AI tools now save time on most individual tasks, but the time often moves to whoever has to check or maintain the output. Most of today's pieces decide in advance what an AI agent may do on its own and what waits for a person. One asks who will own a tool after AI has built it.
Maria Rosala and Rachel Banawa of Nielsen Norman Group, a user experience research and training firm (24 Oct · 01), surveyed 521 designers, researchers and product managers in August and September 2026. Respondents rated AI's effect on task time at 4.2 out of 5, yet their feeling about AI at work averaged only 3.57. Of those who left comments, 69 per cent expressed some negative sentiment, most often about unreliable output and the work of checking it. Some designers described repairing unusable prototypes that product managers had made with AI. The authors argue that organisations should judge AI by whether work reaches users faster and better, not by how many people use it.
Microsoft's Inside Track blog, in a piece by David Hirning (24 Oct · 02), describes how the company's internal teams govern thousands of AI agents. Alberto Carrozzo, who leads responsible AI reviews for Europe, the Middle East and Africa, says his team has reviewed more than 2,000 internal tools. Jeevan Pathuri runs software engineering for the cloud supply chain group, which has more than 100 AI agents in production. The group tests each one to confirm it stays inside the job it was built for. The legal department's AI agent must give a written reason for each judgement on a document, so that attorneys can check it against the source. The article's point is that an AI agent with one narrow job is easier to test, to trust and to work beside.
Tom Dale of Vercel, a company that hosts and deploys websites and web applications (24 Oct · 03), explains how its AI agent for fixing broken production deployments gets permission to act. Broad access lets a mistake reach unrelated systems, and approving every single step turns review into a habit of clicking approve. Vercel's AI agent therefore investigates with read-only access first, then presents one plan that lists the work and the access it needs. The person approves the plan as a whole, and the AI agent receives short-lived credentials limited to that plan and to the person's own permissions. This is Vercel's account of its own product.
Patrick Neeman (24 Oct · 04) is a user experience designer who has spent decades observing expert users of business software. Writing in UX Collective, he argues that a chat box often slows down the people who know a product best. A claims adjuster or a contract lawyer already knows the fields, filters and keyboard shortcuts, and a prompt makes them describe that knowledge in sentences. He sorts work into three places. Routine tasks stay in the screens and rare one-off questions go to chat. Continuous monitoring goes to an AI agent that interrupts only when a person must decide. He recommends measuring how long a user takes to trust an answer enough to act on it, not how fast the answer appears.
Tommaso Nervegna (24 Oct · 05) runs a 100-person design studio inside a large company. He sketches how he would run a new studio with 15 AI agents built on Grok Bot, an AI agent product he has tested in beta. He states plainly that the studio does not exist and that the piece is a thought experiment. He would talk to one coordinating AI agent, which directs the other fourteen and sends him each morning a list of what needs his approval. Each AI agent would get its own account with minimal rights and one of three levels of action: act and log it, prepare and wait for approval, or never act. Creative direction, first client conversations, money and judgements about people would stay with humans.
Patrick Neeman (24 Oct · 06) returns in a second UX Collective piece, this one on whether to buy software or build it. Coding agents make an internal tool look like an afternoon's work, but he budgets 30 per cent of the build cost every year for upkeep. On his figures, a tool two people build in two weeks costs $45,600 over three years, against $18,000 for a subscription covering 25 people. He asks teams to follow one real task across every system it touches and log each switch and each retyped value. Because the AI models underneath are retired on someone else's schedule, he keeps 20 to 30 real tasks with known good answers and reruns them after each change. His point is that designers see these costs first and should bring the evidence before the decision is made.
The survey shows the cost when nobody decides who checks AI output: the checking lands on whoever cares most about quality. Microsoft, Vercel and Nervegna each settle that question in advance, by limiting what an AI agent may do and naming the actions that wait for a person. Neeman applies the same test to product design, placing AI only where it beats the screens an expert already uses. His second piece carries the question past launch: a tool built quickly with AI still needs a named owner and a budget for its upkeep.
Two of these accounts come from companies describing their own products, and one describes a studio that does not exist. None of them measures whether the approval steps actually catch mistakes. A single approval of a whole plan may also become as automatic as many small approvals once people trust the AI agent. Neeman's cost figures illustrate his own rule of thumb and are not a measured case.
A Nielsen Norman Group survey of 521 UX practitioners found that AI speeds up most tasks, but the time saved often goes to checking and fixing other people's AI output.
Microsoft's internal teams govern thousands of AI agents by giving each one a narrow job, reviewing it before launch and requiring explanations that people can check.
Vercel's AI agent for fixing broken deployments investigates with read-only access, then asks a person to approve one plan, avoiding the reflexive clicking that step-by-step prompts cause.
Patrick Neeman argues that chat boxes slow down expert users of business software, and that routine work belongs in screens, rare questions in chat and monitoring in AI agents.
Tommaso Nervegna, who runs a 100-person design studio, sketches a studio run with 15 AI agents in which humans keep creative direction, client meetings, money and judgements about people.
Patrick Neeman argues that coding agents make internal tools cheap to build but not to own, so designers should weigh in before a team decides to buy or build.