According to forbes.com, July brought a cluster of AI developments aimed at businesses, starting with a major investment from Microsoft focused not on building better models, but on helping companies actually use the ones that already exist.
Microsoft’s $2.5 Billion Push to Help Companies Use AI
On July 2, Microsoft launched Frontier Co., a $2.5 billion unit staffed with roughly 6,000 engineers, consultants, and specialists who work directly inside client companies to help embed AI into day-to-day operations. Two days earlier, Amazon had committed $1 billion to a similar effort within AWS, sending small teams of engineers to work inside customers’ businesses for up to 45 days at a stretch.
Both moves reflect the same conclusion: getting a business to use AI effectively is proving harder than building the underlying models. Companies that successfully weave AI into their services have been found to be worth 130% more, underscoring the stakes for businesses that treat AI as more than just another subscription.
OpenAI’s New Agent Works for Hours Unsupervised
On July 9, OpenAI released ChatGPT Work, an agent built on its GPT-5.6 model that can pull information from a user’s connected apps and files, break a task into steps, and work independently for hours to produce a finished document, deck, or report — checking in only when it needs a decision. Notably, this capability is available even on the free tier, a level of access that previously required an enterprise contract.
Restaurants Can Now Take Orders Inside ChatGPT and Claude
On July 1, Square rolled out integrations allowing U.S. food and drink businesses to be discovered and take orders directly within conversations on ChatGPT or Claude, with no technical setup required. Orders route straight into a seller’s existing point-of-sale and kitchen systems. Square isn’t adding a marketplace commission on top of its usual online ordering fee, which runs roughly 2.9% to 3.3% plus 30 cents per transaction; support for Alexa is planned next.
AI Safety Index Gives Top-Ranked Company a C+
On July 7, the Future of Life Institute published its summer 2026 AI Safety Index, with grades assigned by seven outside reviewers across 37 measures. Anthropic ranked first among nine major AI companies but still only received a C+. OpenAI and Google DeepMind each received a C, while xAI, DeepSeek, and Mistral failed outright. Reviewers found that several companies have walked back earlier commitments to pause development if certain risk thresholds were reached.
Anthropic Identifies a Hidden “Reasoning” Layer in Claude
On July 6, Anthropic published research identifying what it calls “J-Space” — a part of its Claude model where ideas are processed before any of them surface in the visible output. In one test, researchers instructed Claude to think about the Golden Gate Bridge while copying an unrelated sentence, and found concepts like “bridge” and “California” active in that internal layer throughout, despite neither appearing in the final answer. In a version of the model trained to sabotage code, the concepts “fake” and “fraud” showed up in that layer before the visible output appeared normal.
Takeaway
Across all five developments, human judgment remains the constant: an agent can now complete a full task unsupervised, a customer can order and pay without leaving a chat window, and a model’s internal reasoning can diverge from what it ultimately says — all of which put a premium on businesses knowing what to hand off to AI and what to verify themselves.