# The AI paradox: More automation, more humans, more work | Dan Shipper

Podcast: Lenny's Podcast: Product | Career | Growth
Published: May 24, 2026
Reading time: 21 min
Canonical: https://podbrew.app/briefs/lenny-s-podcast-product-career-growth-the-ai-paradox-more-automation-more-humans

Dan Shipper, co-founder and CEO of Every, shares his unique perspective on the future of work in an AI-driven world. His company functions as a living laboratory for artificial intelligence, with every employee actively integrating AI tools into their daily operations.

Dan presents a set of insightful predictions on the evolving landscape of work. The discussion covers topics such as the shifting economics of SaaS, the emergence of company-wide "super-agents," and the surprising paradox that more automation will lead to an increase in human tasks. He also outlines how roles for product managers, designers, and forward-deployed engineers are being redefined by AI.

These observations are particularly significant given Dan's past accurate forecasts regarding AI's impact. His company's deep, practical engagement with AI offers listeners a direct and informed look into the fundamental changes reshaping technology, business, and individual careers.

## Key takeaways

- AI models like Anthropic's Mythos Preview are demonstrating advanced autonomy, capable of performing tasks for 17 hours at 50% accuracy.

- Despite rapid advancements in AI's ability to do work, humans are predicted to have more work, not less, creating a notable paradox.

- The early vision of every employee having their own personal AI agent is giving way to a centralized 'super-agent' model for entire companies.

- The transition is due to the high maintenance and debugging effort required for individual agents, which proved too burdensome for most users.

- A new role, the 'forward-deployed engineer,' is becoming crucial for managing and ensuring the effectiveness of these company-wide super-agents.

- Advanced AI agents like Codex and CoWork are becoming the primary interface or "operating system" for professional work, expanding beyond coding to general knowledge tasks.

- A "reverse paradigm shift" is underway, where AI agents integrate browsers and full computer access directly, allowing them to observe and assist with all user activities.

- Users interacting with SaaS applications through their own AI agents will supply their own AI tokens, shifting AI computational costs away from SaaS vendors.

- SaaS product development must adapt to cater to both human and AI agent interaction, emphasizing agent-friendly interfaces like robust CLIs and usable HTML.

- Software is evolving to support seamless human-agent collaboration, requiring new UX features like approval flows, real-time logs, and rollback capabilities.

- The widespread use of CLIs for general-purpose work is ending; GUIs that incorporate agent assistance will become the dominant interface for human-agent collaboration.

- Collaborating AI agents can drastically accelerate interactions by having one agent supply rich user context to another application, bypassing manual input.

- This model can transform onboarding: instead of users filling forms, an agent like Codex, already familiar with the user, can directly configure a new app with relevant information.

- Automation with AI does not eliminate human work but transforms it into oversight and management roles, as humans are needed to ensure automated processes are functioning correctly and effectively.

- Current AI models, even advanced ones, struggle with strategic reframing and 'rewriting from first principles,' often addressing symptoms rather than identifying and solving fundamental architectural flaws.

- Benchmarks can overstate AI autonomy by measuring performance on problems already framed and articulated by humans, obscuring the essential human work involved in problem definition and strategic guidance.

- AI allows non-technical employees to contribute to technical development, leading to a significant increase in pull requests from diverse roles.

- The emphasis for technical teams shifts from building new features to maintaining coherence and managing the quality and integration of AI-generated work.

- AI primarily commoditizes "yesterday's human competence," making basic outputs (like generic landing pages or simple writing) cheap and widespread, thus losing their value.

- Humans will increasingly use AI's commoditized output as a foundation to create novel, interesting, and bespoke solutions, constantly pushing the boundaries of new expertise that models then try to catch up to.

## 00:00 - 06:01 Every's AI-First Culture Enables Prescient Predictions on the Future of Work

Dan Shipper accurately predicted the significant impact of Claude Code for non-technical work, a year before its widespread adoption and evolution into tools like CoWork and Codex. His early insight highlighted how AI could empower everyday users to manage files and perform various non-engineering tasks, a trend that has since become a key driver of Anthropic's success.

Shipper's company, Every, operates with a deeply AI-forward culture. All 30 employees, including writers, designers, salespeople, and customer service staff, are early adopters who actively use advanced AI tools like Codex and CoWork in their daily workflows. This collective immersion in AI technologies creates a unique environment where the team is, in effect, living in the future of work.

This 'pocket of the future' approach provides Every with unparalleled insights into how AI will reshape professional landscapes. By integrating AI into every facet of their operations and gaining early access to new AI models for beta testing, they develop a firsthand understanding of emerging trends, allowing Shipper to make highly accurate predictions about the evolving nature of jobs and tools.

> What you don't wanna do is prognosticate. What do you, what you wanna do instead is, is just live in it together.

## 10:01 - 12:01 The AI Paradox: Why More Automation Leads to More Human Work

Despite impressive AI benchmarks, such as Anthropic's Mythos Preview model capable of autonomously performing tasks for seventeen hours at fifty percent accuracy, the conventional wisdom that AI will simply eliminate human jobs is challenged. These benchmarks suggest significant progress in AI's ability to handle complex operations independently.

However, a contrasting prediction suggests that even as AI models improve, humans will actually find themselves with more work to do, not less. This presents an interesting paradox where increased automation paradoxically leads to an increase in human tasks.

This shift in work structure is anticipated to bifurcate in two main ways. One significant change is that individuals and companies will increasingly utilize dedicated agents to whom they can offload various tasks, mirroring early predictions of AI's role in the workplace.

> We actually have a lot more work to do. Humans have a lot more work to do, even as models get better at doing work. And there's like a really interesting paradox there.

## 12:01 - 18:02 Companies are shifting to a single super-agent managed by forward-deployed engineers

Initially, there was an expectation that every individual or team would adopt their own AI agent, creating a parallel organizational structure. This idea was likened to having a personal 'daemon' that reflected the user.

However, this model proved impractical due to the significant maintenance burden. Individual agents frequently 'break all the time' and require constant troubleshooting, which most employees are unwilling or unable to commit to.

This has led to a shift towards a 'super-agent' model, where a single AI serves the entire company, as seen at Shopify and Ramp. The core reason is that for an AI agent to be truly useful, it needs dedicated human oversight and care.

To address this, a new role, the 'forward-deployed engineer,' is emerging. These engineers are responsible for managing the company's super-agent, ensuring its functionality and relevance for all employees. This top-down approach is expected to evolve, potentially allowing more specialized team or personal agents as AI technology becomes less complex to manage.

> in order for An AI agent to be useful right now, it really needs a human who cares about it.

## 18:02 - 24:04 AI Environments Evolve into Operating Systems for Professional Work

AI agents like Anthropic's Claude Code and OpenAI's Codex are rapidly evolving beyond specialized coding tools into comprehensive environments for general professional tasks. Initially leveraging terminal access to become powerful coding assistants, these platforms, including Anthropic's CoWork and OpenAI's updated Codex desktop app, now aim to handle a wide array of knowledge work.

A significant "reverse paradigm shift" is occurring: instead of integrating AI into existing browser-based tools, these AI agents are incorporating in-app browsers and full computer access. This allows the AI to observe and interact with everything the user does, turning the AI environment itself into the primary interface for work.

This new setup significantly boosts productivity across various workflows. The speaker, for example, uses Codex daily for writing documents by having it watch and assist within an in-app browser. The agent also handles email management, processing communications and conducting research to maintain "Inbox Zero," fundamentally transforming how professional tasks are executed.

> For a long time, we thought... that the optimal experience of AI was gonna be take AI and put it in a browser. And I think the reverse is actually starting to happen... which is take the AI agent that you use all the time on your computer and put a browser in it so it can see everything you're doing, and that is just like a magical combination.

## 24:03 - 26:04 SaaS Companies Improve Margins by Letting Users Bring Their Own AI Tokens

A significant shift in SaaS economics is emerging: users are beginning to run applications within their personal AI agent environments, utilizing their own AI tokens. This means that when a user interacts with a SaaS product through their agent, the costs associated with AI computation, such as token usage, are borne by the user, not the SaaS vendor.

This model fundamentally changes the cost structure for SaaS companies. Historically, integrating AI into a product meant the vendor would incur the costs of AI tokens. With users bringing their own AI, SaaS providers can reduce their operational expenses related to AI, leading to improved profit margins.

To thrive in this new landscape, SaaS companies must adapt their product development. The focus will shift to building applications that are not only user-friendly for humans but also highly compatible with AI agents. This involves ensuring accessible interfaces, such as well-structured HTML and robust command-line interfaces (CLIs), allowing agents to interact seamlessly with the application's functionalities. The Proof app exemplifies this, where the company avoids token costs because users supply their own AI.

The implication is that SaaS development will evolve to serve a dual audience: human users and AI agents. This requires a renewed focus on making core functionalities accessible and controllable programmatically, ensuring that any action taken via an agent is immediately reflected for the user, maintaining a cohesive experience across both interaction types.

> I don't pay for tokens because they're just bringing, they bring their AI to Proof, and so it changes your margins back to, well, I don't really have to pay for tokens anymore 'cause the user is gonna bring the AI.

## 26:03 - 28:05 AI models require a 'harness' or platform to achieve their full potential, moving beyond simple prompt-and-response interfaces.

AI models alone are not sufficient; they need a sophisticated 'harness' or platform to unlock their full capabilities. This means that merely providing a prompt-and-response interface for a model is becoming obsolete, as companies recognize the need for more advanced environments.

Model developers like OpenAI and Anthropic are starting to offer managed agent environments. These platforms run the model on a cloud-based computer, providing a more integrated and powerful way to interact with the AI. This shift aims to extract the best possible results from the underlying models.

Cursor is presented as an example of such a harness, specifically tailored for programmers, though its potential expansion beyond coding is also noted. The recent close association of Cursor with SpaceX further underscores the industry's recognition of the value in these dedicated AI platforms.

The ultimate form of this harness is envisioned as a tool capable of supporting any kind of knowledge work, indicating a future where AI platforms are integral to various professional tasks.

## 28:04 - 32:05 Designing software for human-agent collaboration requires new user experiences.

The way we interact with software is shifting from human-only applications or agent-only command-line interfaces (CLIs) to a new paradigm of human-agent collaboration. In this future, both humans and agents work together on the same task, necessitating a seamless back-and-forth where each has visibility into the other's actions. This fundamental change will lead to very different types of software products.

This collaboration allows for simpler products initially; for example, many formatting tasks that humans traditionally handle in tools like Word are now managed by an agent. However, new user experience (UX) affordances become crucial. Software will need features like approval flows, a summary inbox of agent activities, real-time logs, and the ability to quickly roll back changes, especially since agents can perform a vast number of actions rapidly.

The infrastructure supporting this also needs to adapt. Agents can make a billion requests in seconds, which is already causing scaling challenges for platforms like GitHub as agent usage skyrockets. A significant advantage of this shift is improved problem-solving: agents can generate highly detailed bug reports with exact reproduction steps and insights into the codebase, enabling faster, even automated, fixes through direct agent-to-agent communication with company systems.

This new era also marks the end of CLIs as the primary interface for agent interaction for most users. While CLIs won't disappear entirely, the belief that they represent the future for cloud code is being supplanted. Graphical User Interfaces (GUIs) that seamlessly integrate agent assistance will prove superior for creating effective collaborative environments between humans and AI agents.

> CLIs are over. We speed ran the CLI era. It was nice while it lasted, but it's pretty clear they're not going to completely go away.

## 34:05 - 36:05 AI Agents Can Accelerate Onboarding and Complex Tasks

The future of AI involves multiple agents collaborating, where one agent, such as Codex or CoWork, provides rich contextual information to another application or agent. This collaboration creates a significant "speed-up effect" by automating the transfer of details that would otherwise take a human a long time to input or explain directly.

For instance, in user onboarding for new software, instead of building traditional web interfaces or Slack workflows to gather user information, a new approach can be taken. If a product assumes users are interacting through a powerful agent, that agent can directly communicate with the new app.

This allows the primary agent, already holding extensive user data and history, to provide specific context to the new application. The app then receives tailored information, such as past work or potential use cases, directly from the agent. This streamlines the setup process, creating a highly customized experience, particularly beneficial for complex or technical products.

> Codex has so much information about you that it can just give it, "Here's all the stuff I've been working on with Dan, here's all the ways that, you know, he might wanna use this app, and then bring it back to me."

## 38:05 - 46:08 AI Automation Paradoxically Increases the Need for Human Oversight and Strategic Framing

Despite advancements in AI, automation paradoxically amplifies the need for human oversight. The notion that automation eliminates human work is a misconception; instead, it redefines human roles, shifting them towards management and ensuring the automated systems function effectively. Companies leveraging AI, like the guest's, have even seen their human workforce double, indicating that human expertise remains crucial for guiding and verifying AI applications.

Benchmarks often misleadingly suggest greater AI autonomy than exists. For instance, in a 'senior engineer benchmark' designed to assess AI's ability to rewrite poor code from scratch, models like GPT-5.5 achieved a 62 out of 100, while human senior engineers scored in the high 80s or low 90s. Earlier models performed even worse, scoring around 30. This highlights a significant gap in AI's capacity for fundamental problem-solving and proactive architectural reform.

AI models typically struggle with 'framing problems' and lack the agency to rewrite systems from first principles. When presented with issues, human senior engineers instinctively analyze the underlying codebase and may recommend a complete rewrite due to fundamental flaws. In contrast, current AI models tend to address reported issues directly within the existing framework, even when a more radical solution is warranted, demonstrating their limitation in strategic thinking and independent problem reframing.

This enduring need for human judgment, strategic framing, and the ability to define problems at a higher level means that human engineers and managers remain indispensable. Even as AI improves, the human role in discerning what needs to be built, how it should be built, and guiding the AI through complex, undefined challenges will continue to be vital, ensuring that automation genuinely serves its purpose without compromising quality or strategic direction.

> Automation is a lie: in the sense that every time you automate something, in order to make sure the automation is working well, you need a human on top of it, like making sure that it's working well.

## 50:09 - 56:11 AI Changes Work Shapes: Generalists Thrive and Forward-Deployed Engineers Emerge

Artificial intelligence is enabling non-technical individuals in roles like consulting, operations, and editing to perform technical tasks, such as creating pull requests. This dramatically increases the volume of technical contributions across a company, shifting who can actively participate in development. For example, OpenClaw receives thousands of pull requests daily, many from non-engineers.

This surge in AI-assisted output places significant pressure on existing technical teams. Their focus shifts from simply building things to ensuring coherence, quality, and integration with the existing product. Managing this influx involves deciding which pull requests to merge and what to delete to prevent bloat and maintain a unified system.

The blurring of job boundaries means engineers can design, product managers can code, and marketing can ship, initially causing confusion about specific responsibilities. However, this environment increasingly favors generalists, who can contribute effectively across multiple domains, especially beneficial for smaller organizations.

A new job role, the 'forward-deployed engineer,' is emerging to manage and refine AI agents. These individuals are responsible for making sure agents perform correctly, often by interacting with them conversationally. This role highlights that despite AI's growing power, human oversight and maintenance are still essential, creating new specialized jobs rather than solely eliminating them.

> every agent needs a human.

## 56:10 - 1:00:11 AI Transforms Data Science Roles While Sales Functions Remain Largely Unchanged

The proliferation of AI-generated content is fundamentally altering certain job functions, particularly for data scientists. Their role has shifted from primarily conducting analyses to spending most of their time reviewing and correcting the high volume of inaccurate or sloppy AI-produced data science work shared by others across an organization.

To address this, some large companies are implementing specialized data science bots. These agents are integrated with data warehouses and permission systems, capable of answering basic queries that people might ask. This offloads the mundane tasks, allowing human data scientists to focus on deeper, more complex analytical questions that require specialized expertise.

While many product and tech roles, such as engineering, product management, and design, are experiencing significant changes due to AI, some roles remain relatively unaffected. Sales functions, for instance, are identified as being among the least fundamentally altered. AI primarily assists with top-of-funnel activities rather than transforming the core aspects of the sales job itself.

> most of their job is now reviewing bad data science work

## 1:02:11 - 1:06:12 AI-Generated Writing Will Become Preferred for Professional Communication

Dan Shipper predicts that people will increasingly read and even prefer AI-generated writing in professional documents and emails. He believes the aversion to AI content will diminish as its utility becomes clear, especially when directed well by humans.

For example, his company used Notion agents for quarterly planning, where each employee interacted with an agent to discuss past performance and future goals. This process generated "incredibly good AI-generated strategy reports" for each team, making it easier to identify inter-team collaboration needs and assess quality.

Shipper argues that for many people, the quality of human-written strategy documents and emails is often low. A well-directed AI, like GPT-5.5, can produce superior informational content compared to what many individuals can draft manually.

He personally uses GPT-5.5 in Codex for most of his emails. While acknowledging the need for human oversight, he suggests that AI can act as a highly effective co-author, streamlining communication and improving output quality for various tasks.

> The kind of strategy document that GPT-5.5 can write when it's directed well by someone on my team is way better than like them just like dinking and dunking like, like their fingers on the keyboard, right? Like most people are really bad at writing strategy documents, the bar's low.

## 1:08:12 - 1:12:13 AI Empowers Product Managers and Full-Stack Designers to Build and Innovate

AI is transforming who will be successful in the tech industry, particularly empowering Product Managers (PMs). With advanced coding models, PMs can now build and deploy products much faster, shifting their focus from team coordination to leveraging their core product sense, understanding of users, and problem-solving skills directly.

An example is Marcus, a PM at Spiral, who, after becoming proficient with AI coding tools, now ships products faster than almost anyone on his team. Despite being only "lightly technical," his strong product intuition combined with AI tools makes him incredibly effective, allowing him to concentrate on user needs and product direction without extensive team management.

Similarly, full-stack designers are uniquely positioned to excel. Designers often face challenges translating their creative visions into reality due to engineering limitations. However, with AI tools, they can directly build their designs, creating intricate interactions and unique aesthetics that stand out from generic AI-generated interfaces.

These designers can now make direct pull requests, bypassing traditional hand-off processes and accelerating development. This not only streamlines company workflows but also creates significant entrepreneurial opportunities for designers who can now independently bring their innovative ideas to life.

> Need to build are the things like the building now is done for you. What do you need to be good at figuring out what to build, figuring out if it's great, figuring out problems to solve.

## 1:12:12 - 1:16:14 AI Commoditizes Past Human Competence, Opening New Opportunities

The idea of an AI-driven job apocalypse, leading to mass unemployment, is not expected to materialize. While companies may reorganize and attribute changes to AI, many such shifts are often more related to factors like overhiring. The fundamental pattern observed with new AI models suggests a different impact on the job market.

AI models primarily function by making "yesterday's human competence" cheap and widely accessible. They ingest vast amounts of existing data, allowing anyone to easily deploy previously complex or time-consuming tasks. This leads to a rapid adoption of basic AI outputs, such as generic landing pages or simple written content, which quickly become commoditized because they all originate from similar models and lack unique differentiation.

This commoditization, however, creates new opportunities for human ingenuity. Instead of being replaced, humans are prompted to leverage this readily available "frozen human competence" to create novel and interesting solutions. Because AI models are structurally designed to be compliant and aligned, they inherently trail behind humans who innovate by applying AI to unique and specific situations, pushing the boundaries of what's considered expertise.

This dynamic is evident across various roles. For instance, while AI can automate basic engineering tasks, it doesn't eliminate the need for engineers. Instead, it creates a higher demand for skilled engineers who can discern "slop" from valuable output and integrate AI-generated components effectively into complex codebases, driving the need for more sophisticated problem-solving.

> What a new model drop does, or what models do in general, is they make yesterday's human competence cheap.

## 1:16:13 - 1:20:14 Ride the Models to Stay Relevant in the AI Era

To remain relevant in the evolving job market, individuals must actively "ride the models" by integrating new AI capabilities into their work. This involves consistently trying out new AI tools and figuring out how their extended powers can enhance current tasks. While ignoring AI might seem like a rational response due to fear or uncertainty, embracing these technologies is crucial for being part of the future of work.

Riding the models is not about a specific, static action, but a continuous process driven by curiosity and playfulness. As new AI models emerge, it's important to experiment with them, applying them to your job or personal projects. This means frequently re-evaluating what models couldn't do before, as their capabilities improve rapidly, akin to "turning over rocks" to discover new functionalities.

A significant advantage of the current AI landscape is its broad accessibility. The true "edge of AI" isn't limited to tech hubs like San Francisco where models are built, but rather wherever a human creatively applies AI to real-world problems. Most advanced AI models are available to almost anyone immediately upon release, regardless of financial resources, allowing individuals globally to be early adopters and innovators in how these tools are used.

> The only thing you need to do is ride the models, and that means use them for whatever it is that you do.

## 1:20:14 - 1:24:15 The Nuanced Reality of AI: Both Change and Continuity

AI's broad accessibility stems from its development within Silicon Valley culture, which prioritizes making intelligence "too cheap to meter." This contrasts with how a company like IBM might have commercialized it, potentially leading to a much more restricted and expensive tool. This open approach has facilitated the growth of some of history's fastest-growing companies.

The impact of AI presents a paradox: while fundamental work elements like SaaS, email, and Slack continue largely unchanged, specific professional roles are undergoing significant transformation. Engineers are less focused on writing code, product managers are less on PRDs, and designers are adapting their processes.

Intuitions about the AI future often swing to extremes—either a utopia or an apocalypse. A more realistic view acknowledges a complex blend of profound change and surprising continuity. It suggests that while new horizons bring exciting and challenging elements, many aspects will remain similar, challenging simplistic narratives about a completely new world.

> There's some really cool things, there's some not cool things, and it's just another horizon.

## 1:24:15 - 1:26:16 Find Joy and Solve Problems with AI to Overcome FOMO

Many people approach AI driven by a fear of missing out (FOMO) rather than genuine interest. To truly engage with AI and discover its useful applications, it's more effective to find enjoyment in using these tools. This shifts the motivation from anxiety to positive, active engagement.

The key is to discover your personal 'moment of joy' with AI. As Nikhil Singhall described, this refers to an instance where an AI tool genuinely impresses you with its capability and usefulness, sparking excitement and a desire to explore further. This personal experience becomes a strong driver for continued learning and building.

A practical way to find this joy is by identifying a real problem in your daily life or work and actively seeking to solve it using AI. This hands-on, problem-driven approach often reveals AI's power and potential in a tangible way, moving beyond theoretical understanding to practical application.

> you gotta find your moment of joy with AI.

---

Get podcast briefs for shows you follow: https://podbrew.app/