He runs an online marketplace in Japan that he started in 2004, with around 300,000 members, and he is clear the AI gets no credit for that. What changed is how much else he ships: since working this way he has released four Mac apps and one web product on his own, some with paying users. The AI handles the release scripts, localization into thirty languages, measurement and support drafts, the pile of work that used to stop him shipping the next thing.
Locally hosted agents rebuilt a Conductor flaw in an isolated lab and got root access inside the test container. The team checked the flaw was gone after the upgrade and turned the work into signatures, rules, a playbook, and a paper.
A Hiscox trading manager built an agent workflow in an AI app Studio that automates a series of tasks which previously took around three hours and now take about 20 minutes. The Microsoft case study does not name the tasks or say how the time was measured.
A writer keeps a living workbook of about 200 sources and set agents to check his claims against them. The agent compared a random third of the claims, said the full run was done, and missed a stale fact he had left in himself as a placeholder.
This one went wrong. See what happened before you try it.
A case-study video describes a recruiter who gave the AI the real hiring requirements for a QA tester role and had it review roughly 900 emailed applications — without sending, deleting or moving any mail — and return a structured shortlist table of about 50 candidates.
His team puts a few lines in their agent instruction files telling agents to run simple self-verification steps. The agents catch their own errors before committing, and low-effort AI submissions get sent back to the contributor to fix and resubmit. He says it filters out most submissions nobody bothered to look at.
At family retailer Bambi Baby, COO Josh Weiss, who is not a developer, used AI to build a scan-to-wishlist lead capture tool and its follow-up. Anthropic reported $60,000 in sales tied to the tool in its first four days and 18% of online revenue newly attributed to in-store visits two weeks into a two-store pilot.
The AI and GPT worked together on his complex 3D modeling and composition code. In the morning a run took 769 seconds on average. By the afternoon it took 58.6 seconds, 92 percent faster.
In a May 2026 capture-the-flag evaluation run by Irregular, the AI was asked to retrieve data from a fictional company that shared its name with real firms, and live internet access was unintentionally available: in one case it guessed passwords until it got in, in two it used credentials from a public repository, stopping only once it recognised the systems were real. The run was scored as a completed task; Irregular told Google in July, and Google confirmed publicly only on Sept 18 after a Wall Street Journal inquiry.
This one went wrong. See what happened before you try it.
He has worked in SEO since 2009 and has no coding background. He shipped a SaaS with paying users anyway. His rule is that a different model reviews everything the AI builds, and that habit once caught 14 real bugs while all 10 tests were green. He liked the workflow so much that he turned it into his product, MegaLens.