AI value is organizational, not technological

Strategic clarity plus applied AI yields five times more AI value

BCG surveyed more than 1,300 executives across 20-plus sectors and finds that companies strong in both strategic clarity and applied AI generate five times as much AI value as those strong in neither. Agentic AI's share of total AI value rose from 17% to 22% in a year, and companies with all six recommended AI controls generate three times as much agentic AI value as those with just one. Only 5% of companies have that full control set today.

BCG →

Why most firms get productivity from AI but not profit

McKinsey's 2026 state of AI survey finds 80% of respondents report individual productivity gains, yet only 37% see positive EBIT impact, and just 6% qualify as AI high performers. Nearly three-quarters of those high performers substantially redesigned workflows, versus a quarter of other organizations. Separate McKinsey research finds only 11% of organizations have reached a "reinvention" horizon that rethinks how work gets done.

McKinsey →

Many AI benchmarks fail basic validity tests, Stanford finds

Stanford researchers Sanmi Koyejo and Sang Truong applied measurement science and psychometrics to 56 widely used AI benchmarks. They found the popular BBQ bias test often measures reading comprehension instead, since a model can score as unbiased simply by spotting a trick question. A related study found translated safety benchmarks degrade unpredictably, mixing weaker guardrails with shakier translated questions into one unreliable score.

Stanford HAI →

AI spend is splitting across and beyond the IT budget

BCG's Applied AI 2026 survey finds AI's share of the IT budget has more than tripled, from 5% to 17%. Companies also report spending an average 2.8% of revenue on AI funded outside Tech/IT, on top of 3.3% on enterprise IT. BCG attributes much of this to AI spend moving into business functions and products as programs scale past pilots.

BCG →

CPG firms use agentic AI for speed, not growth coordination

McKinsey surveyed 41 senior CPG commercial leaders and finds only 12 have embedded AI into core workflows; the rest remain at pilot stage. Most target data gathering, pricing analysis and deck creation. McKinsey argues the larger opportunity lies in coordinating retail media, pricing, promotions, assortment and execution together, since fragmented data and slow planning cycles cause lost sales.

McKinsey & Company →

Study maps EU AI Act risk tiers across 628 German startups

The appliedAI Institute for Europe, as part of the Bavarian AI Act Accelerator, classified the highest-risk AI use case of 628 German AI startups and over 2,400 AI use cases into the EU AI Act's risk tiers. The study, combining legal expertise, LLM-assisted classification and manual validation, found compliance challenges concentrated in specific sectors, enterprise functions and AI application areas.

appliedAI Institute for Europe →

A16z: tech drove 76% of S&P 500 earnings growth in 2026

A16z's State of Markets II reports tech contributed about 76% of the S&P 500's total earnings growth in 2026, as of late August. The same report finds only around 2% of S&P 500 companies track any quantified AI metric, and about 2% of US households paid for an AI service as of April.

a16z (State of Markets II) →

Study finds nine AI chatbots share data with trackers

IMDEA Networks Institute tested ChatGPT, Claude, Gemini, Grok, DeepSeek, Perplexity, Le Chat, Meta AI and MS Copilot. Five services shared permalinks to chat histories with trackers, and a third shared AI-generated summaries revealing health, financial or location details. The researchers identified 44 tracking or advertising third parties receiving data, with Google trackers present across all services. Rejecting non-essential cookies still left 80.8% of trackers active.

Heise Online →

OpenAI's Textgrain watermark defaults on only in the EU

OpenAI will add its new Textgrain watermark to ChatGPT and Codex by default only in the EU, meeting AI Act transparency rules, in the coming weeks. API customers worldwide can enable it manually. Detection access is currently limited to selected researchers and organizations by application. Detection accuracy varies by text length and topic, and ranges from 36.5% to 94.3% in OpenAI's own examples.

heise online →

Pentagon finishes Claude removal as Meta, Microsoft usage drops

The Pentagon confirmed it has stopped using Anthropic products, after declaring the company a security risk this spring over its refusal to drop contractual bans on mass surveillance and autonomous weapons use. The Information reports Microsoft's internal Claude spending is now a third below its forecast of at least one billion dollars, and Meta's Claude Code users fell from 60,000 to 30,000 employees.

heise online →