Designing Your AI Practice: Roles, Workflows, and Organizations

In this Design Meets AI issue, I want to touch on a few recent reports on AI’s evolving role in knowledge work. Three caught my eye this month, and read side by side, they keep circling around one question: which parts of thinking and agency do we keep for ourselves, and how much do we take charge of shaping AI’s development?

1. Thinkslop

AI in the Wild has been studying how people use AI since 2024. This year, Marc Zao-Sanders and Sara Biuk coined the term “thinkslop” to describe the lazy, unfocused thinking that can occur when we rely too much on AI. Their data shows some common habits: sending prompts before we know what we want, turning to AI before trying to solve the problem ourselves, copying outputs with little editing, and sometimes mistaking the model’s friendly tone for real depth. On the other hand, using AI as a partner that pushes back can actually make our thinking sharper. The visuals in their report make these patterns easy to see.

AI in the Wild

2. Workers are ready; their orgs aren’t

Microsoft’s 2026 Work Trend Index looks at these behaviors from another perspective. They studied trillions of anonymized Microsoft 365 signals and surveyed 20,000 AI users in 10 countries. One number caught my attention: 49% of Copilot conversations now help with cognitive work like analyzing, evaluating, solving problems, and thinking things through, not just summarizing emails or checking grammar. Also, 66% of users say AI gives them more time for valuable work.

This leads to a paradox mentioned in the HBR article. Thinkslop happens when we let AI do the thinking we should do ourselves. Microsoft’s most advanced users, called Frontier Professionals, stand out for avoiding this trap. 86% use AI output as a starting point, and 43% make sure to do some work without AI to keep their skills sharp. The same tools can lead to very different results. Interestingly, these frontier professionals are only 16% of knowledge workers. The main difference seems to be whether people continue to use their own judgment.

This gap shows up at the organizational level too. Only about a quarter of users say their leadership is clearly aligned on AI. Organizational factors like culture, manager support, and how people are evaluated matter about twice as much as individual mindset in deciding whether AI really makes a difference. People seem ready, but their organizations are still catching up. I like how clearly the report shows this.

Microsoft 2026 Work Trends Report

3. Four modes of working with AI

They also describe four ways of working with AI: asking, exploring, collaborating, and delegating. These modes are mapped along two axes: how much you are directing versus supervising, and how much work the agent is doing. Asking and exploration are lighter, giving you quick responses. Collaboration and delegation mean handing off bigger parts of the workflow to the agent. I noticed this framework is very similar to the one we use for our own AI workflows.

The key point in that section is about discernment: the best users know which mode fits each task and switch on purpose. This is the same kind of “stay in your lane” discipline described in the HBR article, shown as a 2x2. The term “Frontier Professional” also stood out to me. I think we will keep coming up with new names for these new AI-focused roles, just as “forward-deployed engineer” has become common.

4. Three levels of redesign

The report recommends three layers of redesign of the AI integration: redesigning roles, leaders redesigning how work is split between people and agents, and organizations becoming what Microsoft calls Learning Systems. I like this system design view and the idea of pulling different levers of the system to shift organizations toward healthy AI adoption. I wish there were more details on how to put it into practice, but it still offers some useful ideas to consider.

What stands out to me is how much this reminds me of earlier ideas in organizational design. The Learning System concept is a lot like Argyris and Schön’s double-loop learning from the late 1970s. That is when an organization not only fixes its work but also updates how it decides what work should be done. The finding that organizational factors matter about twice as much as individual ones would not have surprised Jay Galbraith. You cannot just add AI to people and ignore the organization around them.

5. When AI starts building AI

Anthropic’s Institute recently published a piece on recursive self-improvement, which means AI helping to build the next version of AI. This takes the same main question and examines it on a larger societal scale. Right now, Claude already does most of the building: over 80% of the code Anthropic merges is written by Claude, and their engineers now ship about eight times as much code per quarter as they did in 2024. What is still human is the judgment about what to build, which experiments to try, which results to trust, and when to see a dead end. The gap between doing and deciding is key, which is why they say we have not yet reached true self-improvement.

When it comes to an AI building itself, they describe three possible futures. In the first, progress slows down and today’s capabilities just become more common, without AI gaining a real sense of research. In the second, which they think is most likely, model development becomes very automated, but humans still set the direction and judge the results. In the third, AI starts designing its own successors, and humans move into more of a supervision role. That last scenario is where things get truly unpredictable.

One point to highlight, since it is easy to get wrong: Anthropic is not asking everyone to slow down. Their argument is more complex. A slowdown only works if all major labs and countries agree and can prove it; if only one group pauses, it just changes who is in front. They are suggesting we build the option to pause, along with the systems needed to verify a real pause. Whether this kind of coordination is possible, given all the players involved, remains an open question.

When AI Builds Itself

Looking at these three reports, a common thread emerges: whether AI is providing emotional support, performing cognitive heavy lifting, or enabling recursive development, we must retain our agency to decide what truly matters. We cannot simply react to these shifts. Instead, we must take a proactive stance by pulling the system levers identified in the Microsoft report—consciously redesigning our roles, workflows, and organizations. If we fail to shape this transition with purpose, we risk repeating the unintended consequences of the social media era, from skill degradation to a diminished sense of agency. I suggest reading these with a designerly lens and considering how we can leverage their insights as beginning points for intentional design.

And that’s a wrap for this issue. Until next time, take good care of yourself and your loved ones.