What a Year of Building Alone Taught Me About AI’s Real Promise

A piece made the rounds recently with a bold claim: AI is pushing us from a top-down economic world toward a decentralized, network-shaped one, and the people running our biggest institutions know it. According to the article, that’s the real reason behind a lot of today’s AI-safety alarm. Executives and policymakers sense the old hierarchies are vulnerable, so they dress up self-preservation as caution. I’m not going to link to the piece, because that claim about people’s hidden motives isn’t something the article actually proves. But strip that part away, and one idea underneath it is worth taking seriously on its own: AI may be lowering the scale at which sophisticated economic activity becomes possible, shifting real work away from large organizations and toward individuals, small teams, and networks.

That sent me back to a book I read years ago and hadn’t thought about in a long time: Murray Bookchin’s Post-Scarcity Anarchism. Bookchin’s argument was more careful than “technology creates abundance, therefore freedom follows.” He was explicit that advanced technology creates a potential — it could reduce drudgery and support smaller-scale, decentralized production, or the same technology under centralized control could just as easily deepen hierarchy and bureaucracy. The technology doesn’t decide which. The social and economic arrangements built around it do.

That’s been sitting with me for weeks now, because I think I’ve spent the past year living inside exactly that uncertainty: building something that depends entirely on tools I don’t control, without having named what that dependency actually means until now.

What I actually built, and what AI actually did

I’ve spent the last year designing and building The Agile Roadmap — a coaching program for professionals navigating AI-era career disruption. Over that year I’ve done curriculum architecture, instructional design, market research, positioning, copywriting, WordPress development, CRM design, workflow automation, webinar development, AI integration work, documentation, a fair amount of security review, SEO, research synthesis, and more rounds of critique and revision than I could count.

Twenty years ago, getting something at this level of sophistication off the ground would have meant either serious capital or a team of specialists I didn’t have. I’m one person, working mostly alone, and I got further in a year than I think I honestly expected to.

I want to be precise about what AI did and didn’t do in that process, because I think the imprecise version of this story is exactly what’s fueling a lot of the current hype in both directions. AI didn’t generate the substance of what I built. It didn’t have the years of engineering and coaching background that told me which technical suggestions were bad, which conceptual pieces didn’t hold together, or which direction was worth pursuing when the first attempt wasn’t right. What it did was radically shrink the distance between having an idea in my head and having something real enough to look at, test, and argue with. A rough draft of a workflow. A first pass at a piece of copy. A prototype I could actually click through instead of just describe. That compression, repeated across a dozen different disciplines, is what let one person operate across all of them before needing to bring in specialists for the parts that genuinely required them.

Checking the feeling against the evidence

I didn’t want to just trust that impression, though. So I went looking for whether the research backs up what it felt like from the inside.

A field experiment at Procter & Gamble, published this year in Organization Science, split 776 professionals into four groups working on real product innovation problems: individuals with AI, individuals without, two-person teams with AI, and two-person teams without. The headline finding was the one that mattered to my situation directly: on the innovation tasks studied, individuals working with AI performed at a level comparable to two-person teams working without it.

An earlier, widely cited study of 5,179 customer support agents found generative AI raised productivity by 14% on average — but that average hid a much bigger story. Novice and low-skilled workers improved by 34%. Experienced, highly skilled workers saw comparatively little benefit. The tool wasn’t making already-capable people modestly faster. It was handing less experienced people access to something closer to what the most capable people already knew.

A 2026 randomized experiment pushed on that same question more directly, and deliberately outside the context of any one company. Researchers gave 1,174 adults with different levels of formal education an incentivized, workplace-style business problem to solve. Without AI, participants with more formal education outperformed those with less by a substantial margin. With AI, that gap shrank by roughly three-quarters. Worth being honest about the limits here too: the underlying difference didn’t vanish. When the researchers took the AI assistant away again in a follow-up round, a real gap came back. AI compressed the difference in what people could produce; it didn’t erase the difference in what people actually knew.

That last point matters more than it might seem, because it’s the honest version of a claim I initially wanted to make too strongly.

Where I’d temper my own optimism

The OECD has been tracking this at the level of actual businesses, not lab experiments, and the picture is more mixed than “AI democratizes everything” would suggest. Adoption among small and medium enterprises is climbing fast, but as of the most recent surveys, still lower than adoption among large firms by a wide margin, and most small businesses using generative AI report using it for isolated tasks rather than anything close to their core operations. The tools being available and the tools being used well are two different things, and right now there’s a real gap between them.

There’s a newer, more pointed finding I think is the most important corrective to my own instinct here. Research tracking business formation since generative AI became widely available finds a genuine surge in solo founders — new companies with fewer shareholders and smaller founding teams than before, consistent with AI substituting for some of what a founding team used to provide. But the same research finds a growing divergence between the quantity of new ventures and their quality. More people can start something. That hasn’t translated into a proportional increase in ventures that actually succeed. The researchers describe this as a shift in where the real bottleneck sits: not at the starting line anymore, but in the development, coordination, and refinement it takes to turn a viable first attempt into something genuinely good.

That tracks with my own experience more than the more triumphant framing does. AI got me to a working first draft of nearly everything on that long list above, fast, and it did real work beyond drafting too — surfacing alternatives I hadn’t considered, pressure-testing an assumption, synthesizing a pile of research into something I could actually evaluate. What it didn’t do was decide which first draft was worth keeping, which was quietly wrong in a way that would have caused real problems six months later, and which needed to be thrown out entirely and rethought. AI could hand me the options and the critique. The responsibility for which one was actually sound, coherent, and worth pursuing stayed mine.

The paradox underneath all of this

Here’s where Bookchin’s distinction earns its keep. At the same time capability is spreading out to individuals, the infrastructure underneath that capability is concentrating hard. The most recent OECD analysis of AI markets puts a specific number on it: as of 2023, the three largest cloud providers accounted for 74% of the global cloud computing market, and Nvidia alone accounted for 90% of the GPU market that trains and runs the models the rest of us build on. Foundation model development has enormous fixed costs, and that tends to concentrate power in a small number of hands almost by default.

So the honest picture isn’t decentralization in any simple sense. It’s something closer to: a small number of enormous companies controlling extraordinarily powerful infrastructure, and, simultaneously, an increasing number of individuals and tiny teams using that infrastructure to do things that used to require organizations far larger than themselves. Those two things aren’t in tension. They’re happening at once, and there’s no guarantee either one wins outright.

That’s actually the more interesting and more useful version of the claim, I think, than the stronger one I started with. It’s not that AI is inherently decentralizing. It’s that AI is lowering the organizational scale at which sophisticated work becomes possible, even while the infrastructure it runs on keeps consolidating. The closer comparison isn’t really the early internet, with its scrappy, genuinely distributed early years. It’s the shift from mainframes to personal computers: the big centralized machines didn’t go away, but the locus of agency moved a lot closer to the individual sitting at a desk.

What I actually take from this

The real shift isn’t just that execution got cheap and judgment got scarce, though that’s the version I started this piece wanting to write. It’s that cheaper execution and cheaper experimentation raise the value of everything that has to happen around them: judgment, synthesis, domain knowledge, taste, the ability to hold a lot of context at once, and the responsibility for actually deciding. AI has made a wide range of execution cheap. What’s left scarce is the person who can look at twenty AI-generated options and recognize which one is actually right, pull the handful of distinctions that matter out of a mountain of research, tell a good first draft from a dangerous one, and keep revising something long after the exciting part is over.

Part of that judgment is knowing when to change your mind. The education-gap study offers a clue here, even if it doesn’t spell out the exact mechanism: the paper found that higher-education participants didn’t just get more from AI because of what they already knew, they used the tool more effectively across several measures the researchers tracked. My own guess, based on nothing more than a year of doing this daily, is that a lot of that difference in effective use comes down to understanding when you need to test an AI answer against what you already know and push back when it doesn’t hold up, rather than accepting the first plausible-sounding response. That’s not a new skill AI created. It’s the old discipline of revising an assumption when the evidence warrants it, and it matters more now precisely because AI can generate plausible-looking answers faster than most people can check them.

I don’t think that’s a comforting story in the way “AI democratizes expertise” sounds comforting. It’s a more demanding one. The tools will meet you further out than they ever could before, if you already have something worth building and the judgment to steer it. They will not supply the judgment itself, and the research on quantity versus quality in AI-assisted entrepreneurship suggests that gap is exactly where a lot of newly launched ventures are currently failing to close it.

A year in, if someone came to me with a big idea and not much else, here’s what I’d actually tell them: the distance between having a good idea and having something real enough to test it against the world has genuinely collapsed, more than at any point in my working life. That part is not hype. But whether what you build is actually worth building was never the part AI was going to do for you, and it still isn’t.

Sources

This piece drew on the following research:

Fabrizio Dell’Acqua, Charles Ayoubi, Hila Lifshitz, Raffaella Sadun, Ethan Mollick, Lilach Mollick, et al., “The Cybernetic Teammate: A Field Experiment on Generative AI Reshaping Teamwork and Expertise,” NBER Working Paper 33641, published in Organization Science, June 2026.
Erik Brynjolfsson, Danielle Li, and Lindsey R. Raymond, “Generative AI at Work,” NBER Working Paper 31161, published in The Quarterly Journal of Economics, Vol. 140, No. 2, 2025.
Guillermo Cruces, Diego Fernández Meijide, Sebastian Galiani, Ramiro H. Gálvez, and María Lombardi, “Does Generative AI Narrow Education-Based Productivity Gaps? Evidence from a Randomized Experiment,” NBER Working Paper 34851 / CEPR Discussion Paper 21299, 2026.
OECD, “Artificial Intelligence Markets,” 2026. Source for the cloud and GPU concentration figures, citing Gambacorta and Shreeti, 2026.
OECD, “The Effects of Generative AI on Productivity, Innovation and Entrepreneurship,” 2025.
OECD, “AI Adoption by Small and Medium-Sized Enterprises,” 2025, and “Empowering SMEs in the Age of AI,” 2026.
OECD, “Artificial Intelligence and Competitive Dynamics in Downstream Markets,” 2025. Source for the large-firm versus small-firm AI adoption gap.
Research on business formation and venture quality since the introduction of generative AI, “Generative AI Fuels Solo Entrepreneurship, but Teams Still Lead at the Top,” 2026.
Murray Bookchin, Post-Scarcity Anarchism, 1971.

I use AI tools to help research sources and draft these posts. I review, edit, and take responsibility for everything published here, including any errors.