Guest essay: Is AI stretching your people?

Reconceiving the work so people keep getting smarter alongside AI

Shelley Evenson | seeSpace

Quick word from Mark and David first.

Don't forget to sign up to our webinar where we will discuss Design and AI next week, based on our recent research series. We have a great set of speakers on the panel...sign up and more details here.

Shelley Evenson, who is on that panel, is a friend and colleague of Mark's going back to 2012. She built and ran an exceptional learning organisation at Fjord and then implemented a whole new process that drove Accenture engagements across the firm. She is an extremely thoughtful and experienced design expert.

We are really pleased to have her write our first ever guest essay on how we learn and the effect AI is having on that. Shelley - thanks and over to you......

Every AI deployment runs two learning loops. The machine's loop closes by default: Models get smarter with every interaction, every correction, every document they ‘read’. The human loop—where people get smarter and more creative by building on the model's output—closes only if someone designs it to. In most organizations today, no one has. Knowledge gets generated, but human expertise doesn't always grow.

My design instincts from many years as a managing director in a major consulting firm tell me this should worry a leadership team more than adoption numbers. Your competitors are buying the same models you are. AI capability everyone can buy does not differentiate anyone. What still separates organizations is what stays asymmetric: the judgment of your people, their creativity to build on that judgment, and the culture that makes good people join, build relationships, and stay. If your people's judgment dulls and their creativity falters as the model learns and adapts, your output converges on what the model produces. So does everyone else's. A sea of sameness, delivered at the same speed and price by anyone with the same subscription. And the people who joined you to grow will leave for a culture where they still can.

Human judgment is the ability to interpret information, weigh competing alternatives, draw on experience and context, and make reasoned decisions to take action and create something new. AI commoditizes access to intelligence. What everyone is calling taste today is really about identifying quality. Judgment determines responsible action. As a designer I believe quality is absolutely critical, but judgment drives organizational action.

The standard hedge—let judgment erode now, hire it back later—assumes you will be able to. If de-skilling is distributed, as a recent C-suite survey suggests, everyone arrives at the same lateral market at the same time, bidding for a shrinking pool of people who still exercise judgment and drive creative services and products (Goel, Martin, & Kaffe, 2026). The firms that keep developing their people won't be buying in that market. They'll be the ones everyone wants to raid.

Even the industry's boldest voices now describe this era as a cognitive loop between people and digital systems—Satya Nadella made exactly that claim in June (Nadella, 2026). Loop, singular. There are two, but the vendors design and build only one. The second loop is yours to design and build.

The problem no dashboard shows

Dashboards measure adoption: seats, tokens, drafts generated, hours saved. These are fast signals, and they are all moving up and to the right. What they don’t show is whether your people are learning and growing or coasting and dulling. The slow signals exist. They just aren’t accounted for on the current dashboard. Theoretical neuroscientist and cognitive scientist Vivienne Ming suggests we need a Hybrid Intelligence Index that measures the impact of human AI collaboration to explore this (Ming 2026).

The fast evidence is already worrisome. Many employees admit some of what they send colleagues is "workslop"—AI-generated work that looks finished but isn't—and the average employee loses hours every month cleaning it up (Liebscher et al., 2026). 61% would ask AI before asking a colleague (Accenture 2026), which could read as efficiency until you notice what it costs: the experience of shared experience that builds teams. That experience is the work—coming to a shared agreement over an individual understanding—a team with a shared mental model and goal. BCG calls the workslop problem "distributed de-skilling" and half of C-suite leaders already see it in their own organizations (Goel, Martin, & Kaffe, 2026).

The data is real. In 2025, a multicentre Lancet study found that endoscopists' detection of pre-cancerous growths in standard, unassisted colonoscopies fell from 28.4% to 22.4% after just months of routine AI assistance—a 20% relative decline in a core clinical skill measured. Even the AI assisted detection rate was below the pre-AI unassisted baseline. (Lancet Gastroenterology & Hepatology, 2025). These were experienced physicians. They began to rely more on the AI and their work skills dulled. There is no reason to believe analysts, underwriters, or engineers’ behavior would be different.

Chris Argyris drew the relevant distinction fifty years ago: single-loop learning corrects errors within the existing frame, while double-loop learning questions the frame and the means (Argyris, 1977). AI adoption as practiced today is textbook single-loop—optimizing the speed and cost of work as it stands. But the real opportunity is the double-loop move—asking whether the work itself (in the large and in the small), not just its pace, should be reconceptualized so that both loops function efficiently and effectively together.

This is not a technology problem; it is a design problem.

Two questions

Whether AI makes a person sharper or duller comes down to two questions you need to ask, task by task.

First: does this work demand and build judgment skills for the person doing it, or is it routine? Second: what role does the AI play? Does it execute the work outright, or does it scaffold or push the person's thinking, pressure-testing decisions, and surfacing patterns without doing the work for them? We don’t have the data to know if scaffolding will improve learning, but maybe we just haven’t designed it effectively yet?

Four positions

The Handoff space is where expense report coding belongs; it is not where a second-year's first market analysis belongs. And the same task can be designed differently depending on who is doing the work—a market model is routine for a partner who has built a hundred, but a critical apprenticeship for the analyst building her first. That is why blanket automation policies fail: work design is design work, done at the level of task and person, not policy work done at the level of tool.

Imagine a single consulting engagement. Tasks like the initial meeting scheduling, the meeting transcript cleanup, or a formatting pass on the proposal pitch deck can fit neatly in Handoff. Have the AI execute but maintain human review.

Creating the market model for the engagement is where the design choice gets real. If you have the model create it then it’s likely the partner gets something polished by Friday. The client is satisfied but positioning it in the Coasting quadrant has consequences. Three years later the analyst that had AI do it is a principal who has never built a market model from the ground up and cannot tell a sound one from a problematic one. Multiply that person by every analyst in the organization and you have bought this quarter's margin with the next decade's partners skills. Some companies are strating to wake up to this.

Now imagine the same deliverable as a Healthy Stretch. The young analyst frames the market and drafts her own assumptions first. Then the AI goes to work with her as a sparring partner: what would have to be true for this to be wrong, which comparable markets broke this pattern, what are the disconfirming cases, where is the evidence? The analyst works faster and is learning and growing more on the same work challenge. Same task, same tools, same deadline.

The thing that changes is the design of the working relationship with the AI and what people bring to it. Vivienne Ming’s research on human-AI forecasting found only that it wasn’t IQ or technical skill that set successful people apart. It was curiosity and intellectual humility (Ming 2026). This suggests a more complex interaction flow for these human/AI relationships and something to design for.

Three strategies turn the 2x2 into practice.

Find the Why
Everything depends on judgment. Know where judgment actually lives—not the steps on the process map, but the exceptions, the judgment calls, the moment someone says we could, but we shouldn't. That is tacit knowledge, and it does not live in systems of record (Polanyi, 1966). Surfacing it takes field research: shadow the work, interview the people, see where good judgment breaks the standard path for a good reason. These moments are the non-negotiable struggle points—the work that must be designed to stretch no matter what an automation case says.

Design the Dance
Decide what stays human, what is shared, what is delegated, and what runs autonomously across all three interaction types within an organization: person to person, person to AI, and AI to AI. Every choice gets tested twice, simultaneously. Does it produce the business outcome needed—the value, and does it protect and grow the human capability the business will need next—human values? A workflow that hits its productivity target by removing the moments that build judgment and joy in the work hasn't succeeded. It has borrowed from the future. Should AI ever go first? If you design the dance to do that there is a real danger that people will stop learning how to begin with a blank page and you won’t achieve the kind of human AI hybrid intelligence Ming describes.

Read the Wind
Work goals and outcomes change as the people and models improve, and yesterday's approaches go stale. Watch the ongoing signals. To continue with our consulting company example: Can this year's second-years analysts do, unaided, what second-years could do three years ago? Do people’s review comments on top of an AI deliverable document reflect an interrogation of the reasoning—a deep review, or only the formatting—a surface review? Are decisions being reversed or escalated more often? And when people get stuck, do they turn to a colleague for advice or to the AI model? None of these answers appear on an adoption dashboard. All of them are observable data points that could be signals for attrition or client dissatisfaction on your horizon.

Who owns this?

Everyone is an AI manager now—setting goals, delegating deliberately, checking outputs, knowing when to step in. It is like that aspect of people management which involves teaching. You have to build context for the AI just like you would for your interns.

What almost no organization has explicitly defined is ownership of the whole system: someone who has a map of struggle points, watches challenge-and-skill fit across teams and designs the plan for continuous changes in the work as the models improve. That last duty is the point. A plan that hard codes the answer could be wrong within a year or sooner. Continuous re-mapping as the technology and people change needs an owner. I call the role the Human|AI Systems Choreographer but it could just as well be a Human|AI experience officer (H|AIX). Give it whatever title fits your org chart. Having the function with the ownership is the point. Without it, decisions still get made, but now by unplanned default: by whichever vendor ships the next model feature.

Where to start

The organizations that thrive with AI won't be the ones that automate the most, but the ones that learn the fastest—that protect the right human effort and design the work to keep people growing as the technology does—closing the second loop.

This is the work to be done. It begins with a diagnostic that baselines and maps one core workflow into the four positions—what needs to be a Healthy Stretch, what can be handed off, what is overkill, and what is coasting right now. From there, you reimagine the work: what’s the best way to interact, the division of labor, the checkpoints, the pace.

And the thesis is testable. Take two matched teams and one quarter: same deliverable, one team in a Stretch relationship, the other delegating to the model. Score the work blind, then score what each team can do unaided. If the Stretch team isn't at least as fast and measurably stronger, the framework fails its own test. Few approaches in this conversation will offer you that.

If your dashboard says AI is going well, and you cannot yet say how your people are getting better your second loop is yet to be designed. Test it.

shelley@seespace.ai

Sources

Accenture. (2026, July 26). The Pulse of Change: Business and technology trends.
https://www.accenture.com/us-en/insights/pulse-of-change

Argyris, C. (1977, September–October). Double loop learning in organizations. Harvard Business Review.

Goel, S., Martin, D., & Kaffe, C. (2026, June 17). When everyone uses AI, companies risk losing critical skills. BCG Institute.

Liebscher, A., Lee, A. Y., Rapuano, K., Kellerman, G., Niederhoffer, K. G., & Hancock, J. T. (2026). Workslop: Examining the prevalence, antecedents and consequences of low-quality AI-generated content at work [Preprint]. Stanford University / BetterUp / Boston Consulting Group.

Ming, V. (2026, April 24). Is AI smarter than people? The Wall Street Journal

Ming, V. (2026, June 8). We need to learn how to argue with AI. Financial Times.

Nadella, S. (2026, June 14). A frontier without an ecosystem is not stable.

Polanyi, M. (1966). The Tacit Dimension. University of Chicago Press.

Budzyń, K., Romańczyk, M., Kitala, D., et al. (2025, August 12). Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study. The Lancet Gastroenterology & Hepatology, 10, 896–90