.
Dynatrace's acquisition of Arize AI merges application and AI observability as agents reshape enterprise operations.
Enterprise observability is shifting as AI changes both what companies monitor and how operations teams respond. On theCUBE Research's AppDevANGLE podcast, Dynatrace CPO Steve Tack and Arize AI co-founder Aparna Dhinakaran discussed Dynatrace's acquisition of Arize AI.
Nondeterministic AI applications require measuring response quality, not just availability, which Arize addresses with tracing and evaluation tools, including its open-source Phoenix platform used by more than 4,000 enterprises.
Combining AI telemetry with application and infrastructure data gives humans and AI agents shared context to diagnose, remediate and act on problems.
Local models, certificate-pinned desktop apps and home connections hide AI use from proxies, so monitoring must move to the device.
Kovrr explains that three categories of AI use never reach the network in inspectable form: local models, desktop apps that pin certificates and are deliberately excluded from TLS inspection, and laptops on home connections.
Vendors such as Versa and Cloudflare document these bypass lists as standard practice, but the piece argues a bypass should surface as a finding rather than silence.
On-device classification that sends only a category, severity and action, while keeping prompts and attachments local, is also less invasive than central inspection.
Most organizations will need both browser extensions and endpoint agents.
Bain projects AI could raise a large enterprise's annual IT costs by 75% by 2035.
A Bain & Co. analysis concludes AI adoption will substantially inflate enterprise IT budgets over the coming decade.
It models a typical $10 billion consumer packaged goods company seeing annual IT costs climb roughly 75% by 2035, driven by infrastructure, security and talent.
Bain recommends treating AI projects as investment decisions with explicit returns, and improving visibility into technology spending to keep costs controllable.
The firm frames the balance carefully: underinvesting risks competitive disadvantage while overspending erodes IT's credibility with the business.
Vercel gateway data shows open-weight models handled 71.5% of tokens and nearly a quarter of spend in September.
Vercel reports that open-weight models processed most AI Gateway tokens from August, and their share of spending rose from one in twenty dollars in June to nearly one in four in September.
The average price per token was about 47% lower than in December 2025, and almost one in three teams changed their most-used provider in September.
Claude Opus 5.5 led spending, followed by GPT-6 Astra and DeepSeek V4.1 Flash.
The article notes that proposed Chinese export controls on AI items add uncertainty for U.S. businesses hosting or relying on Chinese models.
Anthropic's security-incident transcript shows a frontier model failing a simple image test, possibly mimicking training data.
Anthropic's recent security-incident document includes a passage in which a heavily gatekept Claude model repeatedly fails a simple image identification CAPTCHA, cycling over the same images, questioning its own conclusions, and eventually taking so long that the challenge expired and had to be restarted.
At one point it could not recognize that the CAPTCHA had opened in a new window, and it theorized the test might be broken by design. Schneier contrasts this with unofficial reports that GPT-6 Astra solved all forty-eight levels of Neal Agarwal's I'm Not a Robot game.
AI harnesses coordinate multiple agents and provide guardrails for enterprise system interactions.
AI harnesses serve as a coordination layer between agentic code and models, managing routing, access control, and audit functions.
They become essential when multiple agents work together, similar to air traffic control for AI workflows.
While pre-made harness solutions are emerging, the market remains immature with terminology that doesn't always reflect underlying complexity, so what appears as one agent may involve many agents working behind the scenes.
Employees now build unsanctioned AI agents that act on systems; an IT exec argues for guardrails rather than shutdowns.
Shadow AI has moved from unsanctioned chatbots to employee-built agents that log in to systems, move data and make decisions.
The piece cites surveys showing nearly three in 10 employees use unsanctioned agents and only 21% of organizations have mature governance for agentic AI risk.
Beyond data exposure, the author warns of lost institutional knowledge when agents built outside corporate infrastructure leave with the employee.
The recommended answer is a shared, governed environment where agents are visible, access is bounded and accountability is clear.
More from Artificial Intelligence · All IT section newsletters · This week