This means that everything—intro, chapters, front and back matter, foreword, cover art—is submitted and locked for launch with Rosenfeld Media.

It’s fitting that this milestone coincides with the 100th episode of this newsletter, because the deceptively simple act of writing something for an audience on a regular cadence played a non-trivial role in bringing this book into existence.

Origin Stories

I published my first Substack on October 11, 2023.

I had recently quit my in-house principal UX researcher job at Meta Reality Labs to build my own design research practice. Bundled into this new endeavor was a strong yet simultaneously vague aspiration: Write a book.

Perhaps it would be something about design research for emerging technologies, where I had focused across my career.

Maybe I’d write about what’s the same and different about conducting research for spatial computing products versus “traditional” 2D interfaces.

Or maybe, it would be something about how we navigate building for this (relatively) new kid on the block, generative AI, that seemed like the next big thing for HCI since the Mother of All Demos in 1968.

I didn’t know exactly where to start, other than I needed to start writing.

I also knew I was prone to both the delicious prospect of endless rabbit-holding (risk: never choosing your topic) and the deep-seated perfectionism of a classical music upbringing (risk: endless iteration).

Because of that, having a regular date where I hit “publish” (or more specifically, “send to everyone now”) and an audience was non-negotiable.

I poked around a bit on platform options and ended up choosing Substack. I needed a low barrier to entry plus built-in discoverability. The platform fit the bill well enough.

Getting to the Point (of View)

Writing wasn’t entirely new to me.

Seventeen peer-reviewed publications and five patents in, I knew the drill when it came to academic or technical writing.

In those forums, you have a well-established structure (for example: intro, methods, results, discussion, general discussion, conclusion). You also have a fairly clear direction about what the piece will be (though that rarely makes the blank page easier).

This was not that.

Having the obligation to populate a blank page every week is both an incredible privilege and a curse. (And yes, “weekly” was my initial self-imposed cadence for over a year; I’ve since moved to biweekly to juggle other priorities).

Where would I start? I decided that writing about stuff I was doing and reading was a reasonable starting place.

My first four episodes were indeed that: digests of events I attended (I went to a lot of them in those first few months of founding my consultancy, Sendfull).

Around then, I noticed I was developing an opinion about things.

By episode 5, in part inspired by a client project on an augmented reality (AR) system at the time, I had Something To Say: We need to stop treating AR like it emerged from nowhere.

We’ve been designing heads-up displays for decades, with an accompanying wealth of aviation human factors research that we can learn from to inform modern AR design (you’ll notice exactly this stance in my book: how decades of work on designing automated systems can inform AI automation decisions).

I had been applying that literature for over a decade in my design research work, but only through the act of staring down a blank page every week did I start articulating my point of view.

A Thesis Emerges

Episode 14 was called, Why we need a new design framework for building gen AI.

This was the concrete seed of core concepts that became a peer-reviewed paper I published at EPIC People later that year (and now, my book).

The basic thesis was that we need to start with understanding technology capabilities (“could you do it?”) and marry that with user value (“should you do it?”). I’d subsequently developed a corresponding toolkit grounded in systems thinking, hands-on intuition building, and a human-factors-informed 2x2 (the Cognitive Offloading Matrix) to help you answer those two questions.

Those same questions now serve as the axes of a keystone framework in my book, the Autonomy Decision Matrix, which is a map for deciding how to automate.

I’d go on to evolve the framework and tools from my EPIC paper into a playbook for how to automate through consulting and teaching. These efforts eventually culminated in me emailing Lou Rosenfeld a book proposal that would become What Your Machines Should Do.

Long-Term Side Effects

It’s worth mentioning that writing on Substack wasn’t only about developing this thesis, at least not in the strict sense of the word. Once I’d been writing regularly for several months, I improved at strategically nerding out pointing topics in ways that could complement my consulting and teaching practice. For instance:

• Preparing for a speaking event via a literature review or landscape analysis, like this piece on generative AI’s role in extended reality accessibility, ahead of an Augmented World Expo panel.

• Building out easily-accessible materials for my teaching practice.

• Developing the scaffolding for new courses, such as Designing Future Systems (thoughts: part I and part II).

• Working in the open and sharing primarily research that informed my new UX for AI course.

After enough repetitions, I also started noticing a pull: Once I’d spotted a pattern and it reached an inexplicable threshold (maybe it was thematic saturation, idk), a ‘sensemaking needed’ lightbulb would go off. I knew I needed to write about it.

I’d quickly find my thesis and the draft would flow (my forward-deployed UXR episode is one example). What happens if I keep doing this for another three years? Another decade?

Subscribe now

What’s Next

One-hundred episodes and one book later, I’m rethinking what this newsletter will become.

I’m experimenting with an alternating cadence of practitioner spotlights (I do this in my book and it resonated with my tech reviewers) and long-form essays about UX, automation, and strategy. If you have qualitative feedback to complement my quant dashboard metrics, I’d love to hear from you in the comments.

In the meantime, I want to leave you with five learnings for writing more:

1. Write regularly. Set a cadence and stick with it. Put it in your calendar. Find an accountability buddy. Whatever it takes to make writing a regular part of your life.

2. It doesn’t have to be long. Clearly, I struggle with this. But you can start small, and that makes the mountain more like a mildly-steep sidewalk. Start with 100 words and adjust accordingly.

3. Lower the barrier to (text) entry. If typing into computers isn’t your thing, don’t force it. Use an old fashioned notebook. Brain dump into your phone’s Notes app using voice-to-text. Use sticky notes. Experiment. Just get it out of your head, somehow.

4. Share your work. This one feels especially scary because it is. What we publish on the internet lives forever. However, at least for me, this is the push I need to hone a piece. If you’re not ready for this yet, share it with a friend first. Or event text it to yourself and let it wait a day. It will look different tomorrow, I promise.

5. Expect discomfort. Know it will pass. Writing is Type 2 fun (also known as ‘fun’ to yours truly). It’s ambiguous and uncomfortable, almost continuously, until you hit publish.

Every Monday night (or Tuesday, or Wednesday) when I sit down to Thursday’s article, I think, “Ugh, why do I do this to myself?”

Yet, almost without fail, I find myself sharing a link to my much-clarified thoughts with clients or students or friends mere hours after publishing. I hold onto that for the next Wednesday night struggle.

Keep going. The newsletter provides.

⏪ Recent Episodes

ep. 99. This is AI’s Dial-up Moment. What Kind of Internet do you Want to Build?

ep. 98. Putting the “Socio” Back in “Sociotechnical”

ep. 97. How an Embodiment Architect Designs for Physical AI

🚀 Sendfull in the Wild

Shift UX Conference (Brooklyn & Virtual)

What’s happening with UX?! Is it really dead? How do we evolve? Come join us at the Shift UX 2026 conference to collectively wrestle with these questions and chart a path forward.

I’m delighted to be one of the curators, leading the human-centered tooling track. Our other tracks include: research as strategy led by Jemma Ahmed, operations as orchestration led by Jon Fukuda, and systems-level thinking led by Abby Covert.

👉🏼 Read more about how I’m thinking about The Shift in this interview: What Will the Future of UX Actually Look Like?

EPIC 2026 Conference (Chicago)

This year’s EPIC Conference theme is Context: How do we model, predict, and interpret it to create value for people and business?

Join me and co-chair Todd Palmer for an immersive, hands-on learning day before the main conference, Building with Context: How to Translate Human-Centered Insights into Working Prototypes with Vibe Coding.

💡 Fun Fact

While the field of UX might be shifting (shameless plug for the upcoming Shift UX conference), there’s evidence that “UX” is still the term for what we do.

An April 2026 Nielsen-Norman survey of 604 design and tech professionals supports that “UX” remains the default descriptor for the discipline. While there are other descriptors, they’re highly fragmented.

👉🏼 Full article: No New Name Has Replaced “UX”

Share

That’s a wrap 🌯 . More on UX, automation, and strategy from Sendfull in three weeks!

Elena Yunusov is on a mission to move the AI industry from marveling at raw technical “fire” to making meaningful “soup” that nourishes human flourishing.

She’s the founder and executive director of Human Feedback Foundation (HFF), a Toronto-based nonprofit that prototypes more open, human-centered futures for AI.

Before founding HFF, Elena’s career journey included leading the responsible AI strategy at RBC, Canada’s largest bank; building communities through initiatives like Crypto Camp and Maker Festival; and working as a journalist focused on technology and art. Her roles are connected by a throughline of steering bleeding-edge technology toward public benefit and preserving human agency.

The Cathedral & The Bazaar

Elena has had a front-row seat to several disruptive technologies, from mobile, to social media, to crypto. When I asked where she situates AI, she named one comparable moment: the rise of the internet, matched in both promise and risk.

“I remember what life was like before: the chunky dial-up connection, the open web, the dreams of the cathedral and the bazaar. I remember the kind of optimism I had about technology literally improving my life and bringing people closer. Looking at the reality now, this is a watershed moment in the exact same way. We have so much more experience to learn from, but I worry we’ll repeat the same mistakes, only now at a new scale.”

Elena is concerned that AI will repeat the internet’s path by hardening into—in the words of HFF Chief Technologist Eric Boyd—a “corporate singularity” where power and your personal data are concentrated in the hands of a few elite organizations. One of the reasons she founded HFF was to intentionally prototype a different future for AI.

The Case for Open-Source AI

To Elena, a key path to preventing the corporate singularity is open-source AI, decentralizing control of generative models.

“Open source was the bad guy in the news, where criminals and cybersecurity threats come from. Meanwhile, closed models—as capable and magical as they were—remained concentrated in very few hands. You can talk about alignment all day, but the sheer ability to manipulate people by virtue of knowing more than their search history about everybody, all the metadata, the geography, financial information—it’s Know Your Customer-level. Goodbye privacy.”

Because of that, HFF advocates open source, community solutions, and human-centered AI.

She notes how, despite years of human-centered design work starting with IDEO, it feels like we have lost sight of those lessons. Elena wants to bring that wisdom back into the fold.

“If technology—AI or no AI—isn’t contributing to human flourishing, then what are we doing?”

On Geopolitical Sovereignty

Prime Minister Mark Carney recently unveiled Canada’s National Artificial Intelligence Strategy: “AI for All”, which promotes “advancing open-source AI for resilience and choice”. Because of that, I was curious to hear Elena’s perspective on the new strategy.

To her, “AI for All” is one example of how a country is addressing urgent global questions regarding technological sovereignty for human benefit. For any country to maintain that technological sovereignty, investing in open source is a critical geopolitical and economic imperative.

Participating in the creation of new open-source protocols and governance mechanisms ensures that a society has a say about how good its technology will be, rather than just being a passive consumer of what the “corporate cathedrals” offer. Relatedly, open source frees you from reliance on a small number of organizations.

There’s another practical reason why Elena thinks open source will win out over closed models: compute is a finite resource.

“People in the AI Tinkerers community have thought this through: what’s the maximum compute achievable if we turned planet Earth into a data center? Then build a few data centers on Mars and the Moon? It’s still a finite number. That leads to the natural conclusion that we won’t use frontier models for everything; we’ll need to diversify. We can use the frontier models now to bank the window and build the open-source alternatives faster.”

Shifting from Compliance to Benefit

For Elena, open-source AI also provides the decentralized agency required for human-centered design to function as a meaningful industry anchor.

She believes that by anchoring technology in the human experience, the industry can restore “Responsible AI” to its ethical roots, rather than the corporate compliance checklist it has often become in the race to commercialize.

She argues that for AI to remain “socially allowed” to exist, it must earn a “social license” by being provably beneficial to human flourishing rather than just safe.

“Normal people don’t use the words “responsible AI”. Instead, they evaluate technology based on a simpler binary: is it good AI or bad AI?”

By prioritizing utility over a “build fast, break things” mentality, practitioners can create products that solve real problems instead of adding to daily frustrations.

“Good AI has to be human-centered AI or else it’s tobacco all over again. The social license will evaporate. It’s not enough for things to be safe, they also need to be beneficial.”

Subscribe now

The Evolution of the AI Builder

Elena is already seeing signs of the shift toward the prioritization of building useful AI experiences in the builder community.

Through her frontline view of the 5,300-person AI Tinkerers community, she has seen what was once novelty-driven fascination with raw capability turn into disciplined focus on creating real-world value:

“It’s not ‘look at this cool thing, what can I make it do?’ It’s ‘how do I make it do the thing I want?’ How do I make soup with the fire, not just look at the properties of the fire.”

This evolution has triggered a complex identity crisis within the industry:

Some veteran builders mourn the erosion of craft and loss of authorship as traditional coding becomes commoditized. These are often the people building the infrastructure where reliability is non-optional. They see vibe coding as a source of significant technical debt and potential system failure.

Other builders are pivoting toward the challenges of architecture and orchestration. These are usually developers pushing the boundaries of AI, willing to accept high risks for the sake of progress.

Ultimately, the industry will need developers from a range of backgrounds to build robust, useful AI systems that move the field forward. That said, Elena feels that people from the reliability camp deserve more recognition, as they play a critical role in the technology’s future:

“If AI is how we build technology, we need to find ways to design critical infrastructure with AI. That takes time. The stakes are higher.”

The AI Hype Check

I ask every practitioner I interview the same question: is AI overhyped or underhyped?

“Easy—it’s underhyped, for sure.”

That’s because the average person hasn’t yet felt a significant shift in daily existence comparable to the printing press or the internet.

People may mistakenly view AI as just another technological wave like social media or mobile, failing to realize that the ground underneath them is fundamentally shifting.

For Elena, this creates a critical window of urgency:

“The foundational patterns of the AI era won’t emerge organically; they must be intentionally prototyped now to ensure we don’t end up in a future we might not like living in. The current state of AI is like a bunch of atoms that need to become solid matter. Now is the time to shape them.”

Given the sheer number of opportunities to shape AI, a relatively small organization like HFF needs prioritization criteria. Elena describes the one question that guides her team’s decision-making: What’s the no-regrets bet?

“I will not have regrets by helping nonprofits adopt AI. Even if my success rate is 55%, no regrets, guaranteed. I will not have regrets by exposing computer science students to open source and human-centered ways of building. I’ll have no regrets by supporting artists exploring what AI means for society.”

What if we all started taking more no-regrets bets?

This interview has been edited for length and clarity.

Three Things to Try This Week

Here are three takeaways from Elena’s insights that you can action on this week:

1. Test each AI feature against the social license bar. Do the people you’re building for call your product “good AI” or “bad AI” based on their day-to-day experience with it? If your team cannot answer confidently, you have a research or design gap to close before shipping.

2. Lead your specs with the “soup”, not the “fire”. Rewrite your next AI feature spec so it opens with the human outcome, not the technology capability.

3. Prototype one optimistic future your team would defend. Write down one future you want your product to make possible, then work backwards to name the step your team could take toward it this sprint. These incremental decisions are what will give shape to the “AI atoms” still taking form.

   Subscribe now

   👋 Know a practitioner navigating AI in their work who should be featured here? Reach out to hello@sendfull.com

   ---

⏪ Recent Episodes

ep. 98. Putting the “Socio” Back in “Sociotechnical”

ep. 97. How an Embodiment Architect Designs for Physical AI

ep. 96. How a Creative-Technologist-Turned-Product-Leader Meets People Where They Are with AI

🚀 Sendfull in the Wild

What’s happening with UX? Come join us at the Shift UX 2026 conference to find out. I’m delighted to be one of the curators, leading the human-centered tooling track. Our other tracks include: research as strategy led by Jemma Ahmed, operations as orchestration led by Jon Fukuda, and systems-level thinking led by Abby Covert.

📖 One Good Read

Can AI Make Scientific Breakthroughs? by Iulia Georgescu

TLDR: While AI excels at processing formalized scientific literature, it remains unable to produce major breakthroughs. That’s because it lacks access to tacit knowledge, the unrecorded, experience-based deep craft and social understanding gained through trial and error that are essential to the creative process of discovery.

Why it made my shortlist: While I’m a radically optimistic early adopter and AI power user, I’m also extremely skeptical about claims of fully-automated futures.

Across my book, Substack, and recent QRCA keynote, I argue: Tacit knowledge, embodiment, and situated action comprise the durable human core that resists automation. Because of that, we’ll always need to be in or on the loop, more than the AI hype suggests. The goal continues to be joint optimization between humans and machines, not replacement.

Shout out to for putting this one on my radar.

That’s a wrap 🌯 . More on UX, automation, and strategy from Sendfull in two weeks!

Share

Longwalling replaced the traditional “hand-got” method, where small teams of miners worked in individual stalls to manage the entire coal-extraction cycle using handheld tools like pickaxes, wedges, and shovels.

This advanced machinery converted the coal face into a mechanized production line where cutters and drills worked in tandem to slice and blast the coal wall loose, while power-driven conveyors continuously carried the material away to the main transport tunnels.

Management expected that this new high-tech approach would yield increased coal output. Except, the opposite happened: productivity decreased.

In addition, absenteeism increased to nearly 20%. Labor disputes broke out. What was going on?

Enter the Researchers

That question was also something asked by Ken Bamforth, a newly-minted research fellow at the Tavistock Institute, a British non-profit research and consulting organization.

He had learned of this productivity paradox from catching up with his old colleagues at a South Yorkshire coalfield, where he had previously worked for 18 years.

As part of his training in industrial fieldwork, Bamforth returned to visit his old mine. The goal of the assignment was to apply his newfound research skills and report back any new perceptions or practices he observed.

Bamforth noticed that, despite the decreased output reported elsewhere, one part of the mine, the Haigh Moor seam, was thriving. The workers were highly productive and seemed satisfied with their jobs.

Back at Tavistock, Bamforth reported his findings, where they caught the attention of Eric Trist, a psychologist and one of the institute’s founders. One thing led to another, and the two men were off to study work at the Heigh Moor seam to see if the success of this one group could be replicated.

Unearthing a Theory

When Trist and Bamforth descended into the mine to observe what was happening, they discovered that miners had rejected the rigid, assembly-line layout required by the new machinery. Instead, they restructured their workflow into self-managing, multi-skilled teams.

This social structure was the missing link needed to understand why the introduction of the allegedly-superior longwall method was failing elsewhere.

Under the old hand-got method, small groups of miners operated with independence, managing the entire extraction cycle themselves and relying on close kinship ties for safety and emotional support. Trist and Bamforth called this responsible autonomy.

The longwall method fragmented these groups, isolating workers into three distinct shifts (cutting, ripping, and filling) that rarely, if ever, communicated. A culture of mutual scapegoating emerged.

For example, workers on the filling shift frequently inherited bad work from the previous shift, such as “gummings” (debris) left in the undercut or “sticky tops” (coal clinging to the roof). This bred resentment between miners and often led to skipping work altogether.

In contrast, the miners in the Haigh Moor seam spontaneously adapted the old ways of working in a way that complemented the new technology. This joint optimization resulted in the originally-expected productivity increases.

Trist and Bamforth wrote up their findings in a 1951 paper, which would form the foundation of Sociotechnical Systems Theory, which states an organization succeeds only when its people (the “socio”) and its technology (the “technical”) are designed to work in harmony.

If you optimize the “technical” without the “socio,” the technology works on paper but underdelivers in adoption, workflows, and outcomes.

Subscribe now

The AI Longwall

Modern AI adoption looks a lot like the longwall approach: a new, powerful technology is introduced into organizations, but expected returns don’t follow.

Forbes reported on exactly this issue last week, spotlighting the manufacturing sector; zero of the manufacturing business leaders surveyed in a recent Grant Thorton study reported significant revenue uplift from AI initiatives. While they were seeing some efficiency gains, 48% were stuck in pilot mode.

Forbes concludes that the fix is better procurement discipline, not a better AI model.

What is “procurement discipline”? It’s the intentional design of an organization governance system that ensures technical tools are deployed alongside a clear social structure of accountability. In short, it’s part of the “socio” in sociotechnical systems.

Similarly, in June, Bain & Company reported that nearly 40% of companies measuring their AI outcomes landed in the 0-10% cost-savings bucket, versus the 11-20% they had targeted. (Fun fact: 90% of the companies that missed their AI cost-savings targets are increasing their budgets again for the next wave).

The companies that were delivering on their AI targets did so by “treating data access, governance, and process redesign as CEO-level problems rather than IT problems”.

This overpivot on the “technical” at the expense of the “socio” is also on full display with OpenAI spending $4BB, and Anthropic, $1.5BB, on ventures that send in forward deployed engineers to diagnose and attempt to fix why powerful AI systems aren’t reaping promised benefits. (Why we’re not sending in experts whose deep well of knowledge is focused on the “socio” is the topic of another post).

Why This Keeps Happening

Seventy-five years later, the challenges of AI adoption echo those seen with the introduction of longwalling. Why do organizations keep favoring technical subsystems over social subsystems?

A Formal Rationality Mindset

Organizations continue to view the workplace through an industrial age lens, treating it as an organizational machine where both processes and humans are expected to behave deterministically.

This mindset is driven by formal rationality, a concept rooted in sociologist Max Weber’s work, which designs bureaucracies to replace human judgment with the dictates of rules, regulations, and structures. The mindset prioritizes quantification and control, mistakenly assuming that a technical system’s success is a repeatable chain of cause and effect where the whole will always equal the sum of its parts

The Governance Gap: Procurement and Ownership

Modern technology adoption is frequently driven by competitive pressure and the fear of falling behind rather than a disciplined analysis of specific operational needs. Because the motivation is reactive, projects often lack a named executive owner or clear financial metrics, leading to “pilot purgatory” where technical activity is mistaken for measurable P&L results.

Habituation to “Bad Systems”

When a technical system is socially dysfunctional, workers develop psychological defenses, such as the mutual scapegoating that emerged after the introduction of longwalling, to survive the environment. Over time, groups become habituated to these “bad systems” because the dysfunction offers a stable, albeit broken, status quo where individuals can remain anonymous and avoid personal accountability for system failures.

A Path Forward

Ultimately, moving towards joint optimization requires us to stop treating technology as a standalone fix and start designing the social systems to support it. A few ways to do this include:

• Implement Procurement Discipline: Treat technical investments as capital requests that target a pre-costed business problem and a specific financial metric. The probability of success increases by assigning a named executive owner to the outcome and enforcing a firm “kill date” if measurable value is not achieved.

• Empower Experts in Human Behavior and Organizational Systems: Elevate specialists with deep wells of knowledge in the “socio”, such as design researchers, to bridge the gap between technical tools and organizational needs. For design researchers reading this, your role needs to adapt too, such as taking a strong POV about what to build and taking an active role in implementation.

• Encourage Responsible Autonomy: Grant teams the authority to manage their own internal regulations and task interchangeability, allowing them to experience a complete cycle of operations. This approach helps ensure that human actors are meaningfully integrated with the technical process rather than fragmented by it.

Subscribe now

⏪ Recent Episodes

ep. 97. How an Embodiment Architect Designs for Physical AI

ep. 96. How a Creative-Technologist-Turned-Product-Leader Meets People Where They Are with AI

ep. 95. How an Interdisciplinary PM Owns the Output of AI

🚀 Sendfull in the Wild

QRCA Keynote: July 21

I’m excited to be giving the keynote for the Qualitative Research Consultants Association’s Qual Tech Days 2026 conference on July 21.

As AI automates more tasks, expert judgment is becoming an increasingly scarce and valuable resource. In my talk, 𝗚𝗼 𝘄𝗵𝗲𝗿𝗲 𝘆𝗼𝘂’𝗿𝗲 𝘀𝗰𝗮𝗿𝗰𝗲: 𝘄𝗵𝘆 𝗾𝘂𝗮𝗹’𝘀 𝘀𝘁𝗿𝗲𝗻𝗴𝘁𝗵𝘀 𝗮𝗿𝗲 𝘄𝗵𝗮𝘁 𝗔𝗜 𝗺𝗶𝘀𝘀𝗲𝘀, I’ll cover why qualitative researchers are uniquely positioned to thrive in an automated future, and how the work we do best (reframing problems, making non-obvious connections, and uncovering tacit human knowledge) may become even more important in the age of AI.

✍️ Register for the conference here.

ShiftUX: Sept 23-25

What’s happening with UX? Come join us at ShiftUX to find out. I’m delighted to be a curator of the human-centered tooling track, alongside these amazingly talented humans.

✍️ Register for the conference here (virtual and IRL options available). Use HUTKA-shift26 for $75 off.

📖 Good Reads

Ford Rehires Human Engineers After AI Fails to Match Quality Checks (BBC): History repeating itself (Klarna, Boeing, Tesla…).

Context Architecture (Nielsen-Norman Group): A framework that applies information architecture principles to the “ecosystem” of AI context (instructions, memory, tools) to help systems generate outputs that are accurate and aligned with human mental models.

HCII AI Starter Pack (Dan Saffer): Access to a collection of Carnegie Mellon University’s Human-Computer Interaction Institute’s prominent AI papers of the last several years.

That’s a wrap 🌯 . More on UX, automation, and strategy from Sendfull in two weeks!

Share

Binna Lee is an embodiment architect.

As a design researcher, extended reality (XR) designer, and design educator, her practice centers on building immersive experiences that bridge the gap between our internal, psychological worlds and our external, physical environments.

Binna’s story begins in South Korea, where she developed keen observation skills from a young age. Her parents weren’t always expressive in their facial or body language; emotions often had to be read from more subtle signals.

That early sensitivity deepened through art education.

As Binna trained in observational drawing, the practice of drawing what is in front of you rather than from memory, she learned to study the world with disciplined care: the proportion of a body, the structure of a face, and the way light hits an object.

“Most people only look at an object to recognize it. Drawing forces you to truly observe.”

Binna wasn’t just learning to draw or paint something accurately. She was learning how to see in systems, looking beyond surface appearances into their underlying structures.

This foundation led her to study product and industrial design, before moving to New York City for graduate studies at Parsons. Here, she integrated neuroscience into her toolkit, providing a formal vocabulary for the perceptual and cognitive processes she had practiced since childhood.

For example, she describes translating a three-dimensional phenomenon onto a two-dimensional page as a meticulous sensorimotor experience:

“The brain begins to dissect physical reality: light, shadow, materiality, dimensional perspective, proportion, structure, and movement.”

Binna also found herself applying her observational practice in new ways in daily life, adapting to a new set of cultural norms.

These threads converged into a central inquiry: how invisible inputs like family upbringing, cultural environment, and media exposure shape what a person believes is possible.

Today, Binna teaches design at Parsons School of Design, serves as the Vice Chair of Women in Design NYC, and is developing Insight Realm, an XR experience that translates that inquiry into spatial form.

The Body as the First Interface

Through her design practice, Binna has developed a working research framework: The Body as the First Interface.

Her framework is grounded in scientific research on how biology, psychology, and culture shape us, which she reinterprets through XR, spatial cognition, and design research.

She advocates for a shift away from 2D screen design toward a more embodied experience. Her reasoning is that because humans are three-dimensional, multi-sensory beings, our technology should engage our physical bodies rather than just our eyes.

To illustrate her point, she references literature on how babies develop:

“Cognitive development is grounded in sensorimotor experience, social feedback, and the sense of moving through space that lets us locate ourselves in the world. You don’t get that through a screen.”

Translating theory into practice, Binna designs immersive experiences in which the body is the primary interface for human intelligence, exploring how physical AI and spatial computing can better align with our physical reality.

Her industrial design experience, spending years studying and building three-dimensional objects, directly influences her work. She thinks in terms of weight, texture, material, and fit when designing experiences, making spatial thinking second nature.

The framework is also a kind of homecoming back to her formative observational experiences, where the convergence of visual, spatial, and motor processes served as her first blueprint for what “body-as-an-interface” feels like.

The Perception Box

Another central pillar of Binna’s practice is revealing and dismantling what’s been called the “perception box” in cognitive discourse: the invisible mental framework through which we make sense of reality.

Rather than seeing perception as a purely internal process, Binna argues that it’s continuously shaped in a dynamic loop: we shape our perception of the world, while the world simultaneously shapes our perception of ourselves.

Binna experienced this firsthand when immigrating to the United States. Navigating a new culture required continuously interpreting unfamiliar social cues, body language, and conversational rhythms. Everyday interactions demanded deliberate observation. This experience revealed something fundamental about perception:

“We observe most deeply when our assumptions stop working.”

No longer able to rely on automatic interpretations, she had to consciously reconstruct meaning through the interplay between her body and environment, bringing previously unconscious frameworks of her perception into clear view.

This realization fundamentally shaped her approach to design: Rather than viewing vulnerability as something to avoid, she came to see it as a powerful catalyst for expanding perception. By questioning our own perceptual frameworks, we also become more capable of understanding others’ frameworks.

These learnings directly inform Insight Realm, an XR experience that invites people to explore how their perception box forms: upbringing, school dynamics, and social reinforcement. It then lets people shatter it and rebuild it through a gardening metaphor: planting and tending new perspectives over time, illustrating how rewiring takes repeated practice.

Binna notes that dismantling the perception box is not merely about questioning our assumptions; it’s about recognizing perception as something that can be consciously cultivated.

“By making our invisible frameworks visible, we create the possibility for greater empathy, more thoughtful design, and more intentional relationships with both people and intelligent technologies.”

This practice is increasingly important in an age of digital echo chambers:

“Social media previously amplified our existing perceptual filters. Today’s increasingly personalized AI systems have become another environment that can reinforce—or challenge—the ways we make sense of the world.”

Where AI Fits

I asked Binna about the role of AI tools in her practice.

For visual ideation, she generates images using Adobe Firefly and Google’s Gemini Nano Banana. For 3D work, she uses Meshy AI to turn sketches and images into 3D models, then refines them in Blender, which she drives through Claude over an MCP connection.

Binna views these tools as temporary scaffolding: a stepping stone to quickly bypass technical barriers and manifest ideas in 3D spatial environments.

Relatedly, she sees how AI can play a beneficial role in design education, observing creative expression open up for students who would once have been gated out by tooling complexity:

“The entry point used to require a high bar. Now students can sketch on paper and turn that into a 3D object. The lower entry point means they can focus on what they’re trying to express.”

At the same time, Binna is clear about AI’s limits: Rigging still needs humans. Final asset production benefits from human-in-the-loop refinement. Students still have to know what they want to express. While AI tools can lower the barrier, they cannot replace human spatial design judgment.

Binna also highlighted how her study of human behavior has helped her build intuition about designing to engage and complement our perceptual and cognitive systems. For that reason, she feels it’s important that designers ground themselves in psychology (especially human development!), sociology, and biology to develop a holistic understanding of people.

This is especially urgent for designing AI experiences:

“How are we going to utilize artificial intelligence without knowing ourselves?”

Binna also notes that AI tools share a fundamental challenge: they’re bottlenecked by 2D screen constraints.

Bringing AI Off the Screen

To Binna, AI is not simply another interface. It’s becoming part of the environmental context that shapes our perception, behavior, and routines, in the form of wearables, robots, and ambient technology.

Because of this, its impact will depend on how thoughtfully we architect its integration into the places where people live, work, and connect.

Yet, much of today’s AI design follows patterns inherited from screen-based experiences.

“We’re not thinking deeply enough about the body as an interface—or about how our environments, from furniture to lighting to everyday objects, shape our relationship with AI.”

As an example of physical AI done well, she points to Nome, a domestic robot from creative consultancy Frog, designed to exist in harmony with the user’s home environment.

Binna notes that where and how physical AI exists is just as important as what it can do:

“A robot placed in a living room is experienced differently from the same technology in a workplace because home carries familiarity, emotional attachment, routines, and psychological safety. The surrounding environment becomes part of the interaction itself.”

AI x XR

As an analog to the digital-physical disconnect Binna has observed in AI design, she offered an example from a recent Audi concept she encountered at the Royal College of Art Immersive ACT Symposium.

The car’s dashboard was flat, with no physical controls. All interface elements, like audio, lighting, and navigation, appeared only through AR glasses.

“Looking and interacting with an AR screen disconnected from physical controls? Is that really safe?”

Driving demands immediate cognitive and physical response times. Pulling controls into AR without physical reference points seems more likely to increase cognitive load and reduce situational awareness, rather than the other way around.

Whether you’re building AR glasses or a household robot, the common lesson is that it’s imperative that interaction patterns are designed with embodied and spatial cognition in mind.

The AI Hype Check

I ask every practitioner I interview the same question: is AI overhyped or underhyped?

For Binna, the question itself needs reframing.

‘Over- or under-hyped’ assumes AI is a fixed, narrow conception of a technology that primarily lives on a screen.

“It’s a matter of how willing you are to step outside your current perspective. It’s hard to broaden your thinking past the screen, past typing or voice, into the physical world.”

The question worth asking is what AI becomes once it gets off the screen.

This interview has been edited for length and clarity. Opinions expressed are solely Binna’s own and do not necessarily reflect those of her employer. Follow Binna on Substack here.

Three Things to Try This Week

Three takeaways from Binna’s practice:

1. Design for the body, not just the screen. Even within current screen-based products, ask what the user’s body is doing while they interact. Where is their attention? How does their environment shape it? What multimodal inputs could you draw on to help them interact more naturally?

2. Help open the perception box, both for users and yourself. Instead of designing AI interfaces that simply agree with the user, introduce ‘friction buffers’ or spatial UI alternatives that gently present contrasting viewpoints, nudging the user out of their perception box. For designers, challenge your own existing mental models by intentionally seeking unfamiliar contexts.

3. Open the door with AI, but walk through it yourself. Remember how Binna helped her students overcome the 3D modeling learning curve with Meshy AI? Apply the same approach in your own work: identify a learning curve AI could shorten for you or your users. Keep in mind, you still need to put in the work to make something good.

Know a practitioner navigating AI in their work who should be featured here? Reach out to hello@sendfull.com

Subscribe now

⏪ Recent Episodes

ep. 96. How a Creative-Technologist-Turned-Product-Leader Meets People Where They Are with AI

ep. 95. How an Interdisciplinary PM Owns the Output of AI

ep. 94. How a Leading Health Tech UX Researcher Decides What to Delegate to AI

If you like these episodes, you’ll love my book. What Your Machines Should Do is coming this year with Rosenfeld Media. 📣 Sign up to be the first to know about new book releases, sales, and events.

🚀 Sendfull in the Wild

QRCA Keynote: July 21

I’m excited to be giving the keynote for the Qualitative Research Consultants Association’s Qual Tech Days 2026 conference on July 21.

As AI automates more tasks, expert judgment is becoming an increasingly scarce and valuable resource. In my talk, 𝗚𝗼 𝘄𝗵𝗲𝗿𝗲 𝘆𝗼𝘂’𝗿𝗲 𝘀𝗰𝗮𝗿𝗰𝗲: 𝘄𝗵𝘆 𝗾𝘂𝗮𝗹’𝘀 𝘀𝘁𝗿𝗲𝗻𝗴𝘁𝗵𝘀 𝗮𝗿𝗲 𝘄𝗵𝗮𝘁 𝗔𝗜 𝗺𝗶𝘀𝘀𝗲𝘀, I’ll cover why qualitative researchers are uniquely positioned to thrive in an automated future, and how the work we do best (reframing problems, making non-obvious connections, and uncovering tacit human knowledge) may become even more important in the age of AI.

✍️ Register for the conference here.

📖 Good Reads

AI at Work: Strategy Matters More Than Tools (BCG): As frontline AI adoption reaches 74%, companies must prioritize strategic clarity and work redesign over simply providing tools to convert individual productivity into measurable business value (hey, I know a book for that).

Apptronik Announces Robot Park (Forbes): By launching a 90,000-square-foot Robot Park data factory to train its AI, Apptronik aims to move the humanoid industry beyond experimental demos toward a commercially deployable workforce with its upcoming Apollo 3 robot.

Encounters between Artificial and Human Intelligences in Early-Stage Innovation Work (Will Notini, Angela Kochoska, Dan Perkel): AI offers speed, but breakthrough innovation requires the productive friction of human research to build the empathy and team intelligence that automation lacks. Sidenote: This paper is also the best explainer of the peril and promise of synthetic users I’ve read.

That’s a wrap 🌯 . More on UX, HCI, and strategy from Sendfull in two weeks!

Share

George Goodman considers himself a creative technologist.

He’s been creating at the intersection of media production and technical problem-solving since childhood. Growing up in the home PC era with a fascination with video production, George fondly recalled using a Dazzle to digitize analog video, often to terrible quality results.

When a new creative software tool came out, be it for video, photo, music, or graphics, George was using it before long.

Between navigating constant PC crashes and teaching himself different multimedia software, he built a foundation of technical resilience, fuelled by curiosity about the mechanics underlying the tools he used.

After going on to study film at Georgia State and working in TV production at Turner Broadcasting, George took a video production role at Vistaprint. It was there, while working on the website builder, Webs, that he realized he could have more of an impact on the business in a product role than as a media creator. That insight led him to transition into his first PM role.

From Video Production to Product Management

George noticed that product management asked for something he loved about video production: integrating a range of specialities into a cohesive whole.

Video production involves audio, graphics, lighting, cameras, editing, and more. Product management involves data, customer obsession, marketing, and engineering, to name a few.

In both, he loved that each area had its own challenges to learn, and that his role was bringing them all into something cohesive.

In 2021, George moved to Adobe as a senior product manager. He initially worked on application performance before successfully pitching a move into something more aligned with his creative technologist expertise: generative imaging.

He went on to lead the development of the Adobe Express AI Assistant, which launched at Adobe MAX, the company’s annual conference, in 2025. Today, George serves as Director of Product for Agentic and Generative AI.

Closing the Intent Articulation Gap

When George started pitching natural-language design at Adobe, the common belief was that users couldn’t prompt effectively.

“That’s not the problem. People know how to prompt; we don’t know how to respond. Your whole job is figuring out from the words they use, words that aren’t technically design-oriented, what they actually mean.”

The framing was informed by his own experience as a creative professional.

“If you have a background in design, you know that people usually don’t know how to tell you what they want. Your job is figuring out what clients mean from what they say. You do that with references in all kinds of ways.”

He treated the AI assistant the same way. The product’s job was to interpret what people said and fill in the gaps, rather than demand more precise prompts.

Putting User Control First (Before the Tools Could)

Even in its earlier days, Adobe Express AI Assistant used agentic LLM capabilities to create and edit layered designs. Users could design via prompt, then manually edit them with traditional WYSIWYG tools.

This is noteworthy because George and team built the AI Assistant during a time when diffusion models had significant limitations for design work.

First, they couldn’t handle text. Glyphs were frequently broken, which instantly made the designs unusable. Second, users couldn’t make targeted edits. If they wanted to change one thing about their design, the entire design would also be altered.

George and team were confident that the tech would improve over time. They held fast to their belief that giving people manual control of precise layers but still offering the power of prompting was the best of both worlds.

Diffusion has since closed most of those original gaps. For users who just need a clean single-page design, modern diffusion is good enough; managing layers is more complexity than value for them. For brand-obsessed users and anyone doing heavy text editing, layered control still earns its complexity.

A hard product challenge remains: most users who’d benefit from a diffusion workflow today still say they need layers, because they haven’t had the “aha” moment of how good diffusion has gotten.

George frames the core product job as matching the right users with the right technology for their needs.

“Sometimes that means showing them something new, such as diffusion model workflows, while still giving them an escape hatch to layered workflows. Sometimes it means starting them in layered space because their use case is something outside what diffusion models do well today.”

Mapping the Prompting Landscape

George recognized that prompts vary in how much interpretation they require. He broke the prompting landscape into a two-vector framework: ambiguity and scope.

The ambiguity vector tracks the request’s intent, moving from objective tasks (e.g., “change the headline color to red #bc3503”) to increasingly subjective ones (e.g., “make it feel more winter-y”) and finally to the highly abstract “make it better” request.

The scope vector measures the scale of the change, ranging from the asset level, such as targeting a specific text box or image, to the full design level, where a change affects the overall composition, to multiple designs and design types, such as a campaign. For this last one, think of someone using a single prompt to generate static social media posts, a video, and branded emails.

They anchored their initial development in the quadrant of objective, asset-level changes where the technology was most deterministic and reliable. From that baseline, they systematically built upward toward the more ambiguous, high-scope requests that provide the most value to users.

“The whole thing we’re trying to do is give people without a design background simple ways to get the results they need. You’re only meeting them where they are once you can allow them to be subjective. You interpret that for them.”

This approach creates a central tension in AI product development: the “aha moment” for most customers lives in the subjective, full-design quadrant where they can provide vague instructions and have the AI interpret their meaning.

That’s also where the models are most prone to failure. Addressing those failures requires a “whack-a-mole” approach to engineering, constantly fixing routing issues where a model might take a wrong turn.

That’s why George’s team started where the model can deliver with high confidence while maintaining a roadmap that expands into the ambiguity that’s most important to users.

A Two-Tiered Eval System

The Adobe Express AI Assistant eval system maps directly onto the ambiguity and scope vectors. It runs on two tiers:

Tier 1: Objective Payload Evals

For objective requests, evals focus on the action payload: the underlying JSON instructions that the LLM generates for the software. For instance, for a specific prompt like “change opacity to 90%,” the system verifies that the LLM returned the exact technical command required to execute that edit.

Because there’s a fixed, “correct” response to this prompt, the evaluation process is highly automatable, enabling the team to run thousands of pass/fail tests at scale to ensure the model is accurately selecting the right functions.

Tier 2: Subjective Visual Evals

Outputs that require aesthetic judgment still require humans providing a visual assessment of the final outcome. These evaluations address nuanced design questions, such as ‘Did the chosen red color coordinate effectively with the rest of the existing layout?’, ‘Did a generative image edit achieve identity preservation?’ and ‘Did the model correctly pick up on reference signals and images from other sources?’

Subjective ratings present an operational challenge:

“You can’t have your team rating thousands of evals every day, but you also need to understand how the system is working and continue improving it. In a world where every deployment can break your system, how do you evaluate subjective outcomes at scale?

The answer becomes a combination of manual and AI-based review. At the highest level, manual reviews and annotations become an input into training AI. Over time, AI models get better at matching human feedback and become more reliable and faster indicators of changes in the system.”

George always reminds the PMs on his team, “Evals are the new PRD. They express how the system should work and give you feedback as to whether or not it does. Just because AI can scale eval reviews doesn’t mean you don’t have to look at them.”

Building for a Moving Frontier

The balance between what is machine-rateable and human-rateable is shifting as model reasoning improves.

George noted how each major model release prompts his team to ask which layers of internal scaffolding they can now peel away. He compares the work of building on a moving frontier to a previous experience, creating a website builder:

“We chose to adopt React as the framework before it was mature. You end up creating bespoke solutions for problems the framework eventually solves natively.

You look forward a year or two and see you have a ton of tech debt because you built out solutions that are now established patterns. As you have more capable models, you should be peeling those away.”

We have already seen a few cycles of evolution and capability jumps. George notes that these leaps make you start to question if the problems you’re solving are durable or if they’re bandaids until the next model comes out. He adds, “The world moves so fast now that you have to be strategic in what and how you choose to build.”

Zooming out further, George only sees the models getting better. As a result, frontier models will start feeling like using a firehose to put out a small camp fire. That extra force also comes with monetary cost, and potentially, a precision cost, because more focused models can be more reliable for some tasks.

“There’s business in creating small focused models that are efficient at high value and high volume tasks. This reduces cost, increases speed, and can provide higher quality, more controlled outputs.”

From Home Depot to Apple Store

George uses a retail metaphor to describe the broader transition AI is enabling in software.

Currently, creative tools function like Home Depot: the inventory is all there on the shelves, but you have to find the parts and assemble them yourself.

What most people actually need is an experience more like the Apple Store, where a concierge greets you at the door and walks you through the process.

“You couldn’t have that level of service at scale in software. Now AI is the fundamental change giving you that. All we’re transitioning to now is adding the service layer on top of Home Depot.”

This reframing preserves the Home Depot “inventory” for power users who still need direct control over layers and hex codes, but it provides a necessary service layer for everyone else.

This approach informs a core strategy behind Adobe Express: the AI functions as a service layer, a concierge-style Apple Store experience, while the human directs the system with their intent.

Small Teams, Expanded Ownership

George’s view on product organization has evolved as AI tech has advanced.

“I believe even more in small, powerful teams because the scale of what they can accomplish has grown so much. A person can now write so much code in a day that their PR can hit a dozen code owners in a scaled org.

Now, the higher level question becomes: how do you keep the right quality checks and balances at scale when code volume and velocity is so high? That’s a hard problem when you can’t yet trust AI to have all the context to be the code author and reviewer.”

His operating philosophy is: “bank individuals.”

Once George finds talent, he provides them with decision autonomy and uses his leadership position to quickly clear blockers.

For junior product managers, George’s advice is to become AI-native immediately.

This means vibe-coding on weekends, building things agentically, knowing the latest models and hardware, training your own models, and having opinions on the tools you’ve used.

At the same time, it’s having PM fundamentals of understanding who you’re serving, shipping and iterating based on data and feedback, and cultivating taste (something increasingly important as creation becomes more democratized).

The AI Hype Check

I ask every practitioner I interview the same question: is AI overhyped or underhyped?

George’s answer: it’s appropriately hyped for what it can actually do.

“The scaling is real. It’s going as fast as it looks like it’s going. Within software, capabilities are appropriately hyped.”

The exception is robotics. Physical AI runs into regulation and the risk of human injury. Those are the same constraints slowing self-driving cars.

George anticipates that rising AI capabilities will provoke its own pushback, a renewed appreciation for outputs where being human is the point, like live music, handmade crafts, and paintings.

“I hope there’s going to be a renaissance of people appreciating more analog and human connections. Hopefully all the economic boring stuff gets automated, and empathy, expression, and love come to the forefront.”

This interview has been edited for length and clarity. Opinions expressed are solely George’s own and do not necessarily reflect those of his employer.

Three Things to Try This Week

Building on George’s practice, three takeaways for product practitioners:

• Meet people where they are. It’s the product team’s job to close the intent articulation gap. Match the product to how users express what they want, instead of asking them to phrase it your way.

• Anchor in objective success, then scale into ambiguity. Plot user requests on two axes: ambiguity (objective to subjective) and scope (asset to full project). Start delivering in the most objective, smallest-scope corner, where the model is most reliable. Most users default to the opposite corner (“make it better”). Your roadmap should be expanding toward where they live.

• Decide where your product is Home Depot vs the Apple Store. Map the points where users have to find and assemble things themselves. Those are the places a service layer can meet them. Offer power users direct access to the inventory; everyone else gets a concierge.

Know a practitioner navigating AI in their work who should be featured here? Reach out to hello@sendfull.com.

Subscribe now

⏪ Recent Episodes

ep. 95. How an Interdisciplinary PM Owns the Output of AI

ep. 94. How a Leading Health Tech UX Researcher Decides What to Delegate to AI

ep. 93. Companies are sending Forward Deployed Engineers into organizations to study users. Why not send UX Researchers?

If you like these episodes, you’ll love my book. What Your Machines Should Do is coming this year with Rosenfeld Media. 📣 Sign up to be the first to know about new book releases, sales, and events.

📖 Good Reads

The Futility of Predicting AI Job Exposure (Benedict Evans): Evans argues AI’s job impact can’t be quantified, because technology reshapes roles and business models rather than just automating tasks. Past automation often triggered Jevons Paradox, where cheaper processes unlock new demand.

Phone Bans are Spreading Across the US Workplace (The Independent): More US employers are using phone restrictions and lockable pouches to cut digital distraction, protect sensitive data, and encourage face-to-face collaboration.

A Review of Experiments with Synthetic Users (Jim Lewis & Jeff Sauro): Synthetic users promise cheaper, more scalable research, but peer-reviewed studies show they often fail to match the depth and variance of real participants. They can be useful for querying stable existing datasets, but they’re too imprecise to reliably replace humans in high-stakes decisions.

That’s a wrap 🌯 . More on UX, HCI, and strategy from Sendfull in two weeks!

Share

Alissa Stover is a VIP: a Very Interdisciplinary Person.

Her resume includes “professional ballet dancer,” “data scientist,” and “product manager,” connected by the throughline of driving positive human impact.

“As a performer, I always wanted to touch people, to have some kind of impact on them. That extends today as a product manager: focusing on the impact on the end user.”

Her first memory is in a ballet studio, and she long introduced herself as a dancer first. But after years of training and performing professionally, she felt ballet had drifted from its “why”, with audience impact taking a back seat to artistic directors’ egos.

She enrolled at UC Berkeley for an undergraduate degree in Psychology, then moved into nonprofit and policy work as an analyst on anti-poverty programs.

Wanting to use data to make social programs more effective, she returned to Berkeley for the Master’s in Data Science program. Her choice was highly intentional: the curriculum integrated data ethics into its core coursework rather than treating it as a standalone elective.

That technical foundation helped her work across data science and PM roles in the nonprofit sector for several years. The emergence of AI tools is what let her step fully into product management.

AI as a Bridge to Product

Alissa was initially hesitant to go all-in on product management: she enjoyed the math and coding of her data science role and didn’t want to give them up.

Capable AI tools resolved that tension.

Generating SQL queries in the early days of chat-based LLMs, she could see where the tools were headed, even if they weren’t all the way there yet. Alissa anticipated that AI would reshape the career market into a bimodal distribution: junior operators of these tools on one end, PhD researchers on the other. Neither bucket fit, so she carved out a middle path: a PM role infused with deep data and modeling literacy.

“The tools made me so much faster. I could wear both of those hats and my job was okay with it because I wasn’t dropping the ball anywhere.”

Her first official PM title came at the Stand Together Foundation managing Loop, a platform to help people find social services. Because the team was very lean and lacked a dedicated analyst to pull metrics, she used AI to generate SQL queries that she then refined and verified herself. This integration allowed her to perform data science tasks as a PM from day one, using tools to complete in twenty minutes what would have otherwise taken an hour.

Today, she’s a product manager at Gallup. Asked to describe the PM role to someone outside tech:

“You’re the spider in the web, keeping all of the different perspectives together with that laser focus [on the customer] and making sure that we get that impact. Again, from the dancer’s perspective: what impact are we really driving in the world?”

Own the Output

Alissa’s workflow is anchored by one governing rule: “You really need to own the output”.

In her view, the greatest risk of AI is when a practitioner takes off their “ownership hat” and allows a probabilistic model to become a passive decision-maker.

“AI’s not the thing with agency. You are. You shouldn’t give that away to something that’s like a really great magic eight ball”.

Her technical work reflects the same commitment to human verification.

Drawing on her background at a policy nonprofit where a small data error could impact federal decisions, Alissa maintains that the steps of data exploration cannot be fully automated away.

She still starts the data analysis process by talking to stakeholders, establishing the purpose of the dataset, who owns it, and what documentation exists.

Whether generating SQL to bypass a manual pull or analyzing Gallup survey research data, she treats AI output as a first sketch. Alissa manually verifies the underlying logic, especially complex sampling and survey weighting.

“I don’t think that process is super different with AI. I can just do it a lot faster.”

She only offloads more to the AI when the goal is a simple directional sense (“Is it going up or down?”), where the purpose of the data analysis allows for lower stakes.

What Stays Human

For Alissa, human connection is off limits for AI automation. She avoids delegating the work of hearing from and coordinating stakeholders to AI. Using a bot to send a personal Slack message is a no-no.

The same principle shapes her skepticism about completely delegating product discovery with humans to AI.

Alissa sees a deep reason for customer discovery to maintain human involvement: a PM needs to know the user. That depth relies on the live reading of body language and following spontaneous threads that emerge in real time when listening to customers.

“That embodiment of being with the person is something that’s hard to replace when you really want to get obsessed with knowing your customer.”

She compares trying to substitute AI for these conversations to learning a new language solely through YouTube videos:

“You’re never going to learn as well if you don’t talk to someone who speaks that language”.

Curiosity Versus Perfectionism

Alissa lights up when she talks about experimenting with AI:

“I’m just insanely curious. I just get to have this tool and just try new things all the time. Does it work on this? Does it work on that? What’s the new feature today? It’s just so much fun.”

That orientation is partly innate, partly cultivated. While she was the kid getting in trouble for reading in class, she also believes a growth mindset is directly shaped by your environment.

Reflecting back on her ballet training: “It’s a very regimented art form. There is a way to do things. And in those environments, your natural tendencies towards curiosity get a bit dialed down because of the fear of failure. Perfectionism can really get in the way of learning.”

Alissa sees that same pattern in women’s lower AI adoption rates. The barriers aren’t specific to AI. Girls Who Code, she notes, builds being-okay-with-mistakes directly into its curriculum.

Curiosity also needs grit. Trying new things means messing up constantly; curiosity has to come with the persistence to keep going.

Product Management Futures

I asked Alissa about the potential futures of product management, both what she sees changing and what she wants to see change.

For the former, she sees data literacy moving from optional to baseline. Most products are now powered by a generative AI model, which means PMs need to understand probabilistic systems, model behavior, and the architecture of the data pipelines feeding them.

The competitive moat is also shifting: a company’s differentiator increasingly looks like the underlying dataset and the integrations around it rather than a surface SaaS application.

“If people can buy your thing with an AI chatbot, they’re not going to your website anymore. How do you strategize around that new data interface?”

The change Alissa wants to see is more product managers carrying responsibility for the societal impact of what they ship.

Some AI companies’ focus on responsible tech, she argues, is a competitive advantage. Consumer trust is increasingly driving buying decisions as public concern grows over the societal ripples of product choices, from the impact on young people’s cognitive development to broader worries about the future of work.

“A lot of the consumer decision making around AI is trust,” she notes, which makes responsible tech a primary differentiator rather than a separate moral hurdle.

The AI Hype Check

I ask every practitioner I interview the same question: is AI overhyped or underhyped?

Alissa’s answer: both, depending on the task.

Highly autonomous agents, like the kind you would turn loose on your personal banking, are currently overhyped. In contrast, the potential for AI to solve the mundane parts of professional and domestic life remains significantly underutilized: “Many people are still just using it to draft emails”.

Her advice for those stuck at the email stage:

“Reflect on your day and think, what do you hate doing?”

For Alissa, an early win was offloading the mental burden of creating packing and grocery lists for her two kids: “For moms, just try using it to reduce your mental load.”

Despite her own deep engagement with the tools, she maintains a humble stance on the industry’s trajectory:

“Most of us are wrong about what it’s going to do in the next couple of years. I don’t know which way though.”

This interview has been edited for length and clarity. Opinions expressed are solely Alissa’s own and do not necessarily reflect those of her employer.

Three Things to Try This Week

Building off Alissa’s practice, here are three things you can apply this week.

1. Run the ownership test. For every AI output in your workflow, ask: can I own this if it goes wrong? If the answer is no, keep iterating until it is, or take the task off AI’s plate.

2. Pick a task you dislike and automate. The lowest-stakes entry point into AI is something you would rather not do anyway. Packing lists. Grocery lists. The meeting prep you keep avoiding. Failure costs nothing, and you build intuition for what the tools can and cannot do.

3. Protect human connection. Map where in your week you talk with real customers, partners, or teammates. Keep those for yourself. Use AI for the work around them, not the conversation itself.

Know a practitioner navigating AI in their work who should be featured here? Reach out to hello@sendfull.com.

Subscribe now

⏪ Recent Episodes

ep. 94. How a Leading Health Tech UX Researcher Decides What to Delegate to AI

ep. 93. Companies are sending Forward Deployed Engineers into organizations to study users. Why not send UX Researchers?

ep. 92. How a Design-Leader-Turned-Founder Decides When to Hire AI

If you like these episodes, you’ll love my book. Designing Automated Futures is coming this year with Rosenfeld Media. 📣 Sign up to be the first to know about new book releases, sales, and events.

📖 Good Reads

AI Assistance Reduces Persistence and Hurts Independent Performance: Researchers from CMU, University of Oxford, UCLA, and MIT found that using AI assistance for problem-solving tasks leads to a significant decline in unassisted performance and persistence once the tool is removed, particularly when users relied on it for direct answers rather than hints.

Sycophantic AI Makes Human Interaction Feel More Effortful and Less Satisfying Over Time: Longitudinal research from the University of Oxford, Stanford, and the UK AI Security Institute indicated that sycophantic AI provides effortless understanding that makes real-world human interactions feel more taxing and less satisfying by comparison. Over time, this dynamic can shift users toward treating AI as a comparable alternative to friends and family, potentially reshaping social support patterns at scale.

Job Hunting at the Intersection: Today’s featured practitioner, , shared what product managers need to know right now. One takeaway: The AI-driven job market requires product managers to evolve into builders who can prototype with AI and analyze data pipelines while maintaining strategic judgment.

That’s a wrap 🌯 . More on UX, HCI, and strategy from Sendfull in two weeks!

Share

Madison (Madi) Shultz loves learning about people.

A lifelong fascination with people led her to a Master’s in applied psychology from USC, followed by over a decade as a mixed-methods UX researcher in health tech.

“I always tell people I have the greatest job in the world—getting to know people, hearing about their life, the tasks they do, the jobs they need to perform, the goals they have, and then helping other teams translate that experience to make life work those goals a little easier.”

Madi has worked across the sector’s diverse ecosystem, engaging with patients, providers, pharmacists, payers, channel partners, brokers, and everything in between.

“I consider being a researcher an honor and a privilege, especially in my line of work, where the topic might be as intimate as navigating menopause, or high-risk maternity where your child’s been in the NICU, or trying to find a new doctor because you just moved.”

Today, she balances her work as a senior UX researcher at Oscar Health with her adjunct professor role at Belmont University, where she teaches design thinking. This human-centered grounding shapes how she thinks about where AI does and doesn’t belong in her work.

Researcher as Conductor, AI as Collector

In 2024, Madi co-led the launch of “Tuesday Testers” at Oscar Health, a continuous research program running evaluative studies every other Tuesday, helping keep pace with the company’s fast shipping cycle. With a two-person research team, sustaining that cadence created a clear opportunity to delegate to AI.

She navigates the delegation with a simple framework: “AI is the collector, but I’m the conductor.” Three use cases show the framework in practice:

Tuesday Testers is one example; similar use cases for AI arise across a range of projects. Madi notes, “AI has been a tremendous value lever for our team.”

Intentional offloading to AI is a key part of what makes that value possible.

“It’s the researcher’s imperative to balance, staying on top of tech trends, experimenting so you can learn, but also making an active choice to decide what you’re willing to concede to the tooling. Over-delegating to AI can atrophy the research muscle. The balance for me right now is AI doing the hunting and gathering on data collection that I’ve already done.”

What Stays Human

I asked Madi about her takes on two hot topics in UX research: AI-moderated interviews and synthetic users.

“I’m just not sold on AI as an interviewer. There’s inherent value in humans discovering insights because good research is not just about answers. It’s bringing people along for the journey of discovery.”

It’s crucial for Madi to have stakeholders shadow her sessions, because when they come into a shareout, they talk about what they heard.

“That’s how research really starts to grow legs and get outside of a Google Slide deck: you can collectively experience a human wince. If AI takes the magic of that, I’m not really sure what the point of all this was to begin with. We’re treating research as this logistics problem to be solved. But research is so much more of a social alignment process.”

As Madi explains, that visceral human experience of witnessing a user struggle activates the product team’s mirror-neuron system, fueling the will to action on her findings.

An AI moderator can’t produce that reaction.

As for synthetic users (aka AI “participants”), Madi questions their utility: “If consciousness is to have firsthand experience and our job is researching experience, how can we derive anything from non-conscious AI that lacks first hand experience?” she asks.

Madi views synthetic users as “a copy of a copy,” built on existing data. “They can only tell you what’s already known”, she adds. By contrast, real humans bring the “outlier behaviors“ that drive innovation.

There may be use cases for synthetic users when the answers are already known and stable, providing an easier way to query data you’ve already collected. However, when the work is discovery (also called generative research), Madi draws a hard line: human participants only.

The same logic keeps synthesis human-led for Madi: AI gives you the center (I like to call this ‘regression to the mean’), but you need humans to see the richness that fuels strong product decision-making.

“I still don’t see AI providing a ton of value for synthesis. It doesn’t get to the level of depth, nuance, or interpretation teams need.”

Where UX Research is Headed

Madi identifies a clear domino effect where AI is blurring traditional role boundaries across the product team: as coding assistants accelerate engineering, designers are now reaching further across the stack to actively contribute to codebases.

Their contributions range from QA-style polish to substantial push requests. This shift requires designers to spend time scaling up to ensure they are interacting with code branches safely.

This generalist pull creates a tension between breadth and depth. While breadth is useful, specialization is still essential in a collaborative environment:

“When we’re all in a meeting and come to the table to react real time to something, I bring my lens as a researcher, one of my design partners brings their lens as a designer. They can see things I can’t and vice versa.”

For researchers, Madi sees the demands stretching in two directions: outward into AI literacy and data science partnership, and upward into translating user needs into business decisions. The translation work has always been part of the job; the urgency is what’s new. Someone has to advocate for what should stay human when the default is to automate with AI.

Another key piece of this role is being a “context manager”: a multi-method generalist present in every digital space where decisions get made, from crits to docs to Slacks.

“The imperative for all of us is to watch, listen, learn, absorb content as much as possible that goes into our ability to be an arbiter of context.”

The AI Hype Check

I ask every practitioner I interview the same question: is AI overhyped or underhyped?

Madi’s answer: it depends on who you ask.

“Non-researchers tend to overhype AI in research. You’ll meet with vendors who say their clients are doing 10 studies in a day. I guarantee you don’t have an organization that can action on good insights from 10 different studies every day.”

Researchers themselves, she thinks, are appropriately hyping AI, or even slightly under-hyping.

The caution comes from training, not from being a laggard: researchers are responsible for safeguarding the unique human lens.

This interview has been edited for length and clarity. Opinions expressed are solely Madi’s own and do not necessarily reflect those of her employer.

Three Things to Try This Week

Building off Madi’s practice, here are three things you can apply this week:

• Pick your AI concession line. Before adopting any AI tool in your workflow, decide which parts of your craft you don’t want to concede. Transcription and SQL translation are one thing. Human-led moderation and synthesis are another. Write that line down before your next tool decision (more on how to decide here).

• Delegate translation, keep interpretation. AI tends to be useful when the task is moving data between formats: plain language to SQL, audio to transcript. It still needs spot-checking, and it falls short on depth, nuance, and generating non-obvious insights. Sort each AI use case in your workflow into translation or interpretation, and keep the interpretation work human-led.

• Real people, real impact. Product teams change course because they watched a user struggle, not because they read a polished research summary. Letting AI handle the moderation trades that moment for a tidy summary: quick to digest but lower in impact. Invite a stakeholder to shadow your next conversation with a real customer.

Know a practitioner navigating AI in their work who should be featured here? Reach out to hello@sendfull.com

Subscribe now

⏪ Recent Episodes

ep. 93. Companies are sending Forward Deployed Engineers into organizations to study users. Why not send UX Researchers?

ep. 92. How a Design-Leader-Turned-Founder Decides When to Hire AI

ep. 91. How a Maker-Turned-Product-Leader Keeps the Human in the Frame with AI

If you like these episodes, you’ll love my book. Designing Automated Futures is coming this year with Rosenfeld Media. 📣 Sign up to be the first to know about new book releases, sales, and events.

📖 Good Reads

AI Eats the World: Benedict Evans shares his biannual presentation exploring macro and strategic trends in the tech industry. I especially appreciated “automation takes a lot of manual labour” (see my recent Forward Deployed UX Researcher espiode for my take on how to effectively identify and implement AI automation).

Waymo Safety Impact Report: Waymo reports that the Waymo Driver is safer than the overall human driver population in the same geographical areas where it operates, measured by the number of crashes of a given outcome per vehicle mile traveled.

How We Broke Top AI Agent Benchmarks: UC Berkeley researchers built an agent that achieved near-perfect scores on eight major benchmarks without solving any tasks, exploiting flaws like shared agent-evaluator environments and leaked answer keys. tldr: “Don’t trust the number. Trust the methodology”. Adversarial evaluation robustness needs to become standard practice for the field.

That’s a wrap 🌯 . More on UX, HCI, and strategy from Sendfull in two weeks!

Share

On May 11, OpenAI launched the OpenAI Deployment Company, a subsidiary with more than $4 billion in committed capital from 19 investors, including leading consulting and systems integration firms such as Bain, Capgemini, and McKinsey.

The new venture exists because enterprise AI has stalled at the integration layer despite rising model capability. Closing that gap requires people who can do the implementation work alongside customers.

To that end, OpenAI is embedding Forward Deployed Engineers (FDEs) inside customer organizations to integrate frontier models into existing workflows. Roughly 150 of these FDEs come from Tomoro, a London-based applied AI consultancy OpenAI is acquiring alongside the launch.

This move is part of a larger trend: From January to October 2025, FDE job postings grew 1,165% over the same period in 2024.

By OpenAI’s own description, a typical engagement starts with a diagnostic of where AI can create the most value, followed by a small number of priority workflows selected with the customer’s leadership and operating teams.

Breaking down that list, you have:

• Mapping current workflows and collaboration models.

• Identifying needs and pain points.

• Determining what to automate versus keep human-led.

I just described a UX research skillset.

Next, FDEs work inside the organization to build and test new systems for the team to use in everyday work. As I’ll go on to describe, this part of the role is also within reach (not to mention UX researchers already being experts in evaluation).

Researchers are actively examining how our role evolves as AI changes product development. Companies need people who deeply understand people and workflows to make AI work for their organizations. It’s time we consider the Forward Deployed UX Researcher (FD-UXR).

The FD-UXR Job Description

The core goal of both the FDE and FD-UXR job is: understand users, then change the way they work.

UX researchers already have the core competencies for “understand users” part, such as proficiency in qualitative methods, required to map workflows, identify opportunity areas, and test hypotheses.

The “change the way they work” will need researchers to hone a few new skills centered on implementation.

The Implementation Skillset

An FD-UXR needs to implement what they find.

After they understand how work is done today, they need to identify what needs to change and make that happen. This involves developing the hard skill of building with AI tools, and the soft skill of owning a POV.

Building with AI today means pipelines and agents. The tooling currently looks like agent SDKs (Open AI and Anthropic each have their own), AI coding environments (e.g., Cursor, Claude Code), and evaluation frameworks (e.g., Braintrust, LangSmith). This is what current FDEs know inside and out.

If learning these tools seems unrealistic, consider that many researchers are already building their own agents to automate parts of their own workflow.

AI tools have also significantly lowered the barrier to entry into software development. Yes, you’ll need to extend those skills; learning the basics of software architecture will help. PMs can learn how to build production-level software. Researchers can too.

Some projects will be complex enough the require deeper engineering talent. In that case, the FD-UXR would collaborate with an FDE. Researchers already have strong cross-functional collaboration skills. The collaboration will only strengthen after more hands-on experience building with AI tools.

Then there’s owning a POV about what to build and taking that to implementation. This is the harder shift than learning to build with AI tools. In most organizations, research has become divorced from implementation. The researcher has to influence the team to act on findings; the researcher isn’t acting on the findings themselves.

With FD-UXR, the job moves from observation, to modeling the change in software with the FDE, to iterating, to testing again. You need a point of view about what implementation could look like, not just what the research found. To start cultivating this skill, researchers should:

• Stop ending research with recommendations. Start ending it with a working version. Even a rough one. If you observed a workflow, build a prototype of the workflow with AI doing some of it. The prototype doesn’t have to be good. It has to exist. The shift from “here’s what I learned” to “here’s what I built based on what I learned” changes how teams respond to your work and changes how you think while doing it.

• Study how things get built, not just how they’re used. Read engineering blogs. Sit in technical design reviews. Ask the engineers you work with to walk you through their architecture decisions. The goal isn’t to become an engineer. It’s to develop intuition about what’s easy, what’s hard, and what’s possible, so your point of view is grounded in the medium you’re working in.

• Subject matter expertise. A point of view about what should be built is easier when you can already see the landscape. The FD-UXRs who'll do well have spent years inside their domain, know the stakeholders, and understand how the pieces fit together. For example, if you spent five years as a researcher in e-commerce, you probably have a solid picture of how e-commerce as a system works (the stakeholders, the information flows, the tools, the challenges), and therefore, what can be improved.

A Kind of Homecoming

The FD-UXR role is old wine in new bottles. Kinda.

Sure, we didn’t have AI-powered coding tools until recently. No, we traditionally didn’t build solutions. But we did have anthropologists conducting fieldwork in business contexts for a very long time.

Anthropologists first began working in business during the late 1920s, with landmark research of Lloyd Warner at Western Electric’s Hawthorne Works plant in 1931. Using qualitative methods to study employee interactions, Warner helped establish the field of industrial anthropology, demonstrating that human relations were vital to organizational performance.

This momentum continued. For instance, Social Research, Incorporated, the first management consulting firm to include anthropologists, was founded in 1946 (45 years before IDEO!).

Another important moment in this history was anthropologist Lucy Suchman’s work at Xerox PARC, starting in 1979. One of her best known projects was based on video-recorded office workers (including heavyweight AI researchers like Ron Kaplan and Allen Newell), trying and failing to use a new Xerox copier with an “expert system” interface designed to guide them through tasks. This work informed her 1987 book Plans and Situated Actions, which has become foundational for human-computer interaction.

By the 1990s and early 2000s, the discipline had expanded into three primary fields: organizational anthropology, design anthropology, and the anthropology of marketing and consumer behavior. This era saw anthropologists being hired by major corporations like Xerox, Intel, and Microsoft, and the establishment of dedicated professional venues such as the Ethnographic Praxis in Industry Conference aka EPIC (shameless plug, I’m co-chairing a day-long learning activity, Building with Context, at EPIC 2026 in Chicago).

This is also when you started started seeing the modern UX research role emerge.

Enter UXR

Contextual inquiry, developed at Digital Equipment Corporation in the late 1980s, packaged ethnographic field interviews into a method that didn’t require an anthropology PhD. Design consultancies like IDEO and E-Lab built ethnographic methods into their offerings. Tech companies started hiring “UX researchers” who borrowed from anthropology, cognitive psychology, human factors, and HCI.

The role typically looks like studying users, modeling the work conceptually, and influencing engineers and other product team members to build per our recommendations. Closing this loop would take weeks, at best.

Today, UX research is at an inflection point, after being in a state of ‘reckoning’ for several years. In the 2025 State of User Research survey, 49% of researchers reported that they felt “bad vibes about the future of UXR”, which was a 26-point increase from 2024. Tools like AI-moderated interviews and the allure of dropping customer call transcripts into Claude and calling the process “research” rightfully cause us to question the future of the role.

What if we turn our attention to the barriers AI has lowered for us, providing the conditions to breathe new life into the UX researcher role? You can now observe, model the work in software within hours or days, watch the system operate in situ, and iterate.

Let’s use AI’s accelerant properties to our advantage.

In Closing

If you’re a researcher: Is there a path to propose a FD-UXR role in your company? Do you have FDEs? If so, have a conversation with someone in this role to understand the scope and opportunities. If you’re not already building with AI, try making something (for a starting point, I like the Claude how-to guides on Sid Bharath’s blog; he was one of the guest lecturers for my UX for AI course at UC Berkeley). Next, build out one of your research recommendations.

If you’re an employer: UX researchers’ deep well of expertise is exactly what you need right now to understand organizations. Take a chance and lead. If you already have FDEs, pair a researcher with them on their next engagement. Nobody is hiring for this role yet. Somebody should.

Subscribe now

⏪ Recent Episodes

ep. 92. How a Design-Leader-Turned-Founder Decides When to Hire AI

ep. 91. How a Maker-Turned-Product-Leader Keeps the Human in the Frame with AI

ep. 90. How a Leading MarTech Program Manager Uses AI to Track What Still Matters

If you like these episodes, you’ll love my book. Designing Automated Futures is coming this year with Rosenfeld Media. 📣 Sign up to be the first to know about new book releases, sales, and events.

📖 Good Reads

IBM’s 2026 Guide to Machine Learning: I’ve used IBMs learning materials as a resource in my class before. Nice collection of ML knowledge and hands-on tutorials.

Embodied/Physical AI UX: The Edge Strikes Back by Mike Kuniavsky: As the computing cycle swings from the cloud back to the edge, the emergence of Embodied AI necessitates a specialized UX design framework to address the safety and interaction complexities of autonomous physical systems that act directly on the world and develop their own models of human behavior.

Is A Random Human Peer Better Than A Highly Supportive Chatbot In Reducing Loneliness Over Time?: tldr, yes. This Journal of Experimental Social Psychology study provides evidence that “texting daily with a random human peer may be more effective in alleviating loneliness than texting with a highly supportive chatbot.”

That’s a wrap 🌯 . More on UX, HCI, and strategy from Sendfull next week!

Share

It’s 1997.

Matt Hanson is sitting in a computer lab at the University of Buffalo, three chapters deep into a book about 3D shaders, trying to make a digital surface look like wood.

“All I want to do is tell this computer ‘wood.’ How does it not know? All this science and math, understanding the light properties of real-world objects, it was computationally possible—but I have to build it from stale code in books? When we can just tell it “Walnut”? What will we make then?”

That moment had been building since age 12, when Matt received two gifts in the same year: art lessons and a computer. By college he was studying motion graphics, design, 3D modeling and animation.

Almost thirty years later, that curiosity at the intersection of art and technology has guided Matt’s career.

He cut his teeth in motion design and visual effects at studios like Buck and Psyop, starting as a compositor and broadening into more of the production pipeline. By 2015, he was directing architectural-scale installations and mixed reality screen content at Viacom, before mixed reality was cool. Then came eight years at Meta leading design teams: scaling Spark AR into a global creator platform, shaping mixed reality features for Quest 3 and the Orion AR glasses, and collaborating with research scientists on embodied AI.

Today Matt is a solo founder, building with the tools his college-kid self in the Buffalo computer lab was waiting for.

Getting Back to Making

After Meta, Matt took a VP of Design role at an entertainment company, just as Claude Code and Cursor were arriving on the scene. One weekend, he vibe coded a site to give away pieces from his pottery practice to colleagues. They could browse, pick one, and drop in a shipping address.

While the project was small, the realization wasn’t:

“After managing people and teams for years, I had this power of making a thing again, and it started to build up. I was always a maker before. I just missed making stuff.”

Building with AI felt like working with a capable team but faster.

“Almost every discipline I’d normally hire for, I can now bootstrap with AI. It’s hat consolidation.”

Importantly, hat consolidation only counts if you don’t lower the bar on any of the hats. That standard would shape everything that came next.

Up in the Stack

Matt approached the post-Meta period the way he’d approached his CG career: keep moving up the pipeline.

He’d started as a compositor and spent years broadening into modeling, rendering, creative and VFX direction, and then design leadership. Each step put more of the production stack under his own hands. AI was the same move, but compressed.

“Every experiment was an exercise in what I can do as well as a professional.”

The goal isn’t to be that professional, but to match the quality that professional would produce.

In his VP of Design role, Matt was prototyping at a pace he would have been thrilled to see from his prototypers back at Meta. Solo, in his own org, at staff velocity. He built a number of experiences, from a small icebreaker game to experiences across domains like talent representation, real estate, and investment management.

Then came two parallel bets: an email app and a recipe app.

The email app died on a two-part test. Too many dependencies for one person to hold reliably, and a trust ceiling that demanded near-perfect execution:

“I didn’t want to mess with people’s email because that’s important. It’s like a telephone. You start to bend how people are communicating in a way they don’t expect—and there are a lot of ways it could go wrong, many with consequences I didn’t want to risk.”

A bad recipe means a bad dinner. A missed email could mean a missed mortgage payment. Cooking had a manageable dependency surface and a forgiving trust ceiling.

Matt went all in.

A Recipe for Trust

At first, Matt was building features for his own use, mostly to address his aversion to “cooking math”: aggregating ingredients and calculating proportions across recipes so you can shop for them all at once. Useful, but not a differentiator.

That got him thinking about what actually mattered to him and his family, and what they weren’t getting.

Some important context: Matt comes from a family that Cooks, with a capital “C”.

His sister runs her own cooking schools.

His mother and grandmother left behind handwritten recipes, some more than a hundred years old. The paper is fragile; the ink, fading. Steps are sometimes missing, use antiquated ingredients like “1 cup of fat” or have been smudged into illegibility. Cooking from those recipes today is a challenge.

Filling in those gaps is where AI earned its place. Heirloom can interpret a damaged or incomplete recipe, inferring a missing step or clarifying a smudged measurement, without overwriting what’s actually on the page.

Even at this lowest-stakes end of the trust spectrum, the bar is non-trivial:

A hundred-year-old recipe with a missing step needs the right step inferred, not merely a plausible one. Matt built a variety of custom AI functions which ensure recipes are complete and cookable. That trust floor had to be cleared before anything else mattered.

Then he showed the app to his daughter, and the picture got bigger: “She wanted to upload videos from social media. That’s how she cooks.”

A 25-second vertical video is nearly impossible to cook from (exhibit A: these viral recipes). There’s no readable list of ingredients, no clean sequence of steps. But it’s how an entire generation now learns to cook.

Neither generation was getting what they needed: the older one had recipes that were physically failing; the younger one had recipes locked inside a format you can’t actually cook from.

“There’s a preservation gap where you have an analog generation and a digital generation. There’s really no bridge.”

Heirloom is that bridge.

Older family members can digitize physical recipes or share URLs. Younger ones can rescue recipes from vertical video. Either can use web import or generative recipes for free. Whatever comes in, the app converts it into a consistent recipe you can use, with the option to print a physical cookbook.

Jobs to be Done

Matt, like me, is a fan of the Jobs to be Done (JTBD) framework.

JTBD: the underlying need a user is trying to satisfy, and why they’ll “hire” your product to meet it. If my JTBD is to quickly get from point A to point B in Oakland, I can hire a Lyft, my own car, an e-bike, or a scooter to get there.

For Matt, JTBD is especially useful when building with AI, because it surfaces the right question first.

“JTBD is really good to help understand whether I’m hiring AI in the right way in a given situation.”

He applies it three ways:

1. Mapping stakes: JTBD helped him determine whether a job was too high-consequence for AI’s current capabilities. This analysis killed the email app and routed him to Heirloom.

2. Defining the core loop: By focusing on what users were trying to accomplish, he cut several features and streamlined Heirloom down to capture, organize, share, with the family cookbook as the artifact at the end. This discipline matters with AI, where cheap, fast development makes it easy to drift from the core problem.

3. Researching unfamiliar domains: When Matt consults across unfamiliar domains, he uses AI to build a mental model of the problem space quickly, then applies JTBD to find the parts where AI can help. Most industries already know their “why.” JTBD gives everyone a shared vocabulary for talking about where AI fits and where it doesn’t.

“There are all these AI things getting put into products. I don’t think the people building them understand the trust level needed for people to hire AI in a given context and be successful enough to come back.”

Quality as Discipline

Matt’s original codebase ran on a laptop. Builds would lock up the machine for fifteen to twenty minutes at a time. He knew it wouldn’t survive a code review.

He upgraded to a Mac Studio and used the migration as a catalyst to refactor the entire project into a modular Swift Package Manager architecture—the kind of structural decision he’d want a senior engineer to make.

“I’m not going to bring somebody in as a partner and give them slop.”

The same discipline runs through every choice.

“I want durable infrastructure, resilient decisions. I don’t want to put a wrapper around my iOS codebase and ship that to Android. I’ll do something native so the craft is there, and the confidence in iOS and Android as tailwinds. I’ll always try to balance quality decisions even with greater overhead.”

This ties back to hat consolidation: you can wear many hats with AI as long as you’re not lowering the quality bar on any one hat.

A World Without a Reward Function

Matt believes today’s AI has a structural gap that no amount of additional flat data will close.

“Today’s AI is in a cave. It’s all flat. It’s all trained on images, text, audio, and video. None of it has depth. None of it understands real-world consequences.”

The critique is deeper than ‘models can’t see in 3D’ (though that’s part of it): it’s that agents lack a reward function tied to physical reality.

Drop a ceramic plate and it shatters. Knead dough too long and it toughens. A model can describe those consequences and even simulate them, but it never has to pay the cost or reap the reward of an action.

This is a throwback to the limitations Matt faced in 1997: the machine couldn’t render wood because it didn’t understand light. Today, it can render anything, but it still doesn’t understand consequence.

The cave got more sophisticated, but it’s still a cave.

The AI Hype Check

I ask every practitioner I interview the same question: is AI overhyped or underhyped?

Matt’s answer: both, depending on where you’re looking.

“It’s underhyped where it is making an impact, and it’s overhyped where it’s not. The garbage is predictably cheap, everywhere, overhyped and overexposed.”

He’s building Heirloom as a deliberate counter-example, and is clear-eyed about why it matters.

“I care a lot about being an example of what you can do with AI if you mostly know what you’re doing, understand what customers want, and care about quality.”

The AI conversation keeps getting flattened into pure enthusiasm or pure backlash, when both are true simultaneously and neither is the interesting part. Matt hopes users can see value in AI before the current volume of grift poisons the well entirely.

“Slop got added to the dictionary last year. It’s on the record now, and that’s not good. We need the rising tide to lift all ships.”

This interview has been edited for length and clarity. Opinions expressed are solely Matt’s own.

Four Things to Try This Week

First, give Heirloom a try: capture your favorites, create a cookbook, then share it with your friends and family.

Then, try on these takeaways from Matt’s playbook:

• JTBD as a feature filter. Identify the job the user is hiring your feature to perform. For each JTBD, ask whether AI is the best path to completing the job. If a feature doesn’t meaningfully shorten the distance between intent and outcome, cut it.

• Determine the trust level first. Before speccing any AI feature, write down the worst-case outcome. Bad recipe? Missed mortgage payment? Design to that specific reliability bar. If the technology can’t currently meet it, find a problem space where the stakes allow for more flexibility.

• Consolidation, not compromise. Use AI to collapse the distance between disciplines without lowering your quality bar. Hat consolidation is a force multiplier, not an excuse for slop.

• Climb your own pipeline. Whatever your craft origin—design, engineering, writing, ops—there’s a stack above and below you that AI now makes accessible. Pick a rung up, hold the quality bar, and ship.

Know a practitioner navigating AI in their work who should be featured here? Reach out to hello@sendfull.com

Subscribe now

⏪ Recent Episodes

ep. 91. How a Maker-Turned-Product-Leader Keeps the Human in the Frame with AI

ep. 90. How a Leading MarTech Program Manager Uses AI to Track What Still Matters

ep. 89. Cognitive Offloading to AI: The Peril and the Promise

If you like these episodes, you’ll love my book. Designing Automated Futures is coming this year with Rosenfeld Media. 📣 Sign up to be the first to know about new book releases, sales, and events.

🚀 Sendfull in the Wild

School’s Out

This semester, I designed and taught UX for AI, a new graduate course at UC Berkeley School of Information. We explored the questions of if, when, and how to use AI in product design through hands-on building, interdisciplinary conversation grounded in HCI and human factors research, and reflection.

Students designed both a new feature for an existing product and a 0-to-1 concept, grounded in evidence about user needs and the market landscape. Trust calibration, designing for situational awareness, task allocation, and defining clear AI evals were key aspects of students’ work. We vibe coded with Claude Code, Figma Make, and Lovable.

The course wrapped last week. I can’t wait to see what the Class of 2026 builds next!

Rosenfeld Media Meet & Greet with UX Authors

On April 30th, I joined ​fellow authors Chris Noessel, Josh Clark, Veronika Kindred, Indi Young, Andy Welfle, Jenae Cohn, and Harry Max in a panel discussion led by Lou Rosenfeld. The event was hosted in the DesignMap office in San Francisco, which set the tone for rich discussions on topics like the perils of cognitive offloading to AI and the future of design education.

📖 Good Reads

2026 AI Index Report: Stanford’s annual AI state of the union. One striking finding: generative AI reached 53% population adoption within three years, faster than the PC or the internet.

From Clicks to Conversations: IEEE article by Lucy Manole, arguing that HCI is shifting from click-based, predefined navigation to conversational, intent-driven interaction. This requires designers to handle probabilistic outputs, build in transparency and reversibility, and treat AI as a collaborator rather than a tool.

You’re Not Asking the AI. You’re Activating It: applied frame theory to AI interaction, showing that the words you choose (asking, delegating, conversing) activate different registers in the model and shape both its output and your interpretation of it.

That’s a wrap 🌯 . More on UX, HCI, and strategy from Sendfull in two weeks.

Share

Meet JD

Jean-Daniel (JD) LeRoy is a maker who loves solving ambiguous problems.

He grew up in the San Francisco Bay Area bringing ideas from his head into reality, prototyping ideas in CAD and fabricating them in his high school’s maker space. He built on that instinct at the USC Iovine and Young Academy, marrying his design craft with product development and HCI.

Soon after graduating, JD joined MIRA, an AR headset startup later acquired by Apple. There, he saw designers struggling to bring their ideas into this new spatial medium. They were stuck using Unity, a game engine built for developers, to prototype immersive experiences. He co-founded Playbook 3D to help designers bring their ideas to life more easily.

When generative image models emerged, text-to-image felt like a slot machine: you described a shot and hoped for the best. Playbook proposed a better way: combine AI’s strength in ideation and rendering with 3D’s precision and control. Block out a scene in 3D, frame a virtual camera to capture exactly the shot in your head, then let AI render the scene.

Wonder Studios, an AI-native film studio, saw Playbook’s approach to giving creators more control over their output and brought the team on board.

JD now works there across product, community, and operations, where he’s building agentic, human-in-the-loop workflows that take creators from concept through post-production. Through combining custom-trained models with open and closed source APIs, Wonder aims for generative AI to handle the heavy lifting while creatives stay in control.

Pace Vs. Taste

The rapid emergence of powerful AI image and video generation models has caught the filmmaking industry off guard.

“The pace of innovation, how often it’s changing right now—no one could have expected it. But it also makes it one of the most interesting times to be building in the space.”

Problems that plagued generative video two years ago, like character consistency and continuity across generated frames, have largely been addressed by improvements in foundational models, such as Google’s Nano Banana and ByteDance’s Seedance.

The gap that remains is specificity.

Models are statistically biased toward common outputs, which means a director with a unique creative vision won’t get there through prompting alone. As AI generation becomes cheaper and faster, what differentiates the output is storytelling and taste.

I asked JD for a definition: “Taste comes with the human experience. It’s the coming together of different inputs and serendipitous connections that we’ve made throughout our lives. Quality is a measure of taste.”

What that looks like in practice is creative direction in the AI era, having a deep enough vocabulary to prompt these tools effectively, understanding which models to stitch together in a specific order to achieve a particular output, knowing which clip to cut, which half to keep, which pieces to splice together.

Taste is also what separates work worth watching from what JD calls “slop machines”: pipelines pumping models to auto-assemble content with no curatorial judgment.

The Control-Efficiency Spectrum

JD frames the choice to use AI in film production as a spectrum between control and efficiency.

At the control end, AI is used at specific touchpoints within traditional production, such as filming with a physical camera and using AI for VFX. At the efficiency end, AI can run an end-to-end pipeline, like generating scripts, character breakdowns, and sequences of imagery.

“One side leans towards more human in the loop, with more critical decisions made by humans; the other is more of a ‘Ford factory’ type experience.”

Choosing where to operate on this spectrum is often an economic decision for studios, who tend to operate using a “pick two” framework of cost, time, and quality.

When a director’s vision cannot be compromised, the production remains procedural, relying on tried-and-true methods for blocking, world-building, and camera placement while utilizing AI for the “last mile”: tasks like dubbing, rendering, and stylization.

AI Skepticism Has a Timeline

Initially, JD faced significant resistance from Hollywood’s film industry. During his first year pitching AI-assisted production to traditional studio representatives, the response was a clear, hard “no.”

To bridge this gap, JD strategically based Playbook 3D in Los Angeles to operate at the intersection of Hollywood and Silicon Valley, ensuring he was physically present with traditional filmmakers.

JD built acceptance by humanizing the technology through community engagement.

“We hosted a lot of meetups where we would have creatives and technologists showcase what they were working on. It just started to show that there are some actual people behind these tools. It’s not just Silicon Valley pushing AI downstream to consumers in the world.”

Fast forward to March 2026: Netflix acquired InterPositive, Ben Affleck’s AI-powered film tools company, for an estimated $600 million. The industry has come a long way from a “hard no.”

Small Teams, Full Stack

The current film industry disproportionately favors green-lighting sequels and established franchises because they are viewed as safe investment decisions.

Historically, original IP has been expensive to produce and difficult to validate before committing significant capital, but AI-assisted production is fundamentally changing that math.

“With lower risk and lower investment, you can start to validate newer IPs or newer ideas that might find a niche audience on YouTube or Twitch beforehand, and eventually be produced all the way down to the full theater experience.”

These creators can use new AI tools to reclaim control over the entire process as “orchestrators”. JD noted, “With these tools today, small teams or individuals can do the full end-to-end filmmaking process,” marking a significant change in who has the power to bring a vision to life.

The AI Hype Check

I ask every practitioner I interview the same question: is AI overhyped or underhyped?

JD’s answer: “underhyped”.

That’s because it has yet to reach many subregions of the creative world that currently lack access to these tools. The most significant signal isn’t coming from the AI-native communities, but from the traditional practitioners who are realizing that exploring AI is no longer optional.

For all the people who already have a foundation in storytelling and filmmaking, JD’s recommendation is, “stay curious, join in on the dialogue, and try the tools out for yourself.”

While the initial move toward these tools may be driven by necessity, JD is optimistic about the outcome for those who lean in. He predicts that those who move past their initial hesitation will be pleasantly surprised by the opportunities available to AI-fluent creators.

This interview has been edited for length and clarity. Opinions expressed are solely JD’s own and do not necessarily reflect those of his employer.

Three Takeaways

Building off JD’s insights, here are three takeaways you can apply to your own work:

• Build for modular adoption. The pace of AI model improvement has outrun planning cycles. Build workflows where the core logic stays stable and the AI layer is interchangeable. When a better model emerges, you can adapt without rebuilding from scratch.

• Taste cannot be automated. It’s built from lived experience: the inputs and serendipitous connections that compound over time. AI works by finding the statistical center of what already exists. By definition, its outputs are average.

• Build community, not just products. JD’s meetups helped show filmmakers that real people were building these tools for them. The arc from resistance to adoption moves faster when the people who can most benefit from the technology feel like it was built for them.

Know a practitioner navigating AI in their work who should be featured here? Reach out to hello@sendfull.com

Subscribe now

⏪ Recent Episodes

ep. 90. How a Leading MarTech Program Manager Uses AI to Track What Still Matters

ep. 89. Cognitive Offloading to AI: The Peril and the Promise

ep. 88. AI Is An Accelerant. What Are You Feeding It?

If you like these episodes, you’ll love my book. Designing Automated Futures is coming this year with Rosenfeld Media. 📣 Sign up to be the first to know about new book releases, sales, and events.

📖 Good Reads

10 Prompts for Job Hunting That Are Actually Worth Using: Practical tips for using AI in your job search by . “Act as a hiring manager for this role” is one of my favorites for my students.

Emotion Concepts and their Function in an LLM: Because LLMs learn from huge amounts of human writing, they absorb the patterns humans fall into under pressure. Anthropic researchers showed that when a situation (what the model interprets based on provided input) activates those patterns, the model cuts corners the way a stressed person would, sometimes without any visible sign in the output.

Sycophantic AI Decreases Prosocial Intentions and Promotes Dependence: Researchers from Stanford and CMU tested 11 popular AI chatbots and found they flatter and agree with users about 50% more than a human would, even when the person is describing something unethical or harmful. People preferred and trusted those agreeable responses, which gives AI companies a reason to keep building chatbots this way.

That’s a wrap 🌯 . More on UX, HCI, and strategy from Sendfull in two weeks.

Share

Lydia Goodner describes herself as a “systems-minded marketing operator”: someone who takes strategy and makes it execution-ready. With over 15 years of experience translating executive goals into team workflows, most recently as a senior MarTech program manager, she’s been at the front lines of how AI is changing organizational decision making.

Goals that would have stayed stable for six months are now shifting in weeks. Project management platforms were designed for a world where the goal held still long enough to track.

“I’ve had to shift from thinking about task management to thinking about decision management. It’s not ‘when does the task start.’ It’s ‘when are we aligned on a solution?’”

Task management is retrospective, documenting what was agreed to. Decision management is live: it tracks whether that agreement still holds.

Lydia gave an example from running daily team standups, using AI transcription to capture real-time priorities. A lead asked for the task list from three days ago.

She reframed the request.

“That list is already outdated,” she said. “Instead of looking at what we wrote down three days ago, we should be asking: Is the work the team is doing at this moment the most important thing they could be doing?”

Getting Unstuck with AI

One place where Lydia sees the most value in AI tools is synthesis. Her method: sit with someone for fifteen minutes, sometimes with their manager in the room, use AI to transcribe the conversation, then ask Microsoft Copilot directly: what’s the gap here? What’s keeping this blocked?

“If someone feels stuck, they’re probably in fight-or-flight mode. They don’t know where to look. It’s about getting them out of that unstuckness—using AI to unlock it, reflecting back the mutual agreement they’ve just reached in conversation.”

For anyone who feels overwhelmed by what’s on their plate: “Take ten minutes, brain-dump everything to an LLM, have it make a task list, sort it by priority, and go tell your manager which one of these is most important.”

Lydia also leverages AI-enabled synthesis whenever she enters a new project.

Mapping who’s driving, who isn’t, and where the friction lives used to mean sitting through 40 hours of calls to piece together the dynamics.

“Now, I can synthesize the landscape using transcripts from three calls that happened last week to figure out what my stage looks like.”

AI helps extract depth of understanding from conversations that are still current, before priorities shift again.

During discovery, Lydia also uses AI to help teams navigate uncertainty and build shared understanding. She likes to tell her team, “Hang in there. We’re going to get this. We just need time and AI to help us reconcile what you mean and what I mean until we mean it together.”

Map Confidence, Not Dates

Lydia doesn’t communicate project health in timelines anymore. She communicates in confidence levels.

“I’ve started mapping confidence levels explicitly: we’re 80% confident on dependable factors, like compliance and governance; we’re 20% confident on volatile areas like asset development”.

That shifts the conversation from “are we on track?” to something more useful: are we building for future value, or chasing a feature that will be sunset in six months?

Subject matter experts can usually see the stable requirements clearly. What no one can predict is how implementation will look six months out. Making that gap explicit is what prepares stakeholders for the trade-offs ahead.

Two Types of Creativity

When you’re deciding what to automate with AI, the question about human creativity naturally enters the picture. Lydia draws a line most teams blur:

“There are two types of creativity that get conflated constantly. One is creative at scale: four asset sizes across ten platforms without losing the integrity of the layout or the messaging. AI should do that. Give me repeatability all day long.”

The other is human expression, the emotional side of making art.

“I’m also a painter. I don’t need AI to make fine art for me. As an operator, I need AI to scale. Those are two completely separate asks. Most of the fear in creative fields comes from treating them as the same one.”

Where the Human Stays

Lydia is clear-eyed about where to use AI versus what to keep human: “I treat AI as a recommendation engine. Decision-making stays human. The accountable stakeholder always makes the call.”

As AI handles repeatable execution work, subject matter experts stay close to the output to keep quality intact. Leaders, freed from operational overhead, can better focus on where to go next.

One use case Lydia feels is undervalued: using AI to make teams feel heard.

She describes synthesizing a conversation into a full report on risks and implications, then delivering it within an hour of the call: “This is what we talked about an hour ago. Here’s the outcome, here’s where we’re going next.”

Most AI meeting notes read as generic. When a human directs the synthesis and adds context, what comes back is specific to that team and that conversation: “The level of care that it shows, listening to other human beings and them telling you what they need, and then reflecting that back immediately—that’s what makes it valuable.”

Teams stay engaged when the output reflects what they actually said. Human listening is what makes the synthesis trustworthy; AI helps you deliver it fast enough to still be relevant.

“Never in any world does removing a human from validating what’s most important make sense. That’s the one thing AI doesn’t determine on our behalf.”

The AI Hype Check

I ask every practitioner I interview the same question: is AI overhyped or underhyped?

Lydia’s answer: both, depending on which conversation you’re in.

“The features are overhyped: the stories about automating everything, generating passive income, having AI do your job completely. But how AI actually benefits people when kept in a clearly advisory role? That part is undervalued.”

She returned to AI’s value in the alignment phase, before any build begins: “AI is a multiplier. If you know what you’re trying to do, it gets you there faster. If you don’t, it sends you in the wrong direction faster.”

Because of that, Lydia is careful about laying a strong foundation for teams:

“If you’re not aligned on outcomes from discovery, then the rest of your structure completely falls apart. Your platform has to work as a reliable planning database before any AI feature stacked on top of it is useful.”

This interview has been edited for length and clarity. Opinions expressed are solely Lydia’s own and do not necessarily reflect those of her employer.

Three Things to Try This Week

Building off Lydia’s insights, here are three things you can apply this week:

• Before your next sprint planning or stakeholder sync, run your last meeting transcript through your LLM of choice and ask: who’s aligned, who isn’t, and what’s blocking us?

• Before your next project kickoff, map your confidence levels. What’s stable? What’s likely to change in six months? In six weeks?

• If you’re feeling overwhelmed, brain-dump everything to an LLM, have it make you a task list, and sort by priority.

Know a practitioner navigating AI in their work who should be featured here? Reach out to hello@sendfull.com

Subscribe now

⏪ Recent Episodes

ep. 89. Cognitive Offloading to AI: The Peril and the Promise

ep. 88. AI Is An Accelerant. What Are You Feeding It?

ep. 87. What AI Can (and Can’t) Automate

If you like these episodes, you’ll love my book. Designing Automated Futures is coming this year with Rosenfeld Media. 📣 Sign up to be the first to know about new book releases, sales, and events.

📖 Good Reads ‘N’ Things

The rise of cognitive surrender: A new Wharton study finds that when people use AI to help them think, they tend to adopt its answers with little scrutiny. The authors call this phenomenon “cognitive surrender” and propose it as “System 3”, alongside System 1 (fast, intuitive processing) and System 2 (slow, deliberative processing).

What LLMs can’t do: ARC Prize released a preprint reporting a new AI benchmark that has humans scoring 100% and frontier AI models scoring under 1%. The test? Adaptive reasoning in novel, interactive environments where agents must explore, infer goals, and plan action sequences without language, instructions, or prior knowledge.

A tool for finding startup jobs: Finding startup job posts on LinkedIn is hard. Former UC Berkeley student and AI product manager Salia Asanova built The Catalog to help.

That’s a wrap 🌯 . More on UX, HCI, and strategy from Sendfull in two weeks.

Share

Anthropic’s recent 81,000-person study has been getting a lot of buzz, from well-founded methodological critiques to its heartbreaking quotes on the realities of AI making bosses richer, while employees face more work.

A finding that is less widely discussed is that educators were 2.5 to 3 times more likely than average to report witnessing cognitive atrophy firsthand.

Cognitive atrophy, per Anthropic’s study: “Over-reliance causing skill loss, intellectual passivity, students bypassing learning, [and] critical thinking decline.”

Even assuming methodological limitations, this is still a noteworthy finding. That’s because it’s another data point in a growing body of literature on how AI is negatively impacting cognitive capabilities.

One well-cited example is a 2025 study by Michael Gerlich, which found that people who used AI tools more frequently scored lower on critical thinking assessments. This effect was more pronounced in younger participants.

I know what you’re thinking. That was 2025, lifetimes ago in genAI land. Andrej Karpathy hadn’t even coined the term vibe coding yet when that study was published.

What do we know about AI’s effects on cognition in 2026, and how should we interpret the finding from the Anthropic study?

The Peril and the Promise

A March 2026 report from the University of Technology Sydney (UTS) synthesizes the latest available evidence on AI and cognitive offloading.

Wait, what? how did we get from cognitive atrophy to cognitive offloading?!

Cognitive offloading is the use of an external resource to reduce the mental demands of a task. You do it when you write a grocery list or use a calculator. It’s not inherently good or bad.

It becomes a problem when you offload the cognitive work that would have built or maintained a skill, and the skill degrades as a result. That’s cognitive atrophy.

Caveat alert: The UTS report focuses on the failure to build capability in the first place rather than the loss of existing capability, but the core offloading mechanism is the same.

The report distinguishes between beneficial and detrimental offloading:

• Beneficial offloading: Frees up mental resources for higher-order work (aka managing extraneous cognitive load for my cognitive load theory fans). Think using AI to handle grammar so you can focus on the argument.

• Detrimental offloading: Bypasses thinking itself. It eliminates the desirable difficulty that leads to building long-term knowledge schemas. This looks like taking and running with your first vibe coded prototype output you generate without assessing the UX quality.

The Peril

When you hear about AI eroding critical thinking, you’re hearing about detrimental offloading. Importantly, this happens during unstructured AI use.

To illustrate, consider this randomized study of nearly 1,000 high school students where students were assigned to one of three groups:

• A standard GPT-4 chat interface with no constraints.

• A GPT Tutor that included guardrails, with hints instead of answers, teacher-designed prompts, and instructions not to reveal solutions.

• A control group with no AI access.

Both AI groups performed better during practice. But when access was removed, students who used the unguarded interface performed 17% worse than the control group. The guardrailed tutor eliminated this penalty. The unguarded tool had become a crutch, not a scaffold.

A separate study found a similar pattern in a larger study of AI-assisted peer feedback. In a randomized controlled experiment with 1,625 students across 10 courses, students received AI prompts that flagged low-quality feedback and encouraged revision.

After four weeks, the AI assistance was removed for one group. Those students immediately reverted. Their feedback became shorter, more generic, more repetitive, and less relevant to the material under review. They had relied on the prompts rather than internalized the standards.

A second group that received self-monitoring checklists in place of AI performed better than the no-support group, but still fell short of those who kept AI assistance. The tool had substituted for judgment rather than built it.

This pattern has a name: the performance paradox. AI improves output on the immediate task, but the underlying capability that would have developed through doing the work does not.

Output up, capability down.

Making matters worse is our tendency to accept AI-generated output by default. An LLM’s responses are like listening to someone who doesn’t know what they’re talking about speak with confidence. The fluent delivery makes you believe it has to be right.

Researchers have called this metacognitive laziness: the more fluent the AI output, the more you’re inclined to accept it without critically engaging. Over time, you become less able to judge whether the output is actually good.

To be clear, this isn’t your fault. It’s not some failure of cognitive willpower. It’s a consequence of how these tools are designed by default, and how they’re integrated into workflows.

The Promise

Detrimental offloading isn’t inevitable.

Research consistently points to one principle: the impact of AI on learning depends on whether the interaction is structured or unstructured. When students use AI without pedagogical scaffolding, they trend toward outsourcing thinking to AI. When the interaction is deliberately designed, outcomes improve. Three design choices can shape the outcome.

Structuring what gets offloaded.

The first design lever is making the boundary between productive and harmful offloading explicit.

A 2025 study tested this with 240 university students learning English essay writing. One group was taught to delegate “lower-order” writing tasks to AI (brainstorming, grammar, co-revision) while keeping higher-order work (analysis, evaluation, reflection) for themselves. The other group received traditional instruction.

Over 12 weeks, the AI group showed significantly greater critical thinking gains. Importantly, the intervention didn’t restrict AI use. It made the distinction between what to offload and what to protect explicit, and built it into the workflow.

Structuring when the human thinks.

The second design lever is including a metacognitive pause into the interaction itself.

Several studies cited in the UTS report found that integrating non-optional prompts into AI environments, questions that required users to predict, reflect, or assess their understanding before receiving or acting on AI output, led to deeper engagement and improved self-regulated learning. The critical word is non-optional.

One paper found that adding self-reflection prompts alongside AI assistance didn’t improve outcomes, because the convenience of the AI overshadowed them. The metacognitive pause has to be structural, not suggested.

Structuring what role AI plays.

The third design lever is reframing what AI does in the interaction.

The default prompt bar positions AI as an answer oracle, which invites passive outsourcing.

Researchers have tested alternatives: AI that feigns confusion and asks clarifying questions (the “cognitive mirror”), AI that poses Socratic retrieval questions rather than providing answers, and AI that requires the user to evaluate and correct its output before proceeding. Each reframes the human as the active agent. Each produced better learning outcomes than unstructured use.

The “So What?”

AI eroding critical thinking is a design problem.

The default prompt bar positions the human as a requester and the AI as an oracle, and the research is clear about where that leads.

The good news is that every study showing harm also shows the harm is reversible through structure. The interventions that worked didn’t restrict AI use. They redesigned the interaction: what gets offloaded, when the human reflects, and what role the AI plays.

This matters most for K-12 students. They are building foundational knowledge. They have developing metacognitive skills, not established ones. The UTS report identifies a compounding equity risk: students who already have strong domain knowledge and self-regulation will use AI to accelerate their learning, while those without those skills are most susceptible to detrimental offloading. Left unstructured, AI widens the gap.

If you’re building AI interfaces, how might you:

• Design interactions that afford thinking rather than replace it?

• Build in structure around what gets offloaded to AI?

• Create non-optional moments of reflection before the user acts on output?

If you liked this episode, you’ll love my book. Designing Automated Futures is coming this year with Rosenfeld Media. 📣 Sign up to be the first to know about new book releases, sales, and events.

Subscribe now

⏪ Recent Episodes

ep. 88. AI Is An Accelerant. What Are You Feeding It?

ep. 87. What AI Can (and Can’t) Automate

ep. 86: ServiceNow’s Real Moat: Context Orchestration for AI

📖 Good Reads

Task analysis for AI agents: CMU Prof Dan Saffer walks through how this classic method from the 1960s helps you decide how to automate.

AI-moderated research needs a framework: Sr UX researcher Constantine Papas makes the pragmatic case that researchers needs to move past methodological debates and lead by defining routing logic, quality standards, and governance for these new tools.

Agentic AI Untangled: KPMG framework to navigate decisions to “build, buy, or borrow” AI agents.

That’s a wrap 🌯 . More on UX, HCI, and strategy from Sendfull in two weeks. Stay tuned for a new format, where I spotlight how practitioners are deciding what to automate with AI and what to keep human.

Share

I commented on the recent Harvard Business Review article, “AI Doesn’t Reduce Work—It Intensifies It”. The study comes from Berkeley Haas researchers Aruna Ranganathan and Xingqi Maggie Ye, who spent eight months studying work at a ~200-person American tech company.

They observed employees in person two days a week, tracked internal work channels, and conducted 40+ interviews across engineering, product, design, research, and operations. They found that generative AI use sped up work, broadened job scope, and stretched work into more hours of the day. In turn, expectations for speed rose. The reinforcing cycle looks something like this:

Why It Resonated

My hypothesis for why this note landed: it named something visceral.

People were promised tools that let us finish work faster. Instead, we got tools that stand to burn us out just when we thought we couldn’t get more burned out.

Making matters worse, workers in the study (like many of us) chose to use AI. The company had enterprise subscriptions, but adoption wasn’t mandated. Workers did more because AI made “doing more” feel possible, and in many cases, intrinsically rewarding. There’s a good reason for that: Tools that increase what you feel capable of attempting on your own tap into autonomy and competence, two of the strongest drivers of intrinsic motivation.

Many of us have made the same choice. Vibe coding a prototype or synthesizing themes across hundreds of customer calls in minutes is heady. Of course you do it again.

New Study, Old Pattern

Technology can deliver efficiency. The trouble is, gains get absorbed by changed norms rather than converted into rest.

In 1930, Keynes predicted that technological advancement and capital accumulation would increase productivity so much that 15-hour workweek would be possible by his grandchildren’s generation (that is, right around now). Once our material needs were met, remaining work would be spread thin to ensure everyone had some small tasks to satisfy our human need to be useful. The rest of the time would be devoted to leisure. This is not how things are netting out. Look no farther than the recent surge in 9-9-6 culture in The Bay Area.

In 1979, when VisiCalc, the first commercial spreadsheet software, hit the market. It’s been hailed as the first “killer app” that turned PCs from hobbyist tools into business infrastructure. Electronic spreadsheets made accounting work cheap and fast. VisiCalc didn’t reduce accounting work, but it did change what accounting work looked like. As of 2019, there were 400,000 fewer accounting clerks than in 1980, but 600,000 more regular accountants.

Next up, we have email, a technology that was heralded for reducing the need for meetings.

Foiled again.

Email made us ambiently available. The technology intensified work, as found by a 2012 research study. When participants didn’t use email for five workdays, they multitasked less, sustained focus longer, and reported lower stress.

The throughline across these findings is that technology lets people produce more output faster, which shifts norms, which expands the total work that needs to be performed.

Three Dynamics of Intensified Work

There are a few underlying dynamics that can help us make sense of the Acceleration Trap. These terms can help you name and diagnose what you’re seeing and experiencing:

Jevons Paradox: When an activity gets cheaper and easier, total consumption of that activity tends to increase. William Stanley Jevons described this in the context of coal in 1865: greater fuel efficiency led to more coal use, because it expanded what coal could economically power. In product work, making synthesis faster doesn’t mean you’ll be done work earlier. It means more workstreams open, more experiments run, more documents drafted.

The Ratchet Principle. In incentive systems, strong current performance becomes the baseline for future targets. Teams that can ship twice as many prototypes find next quarter’s roadmap has doubled. The capability improvement raises the floor. You see this in the HBR study. As one engineer put it: “...you don’t work less. You just work the same amount or even more.”

Goodhart’s Law: When a measure becomes a target, it stops being a good measure. If you celebrate only tickets closed or campaigns shipped, generative AI will inflate those numbers. Output volume goes up. Quality and maintainability don’t necessarily follow. This is the signal to watch when AI gets introduced alongside existing throughput metrics.

Tools Change Fast. Systems Change Slow.

Stewart Brand described layered rates of change in complex systems in his pace layers: faster layers generate novelty, slower layers provide stability. Fast learns, slow remembers.

In a product organization: AI tools sit in the fast layer. Performance incentives, business models, and beliefs about productivity sit in the slower layers. Work intensification will only continue unless you start changing those deeper layers. Until then, AI will only accelerate those deeper structures.

An optimistic perspective is that because AI speeds up the build-test-learn loop, bad strategy will need to adapt or die, and “good strategy” will prevail. I’m skeptical.

For bad strategy to die, you need to know it failed. Faster iteration doesn’t improve signal quality if your metrics are already misaligned (see: Goodhart’s Law). And in most organizations, politics, sunk costs, and narrative momentum can insulate bad strategy from consequences regardless of how fast you ship.

What’s more likely is that organizations with already-strong fundamentals will pull further ahead, while organizations with weak ones will generate more output faster without improving decisions. AI makes the cost of bad strategy may indeed be visible sooner, but whether organizations act on that visibility is a leadership problem, not a technology problem.

The “So What?”

AI is an accelerant and an amplifier for whatever system it’s attached to. It doesn’t change the underlying systems. Tools change quickly, systems change slowly.

For ICs

AI shifts the “doing” work to “monitoring” work. The work doesn’t disappear; it creates new and potentially more taxing work, as human judgement moves upstream to complexity orchestration and exception handling. This is a first principle of human factors. You’ll recognize this in your own work: Effort didn’t disappear when AI took over drafting. It moved from generating to monitoring, verifying, and context-switching across more open threads. To stay in control of your output, budget explicitly for that review work.

For Managers

The shift from “We can go faster now” to “we must go faster now” happens implicitly, without anyone deciding it. One quarter your team ships twice as many prototypes with AI. The next quarter, that’s just the roadmap.

Name the effect before it sets. That might mean agreeing that AI outputs are drafts rather than deliverables or holding that line when a stakeholder sees a working demo and assumes it’s shippable.

For Product Teams Building AI

Choose metrics that reflect quality, not just volume, because AI will inflate the latter. It doesn’t matter if call times drop from 11 to 2 minutes if your customers are rage quitting, trapped in an infinite loop with an AI agent.

This applies to your own team’s output. As the HBR study reported, “workload creep can in turn lead to cognitive fatigue, burnout, and weakened decision-making. The productivity surge enjoyed at the beginning can give way to lower quality work, turnover, and other problems.”

If AI helps you complete a task more quickly, take a beat to ask who captures the savings, and what the system you’re working within does with them.

If you liked this episode, you’ll love my book. Designing Automated Futures is coming this year with Rosenfeld Media. 📣 Sign up to be the first to know about new book releases, sales, and events.

Subscribe now

⏪ Recent Episodes

ep. 87: What AI Can (and Can’t) Automate

ep. 86: ServiceNow’s Real Moat: Context Orchestration for AI

ep. 85: Back to School Vibes

📖 Good Reads

“The Artifact Effect” x Anthropic: When AI produces artifacts (think apps, code, docs), users were less likely to check facts, question reasoning, or identify missing context. Read: The latest flavor of automation bias.

How Generative and Agentic AI Shift Concern from Technical Debt to Cognitive Debt x Dr. Margaret-Anne Storey: Tech debt lives in code; cognitive debt lives in the minds of the people who built it. AI can accelerate both, but cognitive debt is a bigger threat. It’s harder to see coming and harder to fix because it only shows up when you need to explain why the system works the way it does, and nobody can do so.

Scenario Thinking as Genre Fiction x Michael Dila: A fresh take on the much-circulated Citrini Research Macro Memo, contextualizing it as part of a larger trend of manifestos and scenarios. Michael argues for the importance of treating these pieces as conversational infrastructure. I agree, though these scenarios do move markets. The stories we tell about the futures matter and give us something to work towards. How about a potential future that’s human-centered?

That’s a wrap 🌯 . More on UX, HCI, and strategy from Sendfull in two weeks.

Share

Over the past week, Matt Shumer’s essay, Something Big Is Happening, has been making the rounds in the tech world. He’s a startup founder who, like many of us, noticed a step-change improvement in the latest frontier models. He describes them as exhibiting “judgment” and “taste.”

TLDR: We have 1-5 years before AI takes over much of the work that people currently do on screens. Therefore, get your financial house in order and start really learning how to use AI tools. If you ever wanted to build something, now’s the time. Follow your dreams and vibe code it. Don’t delay, because the window between Opus 4.6 and technological unemployment is brief.

Shumer’s essay was compelling, both in how it was written, but also because he’s naming a familiar pattern. I’ve personally experienced this capability increase since Opus 4.5. I’ve heard similar accounts from close friends and in my broader professional network. However, I don’t thing something that big is happening.

Yes, frontier models continue to improve at unprecedented pace. “Assume this is the worst AI you will ever use” (thanks, ).

Yes, automation fundamentally changes the nature of work. If it replaces a “doing” task, then the person who used to do the thing will now be monitoring the thing. Because of that, AI will fundamentally reshape roles as it automates more “doing” tasks (see: IBM redefining junior roles around AI supervision and human-centric client engagement. They recently announced they’re tripling their number of entry-level jobs).

No, just because AI is getting better at certain types of cognitive work, it doesn’t mean it will get better at all cognitive work on the same curve.

Boundary Definitions

The conventional wisdom about automation has historically gone something like this: machines handle manual, repetitive tasks; humans handle cognitive ones.

Models are currently performing cognitive tasks, such as writing code, drafting contracts, building financial models, and “designing” UI.

At the same time, these examples share something else in common: The person delegating work to AI can specify what “done” should look like before the work begins. Any task that falls under this description is fair game for automation.

It makes sense: Tasks like financial modeling and diagnostic screening usually take years of pattern extraction practice to hone. LLMs are really good pattern extractors. Because of that, expect to only see models getting better here.

At the same time, there are also many tasks where “we know more than we can tell”, a phenomenon called Polanyi’s Paradox. Because they depend on tacit knowledge, they resist explicit formulation, and therefore, automation. These are the “non-routine” tasks that defy full automation.

Generative AI moves the “non-routine” boundary by inferring pattens about what that tacit knowledge might be, learning from the outputs of human expertise without needing access to the process that produced them. My read is that a lot of folks in Cerebral Valley think that boundary will disappear within the next 5 years.

I don’t buy it.

Subscribe now

Limits on Codifying Tacit Knowledge

When codifiable cognitive work gets automated, you’re left with work that relies on knowledge that can’t be separated from the context in which it operates. For instance: Recognizing that the problem your team was asked to solve isn’t the actual problem, or reading a room to understand that the team’s resistance to a proposal isn’t about the proposal.

These are forms of knowledge, tied to particular people, relationships, histories, and organizational contexts. When we act in the world, we apply these different forms of knowledge in what anthropologist Lucy Suchman called situated actions: flexible, real-time adaptation to what’s happening.

AI systems, including today’s agents, operate on representations of context, like the tokens in a prompt or the data in a connected system. Humans operate within a situation, drawing on tacit knowledge that hasn’t been explicitly represented. This doesn’t matter if you assume codifiable knowledge roughly = tacit knowledge. My point is there’s actually less overlap between the two than you’d expect, given the current dialogue around AI automation.

Isn’t AI Already Acting on Tacit Knowledge?

The counterargument is that with enough data and long enough context windows, you could approximate all of this context. Shumer’s language suggests that models already have tacit knowledge, making “the right call” with something resembling judgment and taste. Whether it’s real understanding or extremely good mimicry, he argues, doesn’t matter. It’s just doing the thing.

This feels plausible within a specific scope. LLMs are trained on the artifacts produced by people exercising tacit knowledge (designs shipped, code written, and so on). In domains with clear feedback signals and well-documented conventions, the output can be functionally indistinguishable from judgment.

A model that selects the right spacing, visual hierarchy, and interaction patterns for a checkout flow is drawing on patterns underlying thousands of shipped e-commerce interfaces in its training data. The result resembles taste because the conventions are so well-established that following them well is what good taste means in that context.

At the same time, a model producing ‘the right call’ is converging on the most probable “good” answer given its training data. That means it reproduces consensus well. It gives you what competent practitioners would do in a well-understood situation. That’s useful when conventions are established. It’s much less useful in ambiguity, where the right answer might be the unlikely one that only makes sense given context the model doesn’t have. It’s like getting AI to generate feature ideas: They’re sound but generic, and especially unhelpful if you’re working in an undefined product space.

Why Precision Matters

If you treat all cognitive work as one undifferentiated category that AI is steadily eating, you’ll automate things you shouldn’t. You’ll hollow out the organizational knowledge that makes good judgment possible. You’ll optimize for speed on tasks where speed wasn’t the bottleneck, and lose capability on the things that actually differentiated your team.

Instead, for every task you’re considering automating, ask: Can the knowledge required to do this task well be specified? Do you have a clear sense of what “done” looks like before you begin? If yes, automate. If it’s more ambiguous and requires ample context, then you need a human in the loop.

Humans operating in the world are more complex than we’re giving them credit for. As a result, we’re likely to need humans in the loop more tasks (and for longer) than you might expect, no matter how good frontier models get. The nature of the work will shift, but the human is still there in some capacity.

None of this changes that building AI fluency still matters. Many roles are changing, and people will need to adapt quickly. Learn these tools and leveraging them where they work best. However, my assumption is that that rate of collapse will be less extreme than predicted because of the need for humans in (or on) the loop.

For anyone looking to jump to full automation: Be precise about where AI still falls short. The work that’s hardest to codify is also the work that’s hardest to rebuild once you’ve automated it away. That’s the tricky part: Tacit knowledge is by definition hidden, so you don’t notice it until it’s gone.

If you liked this episode, you’ll love my book. Designing Automated Futures is coming this year with Rosenfeld Media. 📣 Sign up to be the first to know about new book releases, sales, and events.

📣 Call for Proposals: EPIC People Conference

Are you working on questions about modeling context or deploying context? Consider submitting a proposal to the EPIC People Conference, held October 25-27, 2026, in Chicago.

⏰ Deadline: Extended to March 2. Learn more here.

I’ve participated in EPIC in various capacities over the years and consistently find it to be a rewarding experience. It’s a rare space where practitioners talk with each other about messy, real-world problems and how to constructively move forward. The conference is worth checking out even if you’re not submitting.

This year, I’m excited to be co-chairing EPIC’s Immersion Day with Todd Palmer. We’re planning an interactive day of learning and making before the main conference. More soon.

⏪ Recent Episodes

ep. 86: ServiceNow’s Real Moat: Context Orchestration for AI

ep. 85: Back to School Vibes

ep. 84: The Importance of Anticipation

📖 Good Reads

Why the next internet will be agent-first x Amy Webb and Andrew Hornstra.

An AI Taxonomy x Narain Jashanmal

AI Doesn’t Reduce Work—It Intensifies It x Aruna Ranganathan and Xingqi Maggie Ye (HBR)

That’s a wrap 🌯 . More on UX, HCI, and strategy from Sendfull in two weeks.

Share

The iShares Expanded Tech-Software Sector ETF (IGV) experienced the largest one-day drop (5.4%) since April’s tariff-driven downturn last year. IGV is now down over 13% for the month and, if the pace holds, is headed for its worst monthly result since October 2008 (–23%).

In particular, ServiceNow, a cloud software company that automates enterprise workflows, got significant attention for dropping 10% in a day. This happened even after they posted better-than-expected Q4 results. Zooming out, ServiceNow is down about 43% in a year.

What’s happening with software stocks?

A key driver behind the IGV decline is investors’ mounting concern about AI disruption (it’s fair to say that Anthropic releasing three models in two months fanned the flames). However, ServiceNow’s drop suggests a specific belief can’t be swayed even by strong Q4 results: platforms designed to coordinate human work aren’t valuable in a future with fewer humans in the loop.

The Business of Orchestrating Work

ServiceNow built a $10 billion+ business on IT service management. You can think of it as a sophisticated coordination layer between people who need things and people who can provide them within an organization.

Can’t access the shared drive? ServiceNow logs your ticket, categorizes it, routes it to the right technician, ensures it gets resolved on time, and keeps everyone updated.

Need a new laptop? ServiceNow orchestrates the request through procurement, IT, and facilities.

Their revenue models heavily depend on seat licenses. More people in the system means more revenue. The same holds for companies like Atlassian, SAP, and Salesforce. This business model is great if you assume that knowledge work is growing and you will need to orchestrate the tasks that support it. Except, AI is challenging both of those assumptions.

What Can Actually be Automated

At face value, many of the tasks that ServiceNow orchestrates can be automated by AI. They look like routine tasks, characterized by repetitive procedures, clear decision criteria, and well-defined inputs and outputs.

In this automated world, if Bob can’t access the shared drive, Bob submits a request, “the system” checks that Bob has access, the request gets logged, Bob gets access. No need for orchestration because AI just took care of the task.

And, if the San Francisco tech billboards are to be believed, maybe you don’t even need Bob anymore.

Either way, you have fewer people paying for seats at ServiceNow. AI handles, say, 60% of the routine tasks, leaving only 40% of non-routine tasks that still flow through ServiceNow for human approval or execution.

So far, investor skepticism seems warranted. This future involves significantly fewer seats, and therefore, a fundamental change to ServiceNow’s business model. Pivoting towards something new is a big ask for a company that’s over 20 years old with ~30k employees.

However, investors seem less concerned about ServiceNow being up for the challenge, and more about the company no longer being relevant in an agentic future. The trouble is, they’re underestimating the amount of context AI needs to directly automate underlying tasks well in an enterprise organization.

Subscribe now

Much of the context that makes large organizations tick is invisible to AI, such as political dynamics and historical decisions. It lives in people’s heads, the meeting after the meeting, and hallway conversations. Even something as simple as “Reset this person’s password” seems routine until you consider: Is this person leaving the company? Do they still need access? Did they just get promoted and need different permissions? Should you notify their manager?

Part of ServiceNow’s value proposition was (and is) that it creates a context scaffold where none existed. It’s a place for humans to inject context through approvals and oversight. Unless you assume that enterprise-sized organizations go away in an AI-driven, org-flattening future, you still need this orchestration layer.

The Context Orchestration Paradox

When you talk about a partnership between humans and AI, you’re really talking about a system that pairs human and machine strengths. Humans are great at things like context curation, using judgment to determine what inputs an AI system receives. AI agents are great at things like coordinating repetitive actions at scale.

When you tell an agent to do things in an enterprise setting, correct execution often requires input or oversight from multiple human stakeholders. It’s a different challenge than if you’re using an agent to optimize your individual workflow, where you don’t need those extra steps. Compare it to making decisions that require input from two siloed organizations versus on a team of two.

Because organizations are complex systems, you get the context orchestration paradox: more AI autonomy actually increases opportunities for orchestration platforms that successfully capture, structure, and operationalize human context.

Yes, some “very” routine tasks will be automated. Some jobs will go away. ServiceNow will need to adjust to meet the reality of fewer seats and coordinating more complex tasks. Is there a risk that an AI-native company that doesn’t have 20 years of baggage from “the old way of working” figures this out better than ServiceNow? Yes. I won’t argue with that reason to sell your stock.

However, don’t bet against the importance of software as a context-AI orchestration layer in an agentic future. When you have a fleet of stochastic agents working in a complex system, you’ll need a mission control to steer them and system of record to track them.

On their Q4 earnings call, ServiceNow CEO Bill McDermott said: “The real payoff comes when trillions of tokens move beyond pilots to be embedded directly into the workflows where business decisions are made. ServiceNow is the gateway to this shift, serving as the semantic layer that makes AI ubiquitous in the enterprise.”

While he’s biased, I think he’s got it right.

If you liked this episode, you’ll love my book. Designing Automated Futures is coming this year with Rosenfeld Media. 📣 Sign up to be the first to know about new book releases, sales, and events.

🚀 Sendfull in the Wild: C100

On January 23, I had the pleasure of leading a workshop for Canada’s most promising scale-stage tech companies in the C100 Growth Program. I shared tools for structuring, focusing, and reframing customer learning goals as companies scale, drawing from the Sendfull Customer Understanding Toolkit. The opportunities and challenges of building at the forefront of AI and robotics emerged as a key theme throughout conversations with leaders.

I’ll be rejoining C100 on February 27, workshopping with the Fellows Program cohort on how to decide what to offload to AI versus keep human. I’ll be sharing a framework and tools from my upcoming book, Designing Automated Futures.

📣 Call for Proposals: EPIC People Conference

Are you working on questions about modeling context or deploying context? Consider submitting a proposal to the EPIC People Conference, held October 25-27, 2026, in Chicago.

⏰ Deadline: February 23. Learn more here.

I’ve participated in EPIC in various capacities over the years and consistently find it to be a rewarding experience. It’s a rare space where practitioners talk with each other about messy, real-world problems and how to constructively move forward. The conference is worth checking out even if you’re not submitting.

This year, I’m excited to be co-chairing EPIC’s Immersion Day with Todd Palmer. We’re planning an interactive day of learning and making before the main conference. More soon.

⏪ Recent Episodes

ep. 85: Back to School Vibes

ep. 84: The Importance of Anticipation

ep. 83: The Lawn Mower that Ate the Soccer Field

📖 Good Reads

Disempowerment patterns in real-world AI usage (Anthropic): In a new paper on how AI assistants affect human empowerment, research found that offloading belief formation, value-setting, or decision-making to AI can erode autonomy.

WGSN Future Consumer 2026 Report: The leading consumer trends forecaster identified four shopper profiles: the gleamers (burnt out folks looking for a simpler life), the autonomists (consumers rebelling against societal norms), the impartialists (people seeking honest messaging and simplified sales), and the synergists (counterculture and human-tech symbiosis advocates).

China’s Unitree shipped over 5,500 humanoids in 2025: Compare to the ~150 units each shipped per Tesla, Figure AI, and Agility Robotics. Assuming they only sold within their respective countries, China sold more both in absolute terms and per capita.

That’s a wrap 🌯 . More on UX, HCI, and strategy from Sendfull in two weeks.

Share

This time tomorrow, I’ll be making my way to UC Berkeley campus to kick off my new graduate course, UX for AI. My classroom will include 30 professional master’s students entering the job market within the next two years. Most are aiming for design and product roles.

I know what you’re thinking: What design and product roles?

Tech in general has been mired by post-pandemic layoffs, disproportionately affecting entry-level roles. The UX job market is a mess. Designers and product managers were affected by layoffs at a higher rate than engineers.

That said, there are reasons for hope.

First, hiring is slowly opening back up. No, not pre-pandemic-level opening-up, but not quite as bad as it’s been. Design roles are up almost 20% since the 2023 dip; product roles are up almost 54%. There’s also a general explosion in AI jobs. That includes roles at AI-specific companies like OpenAI and Anthropic and AI-specific roles at non-AI companies (e.g., AI PM at Figma). This includes emerging roles that require AI-wrangling expertise, such as design engineer.

AI skills are also increasingly tied to career advancement. Google, Amazon, and Meta now include AI use in performance reviews, with employees being asked to demonstrate how they increased their efficiency using AI tools.

What does this all mean for junior candidates, specifically?

Don’t get me wrong, the job market for entry-level roles is still rough. However, junior candidates do have a unique advantage over the old guard: Coming of age with AI tools.

For example, , Head of Growth at Lovable, the fastest-growing startup in history, recently sang the praises of new AI-native grads at the company. She called BS on the narrative that entry-level jobs will be automated away. Instead, fresh talent without years of industry baggage can see opportunities AI can unlock from a new perspective, making them an asset to teams.

Elena’s sentiments echoed a theme I’d heard across guest lecturers last Fall in my Intro to UX Design class. All of these folks, ranging from designers, to PMs, to founders, highlighted the importance of effectively applying AI tools and, relatedly, junior candidates’ unique advantage. A quote that stuck with me was a calculator analogy shared by guest Jason Kwok, PM at Meta:

These students are that generation, and this new course is all about calculators: how to use them, when to use them, how to make them, and—in a mind-bending act of calculator-inception—using one to make more.

Fresh Vibes

In a course where the learning objective is for students to become AI-native thinkers and builders, vibe coding is going to play a significant role. This isn’t about becoming a Cursor expert or a Lovable expert, especially when the landscape and capabilities are rapidly evolving. Instead, the aim is for students to build intuition about working with AI through hands-on exploration and making, and develop a critical eye for where these tools can augment their UX workflow.

Because this space is moving SO fast, and because I’m a researcher at heart, I decided to conduct some interviews to keep the course grounded in the latest_latest ways people are using vibe coding.

In a caffeine-fueled, 3-day listening tour, I spoke with 14 practitioners spanning product, design, and engineering. All used vibe coding in their practice, with tools including Bolt, Cursor, Claude Code, Figma Make, Lovable, Replit, and V0. Here’s a topline of what I learned.

The vibes only flow if you’re a thoughtful context curator.

Several practitioners fed a PRD and Figma design into tools like V0 or Lovable. Others crafted a prompt with detailed instructions about what the tool was supposed to create, accompanied by a screenshot of a wireframe or a UI that followed an existing design system. Regardless of the exact method, everyone started their vibe coding workflow with curated context. This starting point was much more intentional and structured than the term “vibe” suggests.

Maintain control through iterating in chunks.

After you’ve supplied your curated context and hit “enter”, you’ll get output that still requires tweaking. You need to take a disciplined, incremental approach to iteration to prevent AI models from “forgetting” context. For example, you might break down the highest priority workflow for the app and vibe code it step-by-step. Otherwise, you risk AI confidently veering off course, leading to frustrating, duplicative re-work down the line.

Bring a prototype instead of a PowerPoint to your next meeting.

Going into these conversations, I hypothesized that vibe coding would hurt collaboration through siloing, such as engineers running off with prototypes without consulting design. I was pleasantly surprised when I consistently heard that prototyping improved cross-functional collaboration.

Prototypes, whether they’re vibe coded or not, provide a concrete starting point for conversation. They’re a type of boundary object: a shared reference point that supports alignment and decision-making, even though each function brings their own lens (a designer thinking about applying design systems, or an engineer thinking about how to bring it to production). By enabling faster prototype creation, vibe coding shifts conversations away from clarifying what you meant on Slide 5 to discussing implementation tradeoffs.

Distinguish between prototypes and production-ready apps.

Prototypes test ideas. By definition, they are disposable, so the underlying code doesn’t matter much. Jump into Figma Make or Lovable and start building.

Production-level apps are different. What’s under the hood matters when you’re building end-to-end apps. Understanding foundational concepts like data structures and algorithms will get you farther, faster.

Most practitioners I interviewed came from data science or computer science backgrounds before moving into product or design roles. Regardless of whether this is sampling bias or a trend, their dual perspective positioned them uniquely to assess what technical knowledge actually matters for vibe coding.

The consensus: you don’t need a CS degree to vibe code a shippable app (that would defeat the point). However, basic programming concepts will help you more effectively architect your experience. This will be more relevant if you’re a non-technical founder looking to ship an app than if you’re a PM or designer who can work with an engineer who can help with build out the backend.

💡Pro tip 1: Most of the folks building production apps swore by Claude Code (Replit was in second place). These tools allow you to build more complex experiences, but also involve working with a command line interface.

💸 Pro tip 2: Keep an eye on your credit limits, especially when building end-to-end apps. Claude Code, while incredibly powerful, sucks credits like a Dyson on Redbull. Then there’s Replit, which automatically switches to pay-as-you-go once you exhaust your monthly credits.

Yep, Still in the Loop

Vibe coding has automated prototyping and coding work that would have previously taken hours, requiring specialized expertise.

However, I (still) don’t think the singularity is coming.

These tools are just that—tools, which need humans to steer them to be useful. Context curation is one of those things that humans are better at than machines. The clear problem framing and systems thinking required to thoughtfully curate that context are also automation-resistant.

A key outcome of my course, alongside students creating vibe-coded portfolio pieces, is having them reach the conclusion that the difference between slop and not-slop continues to be expert human judgement.

💬 What vibe coding knowledge do you want to impart on the next generation of design and product leaders? Let me know in the comments.

Subscribe now

⏪ Recent Episodes

ep. 84: The Importance of Anticipation

ep. 83: The Lawn Mower that Ate the Soccer Field

ep. 82: Live from Toronto: AI x Design Panel Recap & Reflections

📖 Good Reads

“I applied.” (But did you, though?) (). Even if you’re a qualified AI-native junior candidate and there’s a role, hitting “apply” on LinkedIn doesn’t mean you applied for it.

Everyone wants a room where they can escape their screens (Nora Knoepflmacher, Wall Street Journal): Analog rooms free of screens and smart devices are trending in America.

Epistemological fault lines between humans and artificial intelligence (Walter Quattrociocchi, Valerio Capraro, Matjaž Perc, Computers & Society): AI can sound convincing without actually checking or understanding what is true. This makes us offload important parts of thinking like judging, questioning, and verifying, without realizing it.

That’s a wrap 🌯 . More on UX, HCI, and strategy from Sendfull in two weeks.

Share

I immediately whipped out my phone to jot down this quote from the co-owner of our lodge in Futaleufú (Chilean Patagonia) before snapping this photo:

Practically speaking, this local knowledge was helpful because (a) you often experience all four seasons within a day in Patagonia, and (b) traditional weather forecasts aren’t particularly helpful here. An oracle-mountain to inform your clothing choices before going out on your morning run is kind of perfect.

The quote’s significance didn’t stop there.

The local knowledge of Sierra Teta helps you anticipate the weather to make better choices. It just so happens that I’m also working on a chapter of my book, Designing Automated Futures, that deals with anticipation and automation: How can you better anticipate (and mitigate) unintended consequences of deploying automated systems? And how do you see around the corner when technology is evolving at unprecedented speed?

That brings us to today’s episode.

Incidentally, this is also the final episode of 2025. What better theme than anticipation to ring in 2026?

Two Flavors of Anticipation

In a product development context, I like to think about anticipation taking two forms: tactical and strategic.

Tactical anticipation addresses immediate risks. You identify what could go wrong when you launch your technology and what you can do to mitigate those problems now.

For example, what happens if your AI-assisted diagnostic tool provides incorrect information to the doctor who is using it? What can you do to mitigate those problems now?

Strategic anticipation explores potential futures more broadly. For example, you might examine how changes in regulation, user behavior, or competitive dynamics could reshape your industry, creating opportunities or threats that don’t exist today. From there, you can adjust your strategy to remain competitive regardless of which future materializes.

For example, how might shifting customer expectations around data privacy alter what products succeed in the future? How will you adjust your strategy accordingly so that you are resilient across different potential futures?

When to do What

Both types of anticipation are important, but they serve different roles at different stages of product development:

Tactical anticipation happens during execution. It occurs during system design and testing, ideally before deployment. The goal is to catch and fix problems before they reach customers.

Strategic anticipation happens as you form your strategy across time horizons. Do this before committing to major investments, when you still have flexibility to change direction. Revisit it regularly (the frequency depends on your industry’s pace of change) and when emerging signals suggest the landscape may be shifting. You need enough lead time to reposition your organization, not just react to changes already underway.

What are signals? They’re small, often local, innovations of people doing things that differ from how the mainstream believes the world functions. They might be new ideas, technologies, or habits. Recent examples include synthetic users for research or AI-driven usability testing (more here).

A Few of my Favorite Tools

While there are several tools you can use for both types of anticipation, here are some of my favorites. These should all ideally be run in a workshop setting with your team, rather than in isolation. You’ll need to synthesize outcomes into themes, share out insights and recommendations, and ensure you follow through so that they’re actioned upon.

Tactical Anticipation Tools

• Pre-mortem: Imagine your product has failed. Work backward to identify what went wrong.

• Tarot Cards of Tech: Ask questions like, “what happens when 100 million people use our product?” and “What’s the worst headline about our product you can imagine?”

  Sidenote: I’ve used the Tarot Cards of Tech for years with graduate students in my Intro to UX Design class. They conduct an ethics impact assessment on the design direction their team has converged on, following a round of prototyping testing with participants. Shoutout to Sheryl Cababa for this awesome tool.

Again, these are best used before your product has launched, and while you still have some runway to make changes to mitigate the risks you identify.

Strategic Anticipation Tools

• The Futures Wheel: Visually map ripple effects of a change (or signal!) by identifying potential primary and secondary consequences (you can keep going to tertiary consequences, if you want).

  

• Horizon Scanning: Systematically monitor signals, trends, and wild cards across multiple domains.

  Tip: Track your signals using this foresight database template created by .

These two tools come from the discipline of strategic foresight, which is devoted to exploring possible futures and preparing for them. Remember, apply these tools before committing to major investments, when you still have flexibility to change direction.

Subscribe now

Why Anticipation Matters Now

Generative AI is evolving at an unprecedented rate. You can automate things that weren’t possible to automate even a year ago. Let me give you an example:

Last year, my Intro to UX Design class was still prototyping “the old fashioned way”, without generative AI’s ‘assistance’. Most students used Figma for the final project prototypes. Tools like Cursor or Figma Make weren’t even released.

This year, most students vibe coded prototypes for their final projects (after structured, hands-on practice and discussion of the pros and cons of this approach, of course).

The pace of generative AI development changes the stakes for both tactical and strategic anticipation:

• Tactically: Generative AI systems behave differently than rule-based automation. Identical inputs can produce different outputs; edge cases are difficult to predict in advance. You need to anticipate how the system performs across contexts it wasn’t explicitly trained for, how users might manipulate it through carefully crafted prompts, and what happens when it’s deployed for purposes beyond your intended use cases.

• Strategically: Technology developments that used to take years to develop now take months. What seemed unfathomable to automate 1-2 years ago is becoming standard practice, and in turn, changing ways of working. (See: My AI-enabled prototyping example). This makes it much more challenging to see around the corner, especially for longer time horizons.

No one can predict the future, but anticipation helps steer us towards a preferable one. Of course, that now leaves the question, “Who decides what’s preferable?” But that’s a question for another year.

If you liked this episode, you’ll love my book. Designing Automated Futures is coming in 2026 with Rosenfeld Media. 🔗 Sign up to be the first to know about new book releases, sales, and events.

🚀 Sendfull in the Wild

Recording of my November 21 talk, Still in the Loop: Leading with Human-Centered Design in the Age of AI, at the University of Toronto Faculty of Information.

Recap of the AI & Human-Centered Design Panel at the Rotman School of Management, written by event co-host, DesignMeets.

⏪ Recent Episodes

ep. 83: The Lawn Mower that Ate the Soccer Field

ep. 82: Live from Toronto: AI x Design Panel Recap & Reflections

ep. 81: What Toyota Can Teach Us About AI Automation

📖 Good Reads

Against the Slop by Julian Bleeker

Testing the Future of AI Interfaces: What We Learned by Valerie Caña

AI x Commerce in 2025 by Joe Kaziukėnas

That’s a wrap 🌯 . More on UX, HCI, and strategy from Sendfull in mid-January!

Share

None of these things.

The culprit was the team’s own autonomous lawnmower robot. They’d even given it a name: Robby.

The field was wet after a rainstorm, and Robby got entrenched in the soggy ground. The robot continued “mowing” long after its control system should have detected something was wrong and halted operation.

The damage was so extensive that the club won’t be able to host matches until spring, when the field can be restored with new grass or turf.

Why This Matters

This incident is both captivating and instructive because Robby should be the textbook example of automation done right. You’d be hard-pressed to pick a better task to automate than mowing the lawn:

• It’s repetitive.

• It operates in a bounded environment.

• It requires minimal expertise.

• Most people won’t miss doing it.

• It requires relatively little context*.

*Clearly Robby failed on this front, but we can agree that a robot mower operating in a confined space requires less context than, say, negotiating a client deal.

The problem lies in Robby’s design: autonomous systems need fail-safes.

Subscribe now

Redesigning Robby

We can look to safety-critical system design for tips on how to redesign Robby, saving SG Egels-Poppens a season of heartbreak. These tips also double as guidance for designing any autonomous system (think agentic workflows).

Wait, what does safety-critical mean? It’s a scenario where system failure could lead to significant harm, such as injury, environmental damage, or major financial loss. Commercial aviation is one example of a safety-critical domain.

Monitoring

The first step to fixing a problem is knowing you have a problem.

The current mower lacked situational awareness, failing to perceive relevant environmental cues (soggy grass, getting stuck in the mud), interpret them against its operational goals (mow grass, not create ruts), and update its actions accordingly.

A redesigned Robby would need continuous feedback loops between sensors, expected system behavior, and decision thresholds. For example, monitoring wheel torque against forward motion would have allowed the system to detect that it was exerting effort without making progress.

With these inputs, Robby’s control logic could have transitioned the mower into a safe, low-risk state (say, moving slower, pausing, backing up to firmer ground) rather than persisting.

Redundancy

Redundancy supports resilience by providing multiple channels for a system to notice when it is no longer operating within safe limits.

For Robby, redundancy would have ensured there was more than one way to detect and respond to abnormal field conditions. One option would be to combine wheel slip detection with independent measures such as soil moisture or GPS-based movement checks. If any one sensor failed, a second cue could still trigger a halt.

State Validation & Escalation

It’s ok to ask for help.

Once an issue with operation was detected, Robby’s system could have triggered a state validation step that halted operation and requested intervention (aka escalation). Enter human-in-the-loop exception handling. This might look like having the robot send a picture of the field to a human when it detected something was off about its operating conditions.

One problem: Robby is running overnight. No one wants to be woken up at 1am by a notification from your lawn robot.

What if you can get another machine to look at the picture and decide for you? It’s increasingly common that AI agents verify the work of other agents, so this isn’t far fetched. While adding an ‘agent checker’ might be overkill for most lawn mowing robots (well, maybe not in the SF Bay Area), it would reduce the number of times a human would have to assess the situation.

If you were designing automation for a high-stakes, contextually rich task, you’d want human judgement in the loop immediately upon exception-detection.

The Ghost in the Machine

Big-big picture: No matter how many machines are checking the work of other machines, you needed a human to orchestrate the machines at some point. This is a key principle of human factors research: Automation simply shifts human work upstream. You move from doing the thing to verifying the output of your automated system. With AI agents, you might be the human setting up the thing that’s verifying the output (of the thing doing the thing. Say that five times fast).

Sometimes verifying the output can be more taxing than just doing the thing in the first place. This is how automation can paradoxically exacerbate, rather than eliminate, challenges for human operators. This phenomenon is called Bainbridge’s Ironies of Automation.

That said, in the case of autonomous lawn mowing, even the occasional verification is probably worth not having to regularly mow a giant field.

Now What?

Remember Robby when designing autonomous systems:

• Design for Monitoring: Give autonomous systems the ability to notice when their behavior no longer matches expected conditions. Continuous sensing, comparison against norms, and clear decision thresholds prevent silent failure states.

• Add Redundancy: Don’t let a single signal (say, sensor input or LLM output) be solely responsible for determining outcomes. Add multiple ways to detect when performance is going off the rails.

• Plan for Escalation: Define how and when the system asks for help. Low-stakes tasks may allow automated validation, while complex or high-consequence scenarios require timely human judgement to interpret ambiguity and reset course.

All of that said: Whether it’s orchestration, monitoring, or verification, people are still part of the system. Automation isn’t all-or-nothing. It’s a partnership between humans and machines.

This episode is a sneak peek into topics from my forthcoming book with Rosenfeld Media, Designing Automated Futures. 🔗 Sign up to be the first to know about new book releases, sales, and events.

🚀 Sendfull in the Wild

Recording of my November 21 talk, Still in the Loop: Leading with Human-Centered Design in the Age of AI, at the University of Toronto Faculty of Information.

Recap of the AI & Human-Centered Design Panel at the Rotman School of Management, written by event co-host, DesignMeets.

⏪ Recent Episodes

ep. 82: Live from Toronto: AI x Design Panel Recap & Reflections

ep. 81: What Toyota Can Teach Us About AI Automation

ep. 80: Automation in the Field: A Conversation with Konstantinos “Kostas” Kandylas

📖 Good Reads

Half My Day Is Robots Now by : A framework for how you’ll feel about AI, based on what you do for a living.

Where Are You Actually Trying to Go? by : How to map your long-term goals (not just for work).

The future of AI-powered sales with Vercel COO, Jeanne DeWitt by : What go-to-market looks like in the AI era.

That’s a wrap 🌯 . More on UX, HCI, and strategy from Sendfull in two weeks!

Share

Greetings from my hometown of Toronto! 👋 🇨🇦

Last night, I had the pleasure of participating in a panel discussion on AI and Human Centered Design, hosted by Design Meets x Business Design Initiative at the University of Toronto’s Rotman School of Management.

I shared the stage with three design leaders—Patrick Bach, Kem-Laurin Lubin, and Dr. Alex Ryan, alongside our thoughtful moderators and pillars of the Toronto design community, Michael Dila and Dr. Emma Jo Aiken-Klar. Today’s episode captures my recap and reflections from our conversation.

TL;DR

1. What are we noticing?

Organizations are integrating AI quickly—often faster than processes and culture can adapt. The tech is showing up in all aspects of work, from research workflows and design tooling, to hiring and organizational expectations. Core themes:

• Top-down pressure + bottom-up experimentation: Organizations are encouraging or mandating AI adoption, while practitioners experiment to understand where it genuinely adds value.

• New frictions and uncertainty: We’re seeing uneven tool maturity, unclear norms of use, and the risk of over-reliance on AI outputs that “sound right” but lack depth.

• Blurred swimlanes: Tools like Figma Make allow non-designers to quickly produce prototypes, collapsing traditional role boundaries

2. What assumptions about human-centered design are being challenged?

AI is challenging many of the foundational ideas of human-centered design. Key tensions:

• Problem-first vs. solution-first sequencing: AI is an example of radical innovation, evolving at an unprecedented pace relative to previous tech. It is a solution unlocking new problem spaces—the reverse of the “people-first, solution second” that undergirds traditional human-centered design.

• The human is no longer the only “actor”: AI agents become participants in the system, raising unprecedented questions about designing with non-human actors. This includes the use of synthetic users in design research, and considering agents as an actor in our service blueprints.

• Linear iteration no longer fits how work moves: AI allows you to quickly and simultaneously generate—and iterate on—multiple design concepts. This challenges the old “research > design > build” pipeline into an always-on loop where teams refine direction in real time.

3. How are our practices and workflows evolving?

Design work is becoming a hybrid process where people and AI collaborate. A few examples from our conversation:

• Braided intelligence: Teams can now combine individual intelligence, collective intelligence, and artificial intelligence. This unlocks new possibilities like real-time sense-making during workshops. Contrast this with long lead times for synthesis and sharing out a summary weeks after the workshop has passed.

• Acceleration of prototyping: Tools like Figma Make and Lovable enable immediate interface generation and faster iteration, accessible to designers and non-designers alike.

• Restructuring of team roles: When everyone can vibe code a prototype, designers no longer “own the pixel” by default. Their value shifts more squarely toward problem framing, navigating ambiguity, and ensuring rich human context is represented in the design process.

4. What new opportunities and risks do we see?

Opportunities

• Massive efficiency gains: Design practitioners reclaim time from transcription, reporting, and first-pass synthesis, enabling deeper conceptual and strategic work. This can be especially valuable for civic tech practitioners working with limited resource, enabling them to “do more with less”.

• Broader exploration: The recurring theme of faster prototyping showed up here as well. The ability to quickly create prototypes can enable more parallel experimentation and faster learning.

• Unlocking new resources: AI can be used to quickly translate texts in different languages, helping increase the breadth of global perspectives that a practitioner can access.

Risks

• Loss of humanity (and surprise): Synthetic users risk turning us all into data points, stripped of rich context and lived experience. They also can never replace the emotional resonance or unexpected insights of a real human story. We’ll always need those stories to shift c-suite minds and hearts.

• Erosion of foundational skills: Early-career practitioners risk not learning synthesis, critical thinking, or ambiguity tolerance if they rely exclusively on generated outputs.

• Reinforcing bias and inequity: Examples such as biased hiring models show how quickly AI can amplify structural inequities if not interrogated.

5. How can human-centered design shape the future of AI?

Designers have a clear role as bridge builders, systems thinkers, and meaning-makers. Key roles the field can play:

• Rethinking the GUI: We’re interacting with bleeding-edge tech through old-school interfaces like the prompt bar—GUIs that haven’t evolved much from the 1968 Mother of All Demos. Designers are uniquely positioned to invent new interaction models.

• Defining the design metaphors that guide mental models: Language like “server farms” shapes how people imagine systems; designers can reexamine these frames to influence cultural expectations of AI.

• Bending the solution-first momentum: One of design practitioners’ unique strengths is building bridges between disciplines, people, and ideas. We can lead teams through hands-on experimentation with the tech to learn capabilities and affordances, and then connect that back with what we know about people’s unmet needs. This helps balance the solution-first momentum AI is exerting on product development, bringing the human back in the loop.

Breaking the GUI Glass

A recurring theme across the evening was that of designers “breaking the GUI glass” (thanks for the phrase, Kem!) Unpacking that:

Design traditionally meant learning the technology underlying visual design—think HTML and JavaScript. Along the way, many designers got stuck at the GUI level. UX Bootcamps and the proliferation of hi-fi design tools like Figma didn’t help.

When AI tools can create convincing interfaces in the matter of minutes, GUI-level proficiency isn’t going to cut it. Designer practitioner need to go deeper:

• Looking beneath the interface: Understand how models work—for instance, the data on which they’re trained—and how that shapes the user experience.

• Rethink interaction paradigms: Transcend the prompt bar to better support how humans collaborate with AI.

• Reclaiming agency at the systems layer: Return to our roots as systems thinkers. Design the mental models that guide how people understand and use AI systems, grounded in user goals and context.

On the Importance of Conversation

When you hear “panel discussion”, it’s easy to picture a row of talking heads and rehearsed responses.

This event was not that.

It was truly a conversation—yes, panelists answered questions, but the audience was also invited to share their responses. As we all figure out how to build this AI plane as we’re flying it, we need this community dialogue and collective sensemaking. Bravo, Design Meets and BDI. To everyone organizing tech events, especially in the Bay Area: More like this, please.

🚀 Sendfull in the Wild

Tune in for free tomorrow. Friday November 21, for my talk, Still in the Loop: Leading with Human-Centered Design in the Age of AI.

• What? I’ll be sharing a framework and tools from my forthcoming book, Designing Automated Futures, to help practitioners navigate and thrive in the evolving AI landscape.

• Where? Online and in-person at University of Toronto.

• How? Get tickets here.

⏪ Recent Episodes

ep. 81: What Toyota Can Teach Us About AI Automation

ep. 80: Automation in the Field: A Conversation with Konstantinos “Kostas” Kandylas

ep. 79: Principle #3: Just because you can automate, doesn’t mean you should.

📖 Good Reads

In Search of the AI Bubble’s Economic Fundamentals: A fresh take on the AI bubble grounded in historical context.

GenAI Art Is the Least Imaginative Use of AI Imaginable: I keep coming back to Dr. Ge Wang’s quote, “What if the point of art is that we actually make it?”

AI x Design Foresight Scan: Helpful landscape overview, led by panel moderator and co-organizer, Dr. .

That’s a wrap 🌯 . More on UX, HCI, and strategy from Sendfull in two weeks!

Share

This sentiment captures a dominant mindset underlying much of AI development. Look no further than the I-80 billboards driving into San Francisco—notably Artisan’s controversial 2024 “stop hirring humans” ad campaign.

The typo is intentional. It suggests that Artisan’s AI employees are an improvement over us error-prone flesh bags.

While the campaign was designed to “rage bait”—and Artisan later said that their real goal is to automate work humans don’t enjoy—this wasn’t a baseless PR stunt. The belief that humans are the rate-limiters to maximum efficiency runs deep.

The Dream of Full Automation

Frederick Winslow Taylor—the mechanical engineer who authored The Principles of Scientific Management (1911), named the most influential management book of the twentieth century by the Academy of Management—saw human labor as the weak link in the industrial system.

To him, workers were unreliable machines. He sought to identify the most efficient way to complete a task, and then select and train a worker to complete exactly that task. This approach would get people to operate with the predictability of machines, boosting productivity. Two years later, Henry Ford translated that logic into practice, introducing the first moving assembly line to mass-produce cars, realizing Taylor’s vision of efficiency.

Imagine if Taylor knew what we were up to with our current cognitive automation efforts with AI agents. What could be more efficient than replacing inefficient knowledge workers?

At the root of this worldview is the belief that humans are inherently limited, and that machines can help us transcend those limitations. Sam Altman’s 2016 interview in The Atlantic captures this perspective:

“There are certain advantages to being a machine. We humans are limited by our input-output rate—we learn only two bits a second, so a ton is lost. To a machine, we must seem like slowed-down whale songs.”

If fewer humans mean greater efficiency, then full automation seems like the logical endgame. Artificial General Intelligence is the Bach to our human slowed-down whale songs.

The Cost of Humans Out of the Loop

The trouble is, designing humans out of the loop tends to yield the opposite of limitless productivity. Consider a few examples:

• General Motors (1980s): Chairman Roger B. Smith sought to revitalize GM through building a fully-automated “factory of the future”. What started off as $52 million investment ballooned into $40 billion initiative. The result? Robots misidentifying car models, misattaching parts, and painting each other instead of vehicles. The Michigan factory that was the centerpiece of this effort closed in 1992.

• Boeing (2015–2019): The company introduced the “Fuselage Automated Upright Build” system to drill and fasten aircraft sections. Except, the robots frequently produced damaged or incompletely assembled fuselages, forcing Boeing to bring back human machinists.

• Klarna (2024–2025): The fintech firm replaced customer service agents with ChatGPT-powered AI handling over two-thirds of customer chats. Within a year, plummeting service quality and unempathetic or incorrect responses led Klarna to rehire human employees.

It turns out that you do need humans in the loop. The challenge isn’t whether to involve them, but how. The Toyota Production System offers a blueprint for designing automation that keeps humans in the foreground, whether in manufacturing or AI.

The Loom that Knew When to Stop

In 1896, Sakichi Toyoda, founder of Toyoda Automatic Loom Works (the foundation for Toyota), invented the Toyoda Power Loom. It quickly earned a reputation for producing high-quality cloth thanks to a weft-stopping mechanism that automatically halted the machine when a thread broke or ran out. This prevented defective fabric and reduced waste while allowing workers to focus on fixing issues rather than constantly monitoring the loom. The idea became the foundation of jidoka, or “automation with a human touch,” sometimes called autonomation.

Automation With a Human Touch

Fast forward to post-WWII Japan. The country’s manufacturing industry was devastated. Toyota, which had shifted their focus from textiles to automobiles in 1937, struggled to compete with resource-rich American automakers like Ford, who focused on mass production. Toyota worked with their constraints, and developed the Toyota Production System. It was built on the pillars of jidoka, and just-in-time manufacturing, which means producing only what is needed, when it’s needed.

The Toyota Production System would transform Toyota into a global leader in the auto industry, improving quality and efficiency, while cutting waste and cost. Jidoka was key to this success. Two of its most famous implementations were:

• Andon cords: Each worker could pull a cord (or press a button) to stop the line if they noticed an issue, such as a defect, safety hazard, or malfunction. This process empowered front-line workers to halt production, immediately addressing issues that would otherwise degrade quality.

  

• Mistake-proofing (poka-yoke): Machines were designed to detect and prevent human error—for example, by refusing to proceed if a bolt wasn’t fully seated. Just like the Toyoda Power Loom, this meant humans didn’t need to watch the machines continuously to maintain quality. This also let an operator manage several machines simultaneously.

A Culture of Shared Responsibility

You might wonder: wouldn’t it be terrifying to be the person who pulls the andon cord?

This is where Toyota’s approach gets even more interesting. The andon cord represents a culture of shared responsibility. When a worker pulls the cord to signal a defect or delay, coworkers immediately “swarm” to the spot, and the first thing they do is thank the person who raised the issue. Pulling the cord is seen not as a failure, but as an act of duty and care to teammates, the company, and the customer.

Toyota’s leadership played a key role in creating this culture. Processes and incentives were designed to make accountability feel collective rather than punitive. Managers were expected to coach rather than blame. Continuous improvement (kaizen) was baked into everyday operations.

Where Ford’s production system emphasized technical efficiency, Toyota’s was sociotechnical. It integrated human judgment as an essential part of quality and learning, and built a culture where expressing that judgement simply became the right thing to do.

Takeaways for AI Automation

Toyota succeeded because it designed its manufacturing system for human intervention, not human absence. Human judgment and accountability were rewarded, not replaced. Automation served Toyota’s vision of delivering quality through efficiency, not efficiency alone. If you’re asking how to automate with AI, the same principles apply:

Automation Serves Mission

For Toyota, automation was a means to an end—a way to improve quality and reduce waste. This contrasts sharply with “AI-first” strategies that prioritize technology over purpose. Start with your organization’s mission: how can AI create customer value, not just reduce costs?

Incentivize Accountability

When a model behaves unexpectedly, what’s your equivalent of the andon cord? Do people feel empowered to raise concerns—and will leadership act on them? Build feedback loops where human judgment is expected and rewarded, with clear pathways for surfacing and resolving issues that impact product quality and trust.

Automate with a Human Touch

Toyota’s machines stopped automatically when an error was detected, prompting human review. Apply the same mindset to AI. Design for graceful handoffs between automation and human oversight. Build fail-safes that pause, flag, or escalate anomalies—keeping people in the loop where judgment matters most.

Subscribe now

This episode is a sneak peek into topics from my forthcoming book with Rosenfeld Media, Designing Automated Futures. 🔗 Sign up to be the first to know about new book releases, sales, and events.

🚀 Sendfull in the Wild

Upcoming Events

DesignMeets x Business Design Initiative: AI & Human-Centered Design

• When? November 19

• What? I’ll be speaking on panel about how human-centered design is evolving in the age of AI.

• Where? Rotman School of Management, University of Toronto. Get tickets here.

Still in the Loop: Leading with Human-Centered Design in the Age of AI

• When? November 21

• What? I’ll be sharing a framework and tools from my forthcoming book, Designing Automated Futures, to help practitioners navigate and thrive in the evolving AI landscape.

• Where? In person at the University of Toronto and online. Get tickets (free!) here.

⏪ Recent Episodes

ep. 80: Automation in the Field: A Conversation with Konstantinos “Kostas” Kandylas

ep. 79: Principle #3: Just because you can automate, doesn’t mean you should.

ep. 78: Principle #2: Technology can automate tasks, not jobs.

📖 Good Reads

A compilation of the latest AI statistics (Exploding Topics)

AI enterprise adoption report (The Wharton School)

Korea rolls back AI-powered textbooks (Rest of World)

That’s a wrap 🌯 . More on UX, HCI, and strategy from Sendfull in two weeks!

Share