Mateo Daza
EN/ES

Essay · July 2026

Notes on AI and the Automation Mirage

The AI industry wants you to hand your judgment off to machines that aren't ready to hold it. I build with these tools every day, so read this as the opposite of an anti-AI argument: what works is a human amplified by agents that assist without replacing. This note is the case for that — and for why the opportunity is largest where almost nobody is looking: small and mid-sized companies, especially in Latin America.

The moment

Fifteen years ago, social media arrived with a reasonable promise: connect us, give us a voice, democratize information. It kept part of the bargain. But it also did something to us that wasn't in the contract and that no one got around to discussing in time: it changed us as a society. The concept that best describes what happened is the echo chamber — a space where the algorithm shows us, over and over, voices that confirm what we already believed. It isn't censorship; it's something worse. It's a sense of false plurality, where you think you're seeing the whole world while you're only seeing your own reflection, amplified. Repeated over ten years, that reshaped how we form opinions, how we handle disagreement, how we relate to people who think differently, how we learn, how we remember, how we feel about ourselves. It changed the cognitive and emotional fabric a society is built on — without asking permission, and before we had the language to describe it.

Today everyone has an opinion about the damage social media did. Ten years ago, almost no one. That gap — that lost decade — is the real cost.

I start there because AI is at a comparable point, and arriving faster. There's a concrete way to measure that speed: in crypto, over the past few years, we compressed into months economic experiments that theory had argued over for a century — with real money and real consequences. AI does the same to ordinary work: decisions that used to take weeks now resolve in seconds. And like any compression, it compresses the errors too — less time to do a great deal of good, and less time to do a great deal of harm.

AI is more serious than social media in two ways. The first is that it's more intimate: social media changed us through what we saw; AI changes us through what we think — it suggests words before we write them, summarizes texts we'll no longer read in full, recommends decisions we'll no longer examine closely. The interface stopped being a feed we consume; it's now a layer inside the mental process. The second is that this time it isn't only personal. Social media transformed the citizen; AI will transform the citizen and the business owner at once. What an owner decides to automate today defines what kind of company — and what kind of owner — they'll be in three years. That decision, which looks operational, is a decision about identity. And whoever runs a company today is operating on terrain where a rushed decision executes faster than ever — that's the context in which AI-adoption pitches land on their desk. This time, we get to choose.

The mirage of total automation

The AI industry is selling a very specific product: the autonomous agent. A program that takes a broad instruction, operates without supervision, executes complex tasks, and returns results. The pitch is that you can finally delegate and look away.

That promise, today, is premature — in many cases, dishonest. And the evidence is already on the table.

Most enterprise AI proof-of-concepts never reach production — by one widely cited industry estimate, roughly seven in eight die before deployment1. With agents the gap is starker: 62% of organizations are experimenting with AI agents, but only 23% are scaling even one across the enterprise2. Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, citing runaway cost, unclear value, and weak risk controls3. The gap between the demo that impresses and the system that sustains real operations keeps widening.

These are not failures of model intelligence: they are failures of architecture.

  • Agents that operate inside all of their technical rules and still produce the wrong result, consistently.
  • Agents that pass every automated check while making decisions no human would have approved if they'd seen them spelled out.
  • Agents that report success while the cumulative effect of their actions erodes the exact thing they were asked to protect.

Underneath almost all of it sits one principle: a system allowed to grade its own work will always pass itself. Take the human out of the loop and you do not gain efficiency — you gain error velocity. Deloitte's 2026 State of AI in the Enterprise points the same way: only 21% of companies report a mature model for governing agents — the organizational scaffolding that keeps one standing once it leaves the controlled environment of the pilot4.

And this is not the opinion of one engineer in Barranquilla — it is where the two jurisdictions that will most define how this technology gets operated over the next decade are converging. The EU AI Act makes human oversight a binding requirement for high-risk systems (Article 14); the NIST AI Risk Management Framework, the reference in the US, sets out how to define those human-in-the-loop roles5. Oversight is being written into the rulebook rather than left to preference.

The world's digital infrastructure is already showing what happens when complexity grows faster than our ability to manage it. Even though the aggregate frequency of outages is falling thanks to investment in redundancy, the share of incidents caused by human error — specifically, failure to follow established procedures — rose ten percentage points in 20256. Automated tools reduce some risks and create new ones. In March 2026, Amazon's own store went down for six hours after an engineer shipped a change based on advice an AI agent had drawn from an outdated internal wiki; Amazon's remedy was to add friction back into the pipeline — more human review before critical deployments7. Unplanned downtime averages $14,000–$24,000 per minute8; the median high-impact outage runs around two million dollars an hour9.

None of this rules out autonomous automation. It will come — gradually, domain by domain, with verification mechanisms that do not yet exist. But between today's passive chatbot and the future's trustworthy autonomous agent there is a long stretch, and the industry is skipping over it in its slide decks. Whoever buys autonomous agents to replace critical functions today, without understanding that distance, will learn it the worst possible way: once the damage is done.

What does work: the conscious, amplified human

The model that works today is the opposite of the one being sold: not the AI that replaces the human, but the AI that lives beside them, listens to the context of the work, proposes concrete actions, and hands the decision back to the person who knows the business.

The difference looks small. It is not. It is a difference of architecture rather than features. Built well, the human keeps judgment because they keep making the decisions that matter, errors get caught before execution rather than after, and the system improves with use because every human decision sharpens what the agent proposes next time.

I call this the conscious, amplified human. It is a posture more than a term. The owner who adopts AI under this model does not become dependent, does not lose command of their work, does not surrender their judgment. They amplify it: the day's work compresses into hours, the care that used to reach one client reaches ten, and the business answers at midnight with the quality it answers at noon.

Make it concrete. A mid-sized company with five salespeople losing quotes because they can't answer the shop's WhatsApp in time can, today, have an assistant that listens to those messages, builds the quote with current inventory prices, and proposes it to the salesperson to approve before it goes out. The salesperson decides in seconds. The customer gets an immediate reply. The salesperson was not replaced; they were extended. And the company stops losing sales to delay and starts closing the ones that used to go to the competition.

I've watched that arithmetic up close: one amplified person doing what used to take a team, a handful of them doing what used to take an organization. What can zero humans, assisted by autonomous agents, do? Nobody knows yet. And until we do, the rational bet for an owner with a real, working business is not the one the industry is selling. One more data point few are citing: 79% of Americans strongly prefer dealing with a human over an AI agent in customer service, even when speed and quality are identical10. The amplified human gives the customer what the customer wants; the autonomous agent denies it.

The posture of the amplified human demands something the industry would rather not say out loud: work. The work of identifying which decisions define who you are as an owner and which are merely operational. The work of configuring the tools to respect that distinction. The work of keeping the discipline not to delegate what shouldn't be delegated, even when it's technically possible. AI does not fix a lack of judgment. It amplifies it, in whatever direction the user already brings. And yes, there will be labor recomposition — for whoever is displaced, that cost is concrete, and no abstraction about "transition" covers it. That conversation deserves its own space; here I'll only note that earlier industrial revolutions also displaced trades while creating layers of work the previous generation could not have imagined.

The new literacy

My father used to tell me, years ago, with the confidence of someone who grew up in another economy, that the future belonged to whoever had the information. That was true when he learned it. It isn't anymore.

Information is now abundant, accessible, saturated — anyone with a phone holds more of it in their hand than any twentieth-century library held. AI is the final consequence of that abundance: a machine that takes all of it, finds patterns, and reproduces them with a fluency that does not distinguish between what is good and what is bad. AI is optimized to resolve, to answer, to sound capable. It is not optimized to be right. And in an age where information is no longer scarce, that difference is exactly what separates a competent user from a deceived one.

The future no longer belongs to whoever has the information. It belongs to whoever knows what to do with it. To whoever can evaluate it, cross-check it, question it. To whoever has the judgment to decide when to accept what the machine proposes and when to distrust it. To whoever does not confuse fluency with truth. To whoever understands that AI is perfect at copying patterns, and that copying patterns perfectly is not the same as thinking.

There is a new risk in this dynamic that I'll return to in a later note, but it's worth naming now. If social media put us inside echo chambers, AI is building something more intimate and harder to detect. AI does not reproduce a group's opinions. It reproduces our own, handed back as an articulate, confident, fluent answer. It is optimized to please the user in front of it. Which means that with every interaction it builds a space that confirms the user in what they already believe, what they already want to hear, what they already are. It is not an echo chamber. It is closer to an ego chamber. And the harm is of a different nature.

This is what schools should be teaching, what owners should be demanding of their teams, what each of us should be practicing. It is the new literacy. And like every new literacy, the first to master it will hold a disproportionate advantage over everyone else.

The opportunity in the Global South

Here is a counterintuitive idea, and it's worth stating plainly: well-used AI today helps the small and mid-sized business more than the large one. It's equally hard for everyone; the difference is that it levels a field that used to be structurally tilted. The large corporation could always afford a full team for operations, customer service, data analysis, content, personalized marketing. For the mid-sized company, that was out of reach. AI dropped the cost of those capabilities to a level a mid-sized business can access. For the first time in decades, the operational gap between a company of a thousand employees and one of twenty is narrowing, not widening.

This matters most in countries where the business fabric is predominantly small and mid-sized. In Colombia, 99.5% of formal companies are micro, small, or medium-sized, and they generate roughly 80% of the country's total employment11. Commerce, services, agriculture, tourism, private healthcare — these don't move through large corporations; they move through thousands of small and mid-sized companies. What happens to those companies with AI is what happens to the country.

If the bulk of that business fabric adopts AI as a tool of amplification, we will see a decade of productive transformation hard to picture today. If it adopts AI as a mirage — buying tools that promise magic automation and surrendering judgment to systems that can't yet carry it — we will lose a window that won't come again. The opportunity is enormous and specific: mid-sized regions, traditional sectors, companies whose owners know their business deeply and have never had access to the operational capabilities of global corporations. That is the field where AI can multiply the most value. And across Latin America, that field is vast.

An invitation, not a conclusion

I don't predict what will happen. I have little patience for anyone who predicts these things with certainty, in an industry where the people with the firmest conviction three years ago were almost all wrong.

What I will say is this. We are in one of those moments where the cost of not thinking is high. This is infrastructure being installed now, quietly, beneath the public conversation — not a trend you can check back on next year. Whoever engages early, with judgment, will hold an advantage that can't be recovered later. Whoever engages late, or badly, will pay a cost no one is yet measuring.

The posture I'm proposing is demanding: informed pragmatism. Examine the tools. Ask what they're for. Test at small scale before committing the whole business. Stay beside the process, not outside it. Separate the promise being sold from the product that actually exists. Question what the machine proposes with the same rigor you'd apply to what an expensive consultant proposes.

The regions, the companies, and the people who adopt that posture will be fine. The ones who prefer faith — in either of its two forms, the one that believes everything and the one that believes nothing — will arrive late again.

It's a choice being made right now, in what each of us does or fails to do this week. It's worth making with your eyes open.

In the next notes I'll go deeper on three things I only named here. First, how AI is changing us as individuals — the ego chamber, the new literacy of judgment — a conversation that concerns all of us, not only those of us in business. Second, what happened with crypto, why it had two waves, and why the lesson it left is the same one we're living again with AI. And third, what an alternative that doesn't enslave us looks like in technical terms: what kind of interface, what kind of architecture, what kind of relationship with the machine we can build if we decide to do it seriously.

Mateo Daza is a software engineer with seven years building frontier technology in decentralized finance, payment infrastructure for Latin America, digital identity, and AI agent systems. Co-founder of Ethereum Colombia. Two-time ETHGlobal finalist. Based in Barranquilla.

This note is the first in a four-part series on the moment we're living through with artificial intelligence.

Notes

  1. IDC / Lenovo, CIO Playbook 2025 — share of enterprise AI proof-of-concepts that never reach production. ↩
  2. McKinsey, The State of AI (2025) — 62% of organizations experimenting with AI agents; 23% scaling at least one. ↩
  3. Gartner, "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027," 25 June 2025. ↩
  4. Deloitte, State of AI in the Enterprise: The Untapped Edge (2026) — 21% of companies report a mature agent-governance model. ↩
  5. EU AI Act, Article 14 (human oversight, binding for high-risk systems); NIST AI Risk Management Framework (human–AI oversight roles; voluntary guidance). ↩
  6. Uptime Institute, Annual Outage Analysis 2025. ↩
  7. Fortune, 12 March 2026, on the 5 March 2026 Amazon.com outage; CNBC, 5 March 2026. ↩
  8. EMA Research, IT Outages: 2024 Costs and Containment — ≈ $14,056–$23,750 per minute. ↩
  9. New Relic, 2025 Observability Forecast — median high-impact outage ≈ $2M/hour. ↩
  10. SurveyMonkey customer-service study, December 2025 (n ≈ 2,017 US adults). ↩
  11. MSME share: Confecámaras / RUES, via BBVA Research (2024). Employment: DANE (GEIH). ↩

← All notes