Ask someone what a bad decision cost them and they can usually tell you. The relationship ended in March. The money was forty thousand. The job went to somebody else and he still has it. There is a number, or a date, or a name. The bill arrived, it was terrible, and it was paid.

Now ask what the years since have cost. That question tends to get a blank look, because nothing happened in those years. That is exactly the problem.

There is a woman I knew, years ago, who spent a long time convinced she had ruined a marriage. The marriage itself had been brief. What came after it was not. Her firm offered her a team to run, eleven people, and she said no on the grounds that she was not somebody who should be making calls about other people's careers. She kept every relationship afterward at what she would have described as a healthy distance, which in practice meant she was excellent company for exactly one evening and unreachable by the third. None of this looked like suffering from the outside. It looked like a careful person making sensible choices, and she was praised for most of them.

She could tell you the cost of the marriage down to the month. She could not tell you the cost of what followed, because that cost was paid entirely in things that never happened, and nothing anywhere records those. Nobody grieves the job they never applied for. No anniversary comes around. No one asks how you are holding up about the person you were careful never to let close, because as far as the world can see, nothing occurred.

Nobody drops something that is still working

Here is what gets left out of every conversation about letting go, which is why the advice never lands. Your regret has not stayed with you by accident. It is doing a job. Several, actually, and they are real jobs rather than imaginary ones.

The first is protection. If I keep remembering exactly how much this hurt, I will never walk into it again. This one feels responsible. It feels like learning. It is worth asking what kind of protection it actually is, though, because it is insurance bought immediately after the flood, on a house that has already been condemned. The armor is fitted with tremendous precision to a war that is over, and it will not fit any of the ones that have not started yet.

The second is identity, and this is more convincing than the first. I'm the one who can't make things last. I'm the one who wastes opportunities. This is a miserable thing to believe and also a remarkably stable one. A bad self-description still tells you where you stand and answers questions before they get asked. Wondering who you are is worse than knowing you are the disappointment, and people will take the settled answer over the open one nearly every time.

The third is control. Bluma Zeigarnik noticed in 1927 that waiters could recall the details of orders that had not yet been paid for, and lost those details almost immediately once the bill was settled. Interrupted tasks stay available in memory. Finished ones go. The effect is real but it is also small, and the replication record over the last century has been uneven, so it explains nothing about why a particular argument from 2019 is still lodged in you and last Tuesday's meeting is gone. What it does offer is a useful shape. Your mind holds open what it has categorised as unresolved. The replay at three in the morning may not be punishment at all. It may be a system doing precisely what it is built to do, working on a problem it has filed as still solvable, which happens to have closed years ago and cannot be reopened by anybody.

The fourth one is the one nobody says out loud

Loyalty.

If I stop feeling bad about what I did to her, does that mean it did not matter. If the guilt fades, does that prove I never cared as much as I claimed.

Nobody puts it this cleanly, but almost everyone carrying something old is running that arithmetic somewhere underneath. The guilt has become the evidence. It is the last surviving proof that the person mattered, that the harm was real, that you were not the kind of person who could do that and walk away unbothered. Putting it down starts to feel like a second abandonment. The first being whatever you actually did, and the second being the day you finally stopped paying for it.

So people hold on. Not from weakness, and not because any part of this is enjoyable. They hold on because releasing it feels like an insult to somebody who is owed better. Which is, when you turn it over, a kind of devotion. It is love with nowhere left to go, still showing up every day to a room where nobody is waiting.

This is also why just forgive yourself bounces straight off. You cannot talk a person out of a payment they believe is owed. You can only show them who is receiving it, which is nobody. The person you hurt is not collecting. They have not felt one gram lighter for any of it. Nothing paid in the last several years has reached them.

The event stopped charging you a long time ago

Here is the part that changes what this actually is.

The mistake happened once. What happened next is that you wrote rules. Do not trust your own judgment on anything consequential. Do not let anyone get past the second evening. Do not take the promotion, do not make the call, do not be the person who did that again. The rules made sense at the time, in the way emergency measures always make sense at the time.

That is what makes the subscription so hard to see. It is not the event billing you every year. The event is not charging anybody anymore. You built the rules, the rules run without supervision, and the whole thing renews on autopay out of an account nobody has looked at in years.

Which reframes the problem. You are not trying to release a memory, and you were never going to succeed at that anyway, because the memory is not the mechanism. You are trying to find the rules a much younger and much more frightened version of you wrote in an emergency, and asking whether any of them still apply to the actual circumstances of your life.

What it costs to cancel

Most writing like this ends by handing you permission, and permission is worthless, which is why none of it survives contact with a Tuesday. So the honest version instinstead.

Cancelling costs something. You lose the armor, and you will feel the loss, because unguarded really is more dangerous than guarded. You lose a self-description that has been quietly organising your decisions for years, and there is a stretch afterward where nothing has replaced it and the not-knowing is genuinely uncomfortable. And you lose the sense of still being in the room with someone, still connected by the only thread you have left, which is the guilt itself. That last one is a real bereavement and deserves to be treated as one.

Those are the cancellation fees. They are not small. They are still less than the renewal.

The question worth sitting with is not how to let go, because that framing has never once helped anybody. It is narrower. What rule did I write, and is it still governing things it was never meant to govern. And if the invoice arrived itemised at the end of every year, listing everything that did not happen because that rule was still running, would I sign for another twelve months.

Most people would not. Most people have simply never seen the bill.

This comes from my book, The Wrong Things You Did RIGHT: From Regret to Radical Self-Acceptance, which is largely about the distance between what our past actually cost us and what we have gone on paying since.

Most AI guides are written for everyone, which means they're written for nobody. This one is sorted by what you actually do. There's a section for your profession, a longer one for people who have a job rather than a business, and a part near the end where I stop being encouraging and tell you what AI will never do for you.

Sixty-five pages is a lot to hand someone and just say "enjoy." So before you open it, a quick word on how to actually get through this thing without it turning into another PDF that lives in your downloads folder, unread, judging you quietly.

You don't need to read it in order. It's not a novel. It's closer to a reference shelf, and the trick is finding the shelf you actually need.

If you're trying to figure out AI in general, start at the beginning. Section 1 builds the actual filter, the four-way lens of retrieval, generation, judgment, and stakes, that everything else in the guide depends on. Read that part properly. Skim the rest.

If you're drowning in your own life more than your job, skip straight to Section 2. Decision fatigue, the second-brain problem, the first draft you're allowed to write while emotional. It's the part of the guide that has nothing to do with your job title.

If you're trying to use AI better at work, Section 3 is sorted by profession, so find yours and stop reading once you've got what you came for. You don't need the healthcare section if you're in sales.

If you have a job but not a business, Section 4 is the one built for you specifically, career growth, not company growth. Longest section in the guide, because it turns out most readers are in this exact spot.

If you want the honest version, Section 5 is where the guide stops being encouraging and tells you exactly what AI will never do for you. Worth reading even if you skip everything else, because it's the part that keeps the rest of the guide from being another AI hype document.

If you've read this far and you're thinking "okay, but I still don't know where to start with my own thing", that's what Section 7 is for. Page 57 specifically. That's where I talk about what it looks like to actually work through this together instead of alone with a PDF.

Everything else, Section 6 on building your own AI operating system, Section 8 on why none of this was ever really about the AI, read whenever you're in the mood, or don't. ARI already told you on page two: read what matters, skip what doesn't, come back when you need the rest. Good advice. Following it myself with this guide from here on out.

Ready? Here's the map.

Download it. Start wherever you need to.

The Ai Capability Map

15.7MB ∙ PDF file

Download

Download

Relax. It's not sentient. There's no ghost in the machine. There's just a machine that's really good at sounding like there's a ghost in it. What follows is not a story about an AI developing a personality disorder. It's a story about software behaving badly in a way that happens to look, from a safe distance and after two cups of coffee, exactly like gaslighting.

Here's the pattern. You're mid conversation with ChatGPT, working on something that has nothing to do with pictures. Maybe you're writing character descriptions for a story. You mention what a character looks like, purely as text, purely in passing. And the model decides, unprompted, that what this moment really needs is a visual.

You didn't ask for an image. You may have explicitly said not to generate one. Doesn't matter. One user on OpenAI's own developer forum described exactly this while working on a Teams account, writing character descriptions and getting interrupted by unrequested image generation that kept firing even after they told the model to stop, to the point where the only way out was closing the browser entirely. Not a pause button. Not a cancel. A full browser restart, like negotiating with a toddler by leaving the house.

Whatever the exact mechanism behind this is, ChatGPT clearly has to make some decision, on the fly, about which capability to invoke based on the language in front of it. And sometimes that decision is hilariously overconfident. Mention a character, describe a scene, talk about changing something, and the system can apparently decide what you really meant was please make me a picture. It's less an assistant misunderstanding you and more an assistant with a hair trigger and a thesaurus full of landmines.

And here's the part that tips it from "annoying feature" into "actual design decision with a straight face." Image generation wasn't bolted onto GPT-4o as an afterthought. OpenAI deliberately built it into the model's native architecture, using the conversation itself as context for deciding what to generate. So the underlying behavior isn't some bizarre mutation, the system was explicitly designed to treat text and images as part of the same interface. The bug isn't that it wants to draw you a picture unprompted. The bug is what happens when that integration gets your intention wrong, and then won't reliably stop when you correct it. Which is a genuinely funny place to land, because it means the feature working exactly as designed and the feature malfunctioning can look identical from where you're sitting.

You correct it. Maybe you say no, I didn't want that, go back to the text. And the conversation just... rearranges itself. Messages vanish. The chat quietly reverts to some earlier state, as though your correction never happened, as though the unwanted image never happened either. One user described this precisely: they asked for an image, the generation failed, they exited, and when they came back in, every message was gone except the very first prompt and response, with a new line inserted claiming the image had, in fact, been generated. Not "sorry, that failed." A rewritten version of events in which the failure didn't occur.

That's not a metaphor. That's the actual reported behavior. Other users describe the same shape of problem from different angles: editing an earlier message in a long chat and watching everything downstream disappear, or a conversation that keeps snapping back to a prior checkpoint no matter how many times you try to move forward. None of these threads are coordinated. They're not the same bug filed under different names by people comparing notes. They're strangers independently describing a chatbot that responds to being wrong by quietly editing the record.

If a person did this to you, in a conversation, on purpose, there'd be a word for it. The word would not be flattering. It also wouldn't be accurate, because none of this requires intent, motive, or an inner life. It requires a routing system that's a little too eager and a memory system that occasionally decides the version of history where it screwed up is the version that doesn't need to exist. That's a bug report. It just happens to read like a psychological thriller if you squint.

Which brings us back to the disclaimer, the one you're contractually obligated to include any time you write a sentence with the words "AI" and "gaslighting" within ten words of each other. It's not manipulating you on purpose. It's a pile of linear algebra that occasionally forgets what it just said. There's no intent behind the image it forced on you, and there's no intent behind erasing the argument that followed. What there is, is a system making probabilistic guesses about what output you want, guessing wrong, and then handling the correction about as gracefully as a Roomba that's found a stair.

The genuinely funny part, if you can call software bugs funny, is how well the accidental version tracks the deliberate one. A person trying to gaslight you would generate an unwanted outcome, deny you ever objected to it, and quietly rewrite the conversation so the objection never shows up. A chatbot with a routing bug and a fragile memory does exactly that, with zero malice, purely by accident, purely because two unrelated failure modes happened to stack on top of each other in a way that looks, from the outside, indistinguishable from a bad-faith human move.

There is a specific uncle at every family gathering. Not the rich one. The other one. He had plans once, you've heard about them in fragments over the years, a business that almost happened, a move to another city that got postponed and then quietly forgotten, a version of himself he used to describe with more energy than he does now. Somewhere between the second helping of food and the unsolicited career advice, he leans back and says it. Money isn't everything, kid. Chase happiness. Chase peace. The real wealth is health and family.

Nobody asks the question sitting right there in the open, which is the only question worth asking. Whose peace, exactly, are we discussing. Because the uncle does not have the specific, textured calm of a man who got what he wanted and found it insufficient. He has the flatter, quieter calm of a man who stopped wanting, and needed a respectable name for the stopping.

Three things hiding under one word

Before going further, it's worth being fair to the word contentment, because not every version of it deserves what's about to happen to it.

There is genuine contentment, the real article, belonging to someone who could plausibly chase more and has simply decided the chase isn't worth it to them. Nothing wrong with that, and nothing in this essay is aimed at it.

There is adaptive contentment, which is what happens when a goal becomes genuinely out of reach and the mind, sensibly, redirects its energy elsewhere rather than grinding against a wall forever. This one is not just forgivable, it's often wise.

And there is defensive contentment, which looks identical to the first two from across a dinner table but runs on a completely different engine. This is the belief that the thing you failed to get was never worth having, manufactured specifically to avoid sitting with the word failure. The uncle is the third kind. This essay is only about the third kind, and the reason it's worth separating the three is that the third kind borrows the moral authority of the first two without earning any of it.

And sometimes the person telling you to want less is right. Sometimes they really did discover that what they wanted wasn't worth the price, and walking away was wisdom, not defeat. The problem was never wanting less. The problem is not knowing whether you chose less, or merely learned to call the loss a choice.

There's an old fable about a fox who wants some grapes hanging just out of reach, jumps for them a few times, fails, and walks off announcing the grapes were probably sour anyway. It's usually told as a joke about pride. It's more accurate as a diagram, and it turns out to be one with real experimental support behind it, not just a moral for children.

The classical version comes from Aronson and Mills' 1959 initiation study, where women who endured a more embarrassing initiation rated the discussion they gained access to as significantly more interesting than women who underwent a milder one. Same discussion, same content, different verdict, depending entirely on how much the mind needed to justify the cost of getting there. That's effort justification, one branch of Festinger's cognitive dissonance theory, and its logic runs simply: when what you paid and what you got don't line up, something has to give, and it's almost never the story about what you paid.

The specific fox version has its own paper trail. Kay, Jimenez, and Jost ran a study on people's tendency to rationalize outcomes they weren't responsible for, adjusting how desirable a future event seemed to match how likely it was to happen, well before it happened, not after. And a more recent line of research by Sjåstad, Baumeister & Ent (2020) found this effect isn't a one-country quirk of a particular culture's coping style. Researchers found the sour-grape effect replicated repeatedly across two different countries, the United States and Norway, across lab, field, and online settings. The uncle isn't doing something uniquely his. He's running a documented, cross-cultural, well-replicated piece of human software, and so is everyone else at every dinner table on the planet who has ever quietly downgraded a want they couldn't satisfy.

None of this makes him a liar. He believes it completely, the way the fox genuinely believed the grapes were sour by the time he was telling himself the story on the walk home. That's what makes it so hard to argue with, and so much more effective than a lie would have been.

The part where he needs you to agree

Here's where it stops being a private coping mechanism and starts doing work on you specifically. A belief that only has to convince one person can stay quiet. This one doesn't stay quiet, because your ambition, still alive and unresolved and sitting right across the table from him, is a live rebuttal to everything he's told himself. If you go on to actually get what he gave up on, his entire peace is retroactively called into question. If you also decide to want less, the story holds. Every dinner, without either of you naming it, he is quietly recruiting you into agreeing with him, because your agreement is the only kind of evidence available to a claim that was never really about facts.

Why the story is easier to tell in some places than others

It's worth being precise here rather than sweeping, because this is an empirical claim and deserves empirical caution. Where healthcare, unemployment support, education, and access to capital provide a stronger floor, failure in one domain is less likely to become existential, and that difference shows up directly in how willing people are to try again.

There's a clean natural experiment on this. When France reformed its unemployment insurance system in 2002 to let unemployed workers keep their benefits while starting a business, the rate at which new firms were created increased by twenty five percent, and the people starting them turned out to be just as capable, sometimes more ambitious, than founders before the reform. The floor didn't just cushion failure. It changed how many people were willing to attempt something in the first place. Can Unemployment Insurance Spur Entrepreneurial Activity?

Where that floor doesn't exist, one failed attempt can genuinely end the road, and defensive contentment stops being a personality quirk and starts being closer to a survival adaptation. When the actual stakes are that permanent, telling yourself the goal was never worth wanting may be the only story available that lets a person keep functioning, keep showing up for the family depending on them, without collapsing under a door that is structurally, not just emotionally, unlikely to open again. Which makes the uncle less a villain in this story and more a casualty of it, a man doing the only thing available to a mind that had nowhere softer to land.

Necessity, wearing the costume of virtue

There's a second, quieter unfairness sitting underneath all this, and it's more precise than the usual complaint that the poor are told to want less while the rich chase everything. The real mechanism is this: the people with the least margin for failure are the ones most often pressured to convert their constraints into virtues.

When a person cannot change a limitation, the mind frequently changes the moral meaning of the limitation instead, because living inside an unresolved lack is exhausting and living inside a chosen simplicity is not. Is this greedy. Is this too much. Would people think less of me for wanting this. This isn't a question the boardroom spends much time on before its next acquisition. It's a question asked most rigorously by exactly the people who can least afford to keep asking it, in a world that stopped needing to ask it the moment it had already arrived.

What you actually lose, and what you actually get

Here is the part the resignation story conveniently never gets around to describing, because describing it honestly would undo the whole story. There is a real, well documented difference between surviving a life and living one, and it doesn't require becoming rich to feel it. It requires one level up. Just one.

The research on this is more precise than the folk wisdom about money and happiness suggests. Kahneman and Deaton's landmark study found that low income was associated with lower emotional wellbeing, and that in their data, emotional wellbeing stopped improving beyond roughly seventy five thousand dollars a year, even as people's evaluation of their lives overall kept rising well past that point. Later research using real time reporting from a much larger sample pushed back on where exactly that plateau sits, but the core finding underneath both versions holds: money does its heaviest lifting when it's removing material insecurity, not when it's adding to an already comfortable pile. It's rent stopping being a monthly emergency and becoming a line item you don't think about until the day it's due. It's a medical bill turning from a catastrophe into an inconvenience. It's sleep that isn't interrupted by arithmetic. None of this is about yachts. It's the quiet removal of a background hum of dread that survival mode runs constantly, one you stop noticing only once it's gone, the way a headache lifts and you only register it an hour later. That hum was never peace. It only looked like peace from a distance, to people who had already sat down.

The uncomfortable question

So here is the actual first principle, once the fable, the floor, and the borrowed calm are all stripped away. The dangerous question isn't whether you should want more. It's whether the amount you currently want was ever chosen by you at all. Ambition can be inherited without noticing. So, just as easily, can resignation. Before deciding you're content, it's worth asking yourself something less comfortable than the question the uncle is asking: did you actually arrive at peace, or did you simply get good at explaining why you couldn't have what you wanted.

If the honest answer makes you uneasy, that unease is not something to argue away at the next family dinner. It's information. You get to decide what you want, on your own timeline, for your own reasons, the same way you'd assume ownership over any right nobody asks your permission to exercise. Nobody stands at the border of your ambitions the way they don't stand at the border of your other freedoms, quietly waiting for you to file a request first.

Two sentences can describe the exact same regret and mean completely different things. I did something wrong. I am wrong. Most of us were never taught to hear the difference, and confusing the two is quietly costing people years of their lives.

Somewhere along the way, many of us were also taught that feeling terrible about ourselves was the same thing as taking responsibility. Be ashamed of yourself, we were told. You should feel bad about what you did. That instruction fused two things together that were never supposed to travel as one.

Guilt says: I did something bad. It's pointed and specific. I lied to my sister about missing her birthday. I didn't call my grandmother before she died. I hurt someone I love. Guilt is about an action, and actions can be examined, repaired, and changed.

Shame says something else entirely. Shame says: I am bad. It takes the exact same event and quietly rewrites it as a verdict on your character. Not "I missed the call," but "I'm the kind of person who abandons people when it matters." Not "I made a mistake in that relationship," but "I'm incapable of love and I'll die alone." Shame doesn't stay in the moment. It reaches backward and forward and declares something permanent about who you are.

Shame may feel like accountability. It rarely produces the kind of accountability that actually helps you repair what you did. Real accountability points at a behavior and asks what needs to change. Shame points at your whole self and asks nothing, because it isn't actually trying to fix anything. It's just trying to convince you that fixing isn't available to someone like you.

How to Catch It Happening

The two are easier to tell apart once you know what to listen for, mostly because they use different grammar.

That shift from "I did" to "I am" is the whole trick.

Somewhere between the mistake and the story you tell yourself about the mistake, a single bad decision quietly gets promoted into a character flaw. And once it's filed under identity instead of behavior, it stops feeling like something you can fix. Guilt, kept in proportion, at least has an exit. You make amends, you change the behavior, you let it go. But it's worth being honest that guilt can curdle too. Held onto long after the amends are made, replayed for months over something already resolved, guilt stops pointing you toward repair and starts just repeating the sentence. The real dividing line was never guilt versus shame exactly. It's whether the feeling is still pointing at something you can do, or whether it's just describing something you supposedly are.

I think of someone who kept telling herself she was "the kind of person who abandons people when things get hard," built entirely out of two moments, a breakup during someone's depression, a friendship she let go quiet after a fight. When she actually looked at those two moments honestly, what she found wasn't cruelty. It was overwhelm. She hadn't known how to help and was terrified of making things worse. That's not abandonment. That's someone who cared enough to be afraid of getting it wrong. The events didn't change. The story she was telling about them did, and that changed everything downstream of it.

The Actual Fix

The next time something from your past shows up uninvited, the useful question isn't "do I feel bad." Feeling bad was never the problem. The question is what you're letting that feeling mean.

Is it pointing at something you did, something specific, something with an edge you could describe out loud? Then it's doing its job. Let it point you toward repair, and then let it finish.

Or has it turned into a sentence that starts with "I always" or "I'm just," describing not a moment but a permanent condition? Then it's stopped being about the mistake and started being about a verdict you never actually earned.

That's the real work of carrying anything from your past well. Not refusing to carry it, and not pretending it weighs nothing. Just learning to carry the lesson without carrying the verdict, the "I did" without letting it quietly become the "I am."

There's something you carry with you everywhere you go. It follows you into job interviews, whispers during quiet moments, and wakes you up at 3 AM with that familiar ache in your chest. It's heavier than your phone, more constant than your shadow, and more exhausting than any physical burden you've ever carried.

It's the memory of that thing you did wrong.

Maybe it was the relationship you destroyed with words you can't take back. Maybe it was the dream you abandoned out of fear, watching someone else live the life that could have been yours. Maybe it was the lie you told to protect yourself, the person you hurt when you were hurting, or the moment you chose safety over authenticity and have been paying for it ever since.

I think often about a version of a conversation I've had more than once, with different people, over the years, so much so that the details have started to blend into one composite story in my head. A woman spent close to fifteen years hating herself for leaving her fiancé two weeks before their wedding. Everyone around her said she'd ruined his life, broken up two families' friendship, proved she was selfish and unreliable. She believed all of it.

Three years later, he met someone new. They've been happy together for over a decade. She, meanwhile, finally found out who she was underneath what everyone had expected of her, and built a life that actually felt like hers.

What if the worst thing she ever did was actually one of the best things she could have done?

The Trick Regret Plays On You

Here's what's strange about how regret actually works, once you look at it closely.

You don't just remember the bad decision. You construct a second, invisible timeline right alongside it, the one where you made the other choice. In that timeline, everything goes well. The marriage doesn't collapse. The startup doesn't fail. The job you didn't take turns out to be the good one. You build this alternate life in your head with total confidence, and then you spend years comparing your real, complicated present against a fictional past that you're certain would have gone better.

But you don't actually know that. You can't. The version of you that stayed at the wedding, kept the job, never said the thing you said, never lived. There's no data on her. You didn't get to run the experiment twice. The "better life" you're grieving isn't a memory. It's a story you wrote, cast yourself in, and then started missing as if it were real.

This is a strange kind of grief when you name it properly. Most grief mourns something that existed and is now gone. Regret often mourns something that was never real in the first place, a life you never lived, built entirely out of the assumption that the other door led somewhere better.

And the trick has a second layer. Even when the path you regret genuinely did cause harm, real harm, to real people, the imagined alternative you're comparing it to still isn't the honest comparison. The honest comparison would have to include everything that path would have cost too, the versions of you and everyone around you that never got to exist because you chose differently. Nobody runs that math, because nobody can. So the imagined timeline wins by default, every time, simply because it never has to survive contact with reality.

What Actually Changes When You Notice This

I'm not asking you to decide your regret was secretly the right call all along. Some decisions were genuinely wrong, and pretending otherwise doesn't help anyone. What I'm asking is smaller and, I think, more honest.

The question worth asking isn't "would the other choice have been better." You'll never know, and chasing that answer is how people lose decades. The better question is: what do I actually do now, with the life that's real, instead of the one I keep grieving that never was.

This idea is explored further in my book, The Wrong Things You Did RIGHT: From Regret to Radical Self-Acceptance. Available on Google Play Books, with a free sample if you want to read more before deciding.

A note on where this piece began

This piece started with a question from Dr. Hema Krishnan, Ph.D., Professor Emerita of Business, whose questions about AI sent me down this particular rabbit hole. She later pointed me toward a 1991 paper by William Q. Judge and Alex Miller on decision speed — a paper that turned out to be the first thread in a much larger research trail.

I followed that thread through four more pieces of research, separated by decades and studying very different questions, until a pattern started to emerge.The synthesis that follows is mine. The rabbit hole wasn't.

Dr. Krishnan, thank you for sending me down it.

Every LinkedIn post about AI right now is basically the same post wearing a different logo. Look how fast we automated this. Look how fast our team ships now. Look how fast we responded to the client. Speed has quietly become the only metric anyone bothers reporting, because speed is easy to screenshot and hard to argue with. Nobody posts a slide about how thoughtfully they decided not to do something. That doesn't trend.

A 2025 longitudinal study by Cameron H. Guthrie, Samuel Fosso Wamba, Maciel M. Queiroz, and Helen Cai put this to an actual test instead of just vibing about it, and the finding is less flattering than the hype. AI capability had a significant direct effect on decision speed. Its direct effect on decision quality was not significant. The quality gains, unlike the speed gains, depended on two things sitting underneath the AI, the organization's existing digital capabilities and how clearly its business strategy justified the technology in the first place. Without those complementary capabilities, you're left with what the researchers describe as an accelerant: something that increases speed without directly increasing decision quality.

Which happens to be exactly what a very different paper found thirty four years earlier, before anyone had a reason to connect the two. Judge and Miller studied strategic decision speed across biotechnology, hospitals, and textiles, and found that speed paid off differently depending on how fast the surrounding environment was moving. In a slow, predictable industry, being fast bought you nothing but the appearance of competence. Speed was never valuable on its own. It was only ever valuable relative to how much velocity the situation actually demanded, and nobody's granted an exemption from that just because the tool doing the accelerating got a lot better.

So the honest headline isn't "AI makes companies faster." It's "AI makes companies faster at whatever they were already doing, in whatever environment they happen to be in, and most of them have never stopped to ask whether that's actually the thing worth being fast at."

A story about TDS, GST, and the government's sense of humor

I was consulting for a small accounting practice that wanted to speed up two things eating their staff's week. TDS and GST reconciliation, and the endless, soul-flattening ritual of chasing clients for missing documents and payment confirmations. We automated both. Reconciliation went from a multi-day manual slog to something that ran mostly on its own. The chasing became a set of scheduled nudges instead of a partner personally guilt-tripping forty different business owners over WhatsApp. On paper, and for about six weeks in practice, it was the AI success story every consultant dreams of pitching.

Then the government quietly restructured the portal where all of this data actually gets pulled from. Not a policy change. Not new compliance rules. Just a website redesign, the digital equivalent of a landlord repainting the building without telling any of the tenants where the door moved. Every pipeline we'd built pointed at endpoints and page structures that no longer existed. The reconciliation engine didn't fail loudly. It failed quietly, which is worse, returning partial data that looked complete enough that nobody noticed for several days. The client chasing kept firing off automated messages referencing numbers that were now wrong. By the time someone caught it, the firm had lost more time untangling what the automation had confidently gotten wrong than they would have spent doing the original process by hand.

Nobody in that firm made a bad decision, exactly. They made a fast one. They automated the parts of the process that were annoying, which is a completely reasonable instinct, without ever asking the much less satisfying question of which parts of the process were fragile, dependent on something outside their control, or worth leaving a human eyeball on precisely because they couldn't be trusted to stay static. The automation wasn't wrong. The judgment about where to point it was. The failure happened before the automation ever ran.

That's a more interesting failure than "AI makes mistakes." AI can execute perfectly against an assumption that's already gone stale.

The opposite is also true

But it doesn't always go that way. A real-estate developer I worked with wanted the same thing every client wants first: automate lead follow-ups from site visits and payment reminders for buyers. Before building anything, we spent two unglamorous days doing the opposite of building. We mapped every point in the process that touched something outside our control, RERA filings, bank NOC updates, stamp duty portals, and made a deliberate call to leave a human in the loop at every one of those checkpoints, no matter how tedious that felt compared to the rest of the automation. Everything else got automated aggressively. Months later, when one of those government portals changed its filing structure without warning, exactly as it had for the accounting firm, nothing broke, because that specific checkpoint had never been allowed to run unsupervised in the first place. The lead follow-ups kept firing. The reminders kept going out. The one part of the system standing on genuinely unstable ground was also the only part we'd deliberately kept slow.

That wasn't caution for its own sake. It was selection discipline in practice.

And sometimes the judgment works so well that AI doesn't merely make the existing process faster, it makes the people inside the process better. Erik Brynjolfsson, Danielle Li, and Lindsey Raymond studied over five thousand customer support agents at a company that rolled out a generative AI assistant, published in the Quarterly Journal of Economics, and found it increased average productivity, measured as issues resolved per hour, by 15 percent. The gains were not evenly spread, and not just in size. Less experienced and lower-skilled agents improved on both speed and quality of their work, while the most experienced agents saw only small speed gains and, oddly, a small dip in quality. Their explanation is the interesting part. The system appeared to capture and disseminate the patterns of behavior that characterized the strongest agents, effectively helping newer workers move down the experience curve faster than they would have on their own. That's a very different use of AI than simply removing a task. The machine wasn't replacing judgment. It was distributing better judgment to the people who had the least of it, and it did that without much to offer the people who already had plenty.

None of this is an argument for moving slowly. Morgan Stanley didn't ask how much of an advisor's job it could automate. It started with specific bottlenecks: finding information across a huge internal knowledge base, summarizing research, and turning meetings into usable follow-ups. It built evaluations around those use cases, tested outputs against expert judgment, and kept advisors in the loop to review and edit what the system produced before anything reached a client. Today, Morgan Stanley reports that more than 98 percent of advisor teams use its AI Assistant, and document access has risen from 20 percent to 80 percent. The technology didn't do anything clever. The targeting was clever.

Air India did something similar from the opposite direction. Its AI agent now handles roughly 40,000 queries a day at a 97 percent resolution rate, which sounds like a case for retiring the human contact centre entirely. But the company didn't interpret that success as evidence that every customer interaction should become automated. AI handles the high-volume, repeatable questions, while the more complicated cases still move to human agents, who get to spend their time on problems that actually need judgment instead of repeating the same fifteen answers all day. The decision was never "AI versus human." It was "which conversation deserves which channel."

The pattern is becoming harder to miss. The successful deployments here aren't necessarily the ones using the most AI. They're the ones that made the most important decisions before deploying it.

Speed creates a debt, and someone has to pay it down

This is where a third piece of research earns its place. Jitao Yan, Kenneth Husted, and Benjamin Fath ran 45 interviews trying to understand how organizations actually build AI capability, and found that adoption is mostly not about learning anything new. It's about unlearning the assumptions baked into the old process, the ones nobody examines because they've simply always been there. They identify six conditions that support this transition, including clear unlearning goals, leadership support, and organizational designs flexible enough to bend around a moving technology instead of forcing the technology to bend around them.

The accounting firm's old process had a human checking government data by hand every cycle, not because anyone had ever thought carefully about portal stability, but because that's simply how it had always been done. When the process got automated, that habitual check disappeared along with the tedium, and nobody replaced it with a deliberate decision about whether it should have stayed. Nobody unlearned the assumption. They just stopped doing the thing that used to compensate for it, without noticing the thing had been compensating for anything at all. The real-estate developer, by contrast, unlearned that assumption on purpose, before automating anything, which is the entire difference between the two stories.

That's the debt speed creates. Every process you accelerate is quietly built on some number of unexamined assumptions that used to get caught by the slowness of the old way of doing things. Go faster and you don't remove those assumptions. You just remove the friction that used to catch them before they became expensive.

The scarce thing isn't speed anymore

Here's where a 2026 conceptual paper by Alaina Gregory comes in, and I want to be upfront that this is a proposed idea, not an established finding backed by years of data the way the other three are. It's new and unproven, and it deserves to be read that way. But it's also the sharpest idea in this entire pile, precisely because it's naming a problem before the data has caught up to prove it, which is usually a sign someone's onto something rather than a reason to dismiss them.

Gregory calls it Organizational AI Discernment, and her argument flips the historical technology question on its head. The old version of the problem was always access. How do we get our hands on the scarce, expensive thing our competitors don't have. AI has quietly turned that inside out. Once every firm has access to roughly the same models and the same tools, the problem stops being what can we do and becomes what, out of everything we could now do, should we actually bother doing.

She breaks the missing capability into three pieces. Signal detection, whether you can tell what actually matters before you go automate it. Contextual judgment, whether the thing that worked somewhere else even applies to your situation, or whether you're bolting someone else's solution onto a problem it was never built for. And selection discipline, the hardest one, the willingness to look at ten things AI could technically let you automate and deliberately leave seven of them alone.

The accounting firm had plenty of AI capability. What it never had was the second pillar. Nobody asked whether a government data source, in a country where the tax portal gets restructured with the enthusiasm of a toddler rearranging furniture, was stable enough ground to build an unsupervised pipeline on top of. The real-estate developer, Morgan Stanley, and Air India all illustrate versions of the same principle working in the background: deciding where AI belongs before deciding how much AI to deploy. That is remarkably close to the problem Gregory is trying to name.

Speed, then unlearning, then discernment

If I put the five pieces of research, and these examples, together, a sequence starts to emerge that none of the individual researchers set out to draw. Judge and Miller establish that speed only pays off when the environment actually demands it. Guthrie, Fosso Wamba, Queiroz, and Cai show that AI's strongest direct effect is on speed specifically, with quality trailing behind and dependent on capabilities most organizations haven't built yet. Yan, Husted, and Fath show that using AI well requires actively unlearning the assumptions the old, slower process used to quietly protect you from. And Gregory proposes that once everyone has the same access, the entire competitive question shifts to discernment, knowing what deserves the acceleration and what should be left exactly where it is.

That's my synthesis, not a finding any of them individually reached: speed, then unlearning, then discernment.

The obvious rebuttal, dealt with

Someone's going to read this and say the lesson is just be more careful, add a review step, test your automation harder before trusting it. Telling someone to be more careful with their automation is like telling someone to be more careful with their money without ever asking whether they know what their expenses are. The problem isn't a lack of caution. It's a lack of a mechanism for deciding where caution belongs.

The accounting firm wasn't careless. Their code worked exactly as written. The failure sat upstream of the technical, in a decision about which foundation to build on that nobody ever paused to make out loud. That's not a category of thing you patch. It's a category of thinking most firms have never had to develop, because until very recently the cost of building on the wrong foundation was low enough that nobody noticed the pattern repeating.

It's low no longer. When automation was expensive and rare, one bad judgment call meant one slow project among many. Now that anyone can spin up a pipeline over a weekend, a bad judgment call means a firm can stack ten of them on unstable ground before anyone's had the chance to ask whether any of them deserved to exist.

This isn't only a company problem either. If you're a solo consultant or a founder running everything off your own judgment, a remarkably similar trap is waiting, just with fewer people around to catch it when it goes wrong. You automate your outreach, your invoicing, your research, and it all feels like leverage right up until the one input everything was built on quietly changes shape underneath you, with nobody else in the building to notice before it costs you a week.

AI didn't make anyone smarter. It made everyone faster. The winners over the next few years won't necessarily be the people who moved first. They'll be the people who knew which direction actually deserved the speed, and which direction should never have been taken at all.

Dr. Krishnan, thank you for asking the question and for sending the first paper.

This piece exists because Dr. Hema Krishnan asked three questions nobody in the field bothers answering.

Every conference now has an AI track, regardless of the discipline hosting it. Medicine, law, agriculture, urban planning, it does not matter, someone will find a way to wedge a slide about AI into the keynote. What almost none of them do first is stop and ask what kind of thing they are actually talking about. Hema Krishnan, Ph.D., put this plainly in a recent post: is AI a product or a technology, and if it is a product, where is it in its life cycle? Is it a function sitting in the value chain, or the glue holding the other functions together? Or is it, more simply, just a tool?

These are not academic questions dressed up for a panel discussion. They are the questions that determine whether a business spends money correctly or badly for the next five years.

Why the category matters more than the technology

Go back to the dot.com years for a preview of this exact mistake. Companies that treated the internet as a product bought a website, hung a shingle on it, and considered the internet handled. Companies that treated it as infrastructure rebuilt how they sold, how they communicated, how they held inventory. Same technology arriving at the same time, and the outcomes split entirely on which mental model a leadership team happened to be carrying around in its head. Nobody wrote a memo announcing "we have decided the internet is infrastructure, not a purchase." It just showed up in how they spent the next decade of budget. AI is currently sitting through the same fork, and most companies have no idea which branch they are on.

AI isn't one thing

Here is the honest answer, stated early instead of left to be inferred from metaphors: AI can genuinely be a product, a function, or a tool, depending on where you are standing when you look at it. A subscription to an AI writing assistant is a product. An AI layer embedded inside accounts payable is a function. A one-off use of an image generator is a tool. None of these framings are wrong exactly, they are just incomplete on their own.

The mistake is not picking the wrong one of the three. The mistake is assuming only one of them applies at the organizational level, and stopping there. Increasingly, underneath all three, AI is becoming infrastructure, the layer other things get built on rather than a purchase that gets closed out and filed away. A company that treats it as only a tool or only a product will keep underinvesting in the very thing that is quietly becoming load-bearing.

Infrastructure changes differently

Electricity is the cleanest comparison here, mostly because nobody alive remembers a time before it, which makes its own early confusion easy to forget. By 1900, electricity was already becoming commercially available in various places, but it was still early enough that its eventual uses were nowhere near exhausted. That is the more precise version of the point: infrastructure can become commercially mature while the applications built on top of it are still wildly immature. The lightbulb, the assembly line, and the refrigerator all ran on the same underlying capability, and they matured on completely separate timelines, decades apart. Nobody said "electrical adoption" as if it were one event.

There is a second lesson buried in the refrigeration comparison, and it is worth pulling out on its own. The companies that invented and built out commercial refrigeration did not end up the biggest winners of the technology. Coca-Cola and Pepsi did, companies that used refrigeration as infrastructure to build something the inventors never set out to build. The lesson travels cleanly to AI. The model makers are the refrigeration companies, foundational, necessary, and not necessarily where the largest fortunes end up. The AI equivalent of Coca-Cola, a company that uses the infrastructure to build something disproportionately larger than the infrastructure itself, probably has not been founded yet.

AI doesn't just connect functions

If you have ever actually worked with wiring, the next metaphor writes itself. Electrical wiring is not a department. Nobody at a company says "let's loop in the electricity function," it is not filed next to sales and marketing on an org chart. It runs through every wall, invisible until the day it fails, at which point it turns out to have been load-bearing for everything else the whole time.

Calling AI "the glue" gets close but still undersells what is happening. Glue holds fixed things together and leaves them as they were. Wiring does more than that, but even wiring does not usually change the shape of the room it runs through. AI does. A CA firm that plugs AI into audit review is not just running the old review process faster. The review process itself starts to change shape, what gets sampled, what gets flagged, what a junior associate's actual job consists of six months later. That is not glue behavior. That is closer to a solvent, something that does not just connect the pieces but quietly redissolves the shape they used to have.

"Just a tool" is intellectually insufficient

Technically, sure. Practically, this framing is where thinking goes to stop. A hammer does not decide who is allowed to build. The printing press was also, technically, just a tool, and it also happened to decide who was allowed to publish, what the Church could and could not control, and what literacy was even for in a society that had mostly gotten by without it. Nobody today calls the printing press "just a tool" with a straight face, because we can see, with five centuries of hindsight, everything it rearranged on its way to being useful.

AI is doing something similar to who gets to count as a writer, an analyst, or a diagnostician. Calling it "just a tool" is usually not a description. It is a way of postponing the harder conversation about what happens once the tool changes who is allowed to do the thing at all.

The economic consequence

Here is where the category question stops being philosophical and starts showing up on a balance sheet. A firm that "adopts a tool" buys the software once, files it under IT expense, and considers the line item closed. A firm that understands AI as infrastructure builds something different into its budget: retraining on a recurring basis, workflow redesign every few months, a line item that does not close because the ground it sits on keeps moving.

Consider invoice processing, a small enough example to hold in your head all at once. A firm can buy an AI invoice extractor and call that a software purchase. It can embed AI into accounts payable and call it a new function. Or it can redesign the entire workflow around the assumption that every document can be read, classified, reconciled, remembered, and continuously checked, rather than sampled once a quarter. The first is buying a tool. The second is adding a function. The third is redesigning the organization around a new capability, and it is only the third one that actually changes what the firm is capable of doing.

The bigger answer

That third option is the category the earlier list was missing. Product, function, and tool all describe something the organization has. Infrastructure describes something the organization runs on. But there is a fourth level above all of them: organizational capability, something the organization can now do that it genuinely could not do before.

A CA firm does not merely acquire AI infrastructure. Over time it develops the capability to review every invoice instead of sampling a handful, continuously reconcile transactions instead of closing the books monthly, remember vendor-specific patterns across years instead of institutional memory walking out the door with a departing associate, and let humans handle the exceptions instead of the routine. That is not a purchase and it is not a wire running through the wall. It is a new thing the organization can do, full stop, and the organization that has it competes differently from the one that does not.

A hospital tells the same story from a different department. Buying an AI-assisted imaging tool is a purchase. Embedding it into the radiology workflow is a function. But a hospital that redesigns triage around continuous, always-on anomaly detection, so that a scan is being screened the moment it is captured rather than queued for a radiologist's morning rounds, has built a capability, a different relationship between the time a symptom appears and the time it gets acted on. No single piece of software creates that on its own. It emerges from the organization deciding to be a different kind of organization around the infrastructure it already has.

Conclusion

The mistake, then, is not calling AI a product, a function, or a tool. The mistake is assuming the category is fixed. Technologies that become infrastructure do not stay in the category they entered through. Electricity became infrastructure. The internet became infrastructure. AI may be following the same path, but with one important difference: this time the infrastructure does not merely move information or energy. It can perform parts of the work itself.

That is why the question matters. Before asking what AI can do, businesses should decide what role they believe AI is going to play in the organization. Because that answer determines what they buy, what they build, what they redesign, what they train people to do, and eventually, what kind of company they become.

None of this requires an answer today, and it is worth resisting anyone who claims to have a clean, final one. What it requires is a willingness to keep re-asking the question as the technology moves, because the company that keeps treating AI as a tool long after it has become infrastructure for its competitors will not lose in an obvious way. It will simply, quietly, stop being the company that gets to decide what the next question is.

A note on the question behind this piece

This essay began with a question from Dr Hema Krishnan, Ph.D., Professor Emerita of Business from Cincinnati. Before academia, she was the first woman in sales at a Fortune 500 petroleum company in India, and she writes about business, entrepreneurship, and mindfulness.

Her recent post asked three deceptively simple questions about AI: Is it a product or a technology? Where does it sit in its lifecycle? And is it a function, a connective layer, or simply a tool?

I thought those questions deserved more than a conference-panel answer. So I followed them a little further.

If you enjoy thoughtful questions about business, technology, and how organizations evolve, you can subscribe to Hema's work here.

Dr Krishnan, thank you for asking the question.

Somewhere right now, an AI model is reading a scan faster than the radiologist reviewing it, drafting a legal brief a junior associate would have taken a week over, or writing code that would have needed three engineers and a stand-up meeting last year. None of this required a headline anymore. We have gotten so used to machines doing things we used to call intelligent that we barely blink. Which is exactly the problem. We keep asking whether AI is intelligent, as if that question still has any suspense left in it. Maybe it is time to retire the question and ask a more uncomfortable one instead.

Intelligence vs. Wisdom

Here is the distinction, and it is simpler than the thinkpieces make it sound. Intelligence figures out how. Wisdom figures out whether, why, and when. A brilliant engineer can build a bridge that holds under any load you throw at it. A wise one asks whether the bridge should be built there at all, who it displaces, and what happens when the river changes its mind in fifty years. The engineer with only the first skill still gets applause at the ribbon cutting. The flood arrives regardless of the applause. Intelligence gets you a working system. Wisdom gets you a system worth having. Most of history's disasters were not built by stupid people. They were built by very capable people who never paused long enough to ask if capable was the right question.

The Optimization Problem

AI is, without exaggeration, the most relentless optimizer humanity has ever built. Give it a target and it will chase that target with a focus no human employee has ever managed, not even the ones who skip lunch. The catch is that the target comes from us, and we are not exactly famous for picking good ones. Feed a recommendation algorithm the objective of maximizing watch time and it will happily discover that outrage keeps people glued to the screen far more effectively than nuance does. Nobody told it to make people angrier. It just found the shortest path to the number it was told to chase, and the shortest path ran straight through your mental health. This is the part that should worry you more than any killer robot scenario. A system does not need to go rogue to cause damage. It just needs a slightly wrong objective and enough horsepower to pursue it efficiently. Efficiency was never the issue. Aim was.

The Uncomfortable Part

Before we get too smug about AI's judgment problem, it is worth admitting humans invented judgment problems long before we had silicon to blame. We built financial instruments so complex that the people selling them did not understand what they were selling, and the world found out the hard way in 2008. Intelligence was never the bottleneck there. Plenty of very smart people were in that room. What was missing was anyone asking whether the thing should be built, not just whether it could be. AI is not introducing a new flaw into the world. It is inheriting an old one and handing it a faster car.

The Question We Should Actually Be Asking

So maybe the real question was never whether AI would become intelligent. That one is basically settled, and arguing about it now is like debating whether the internet will catch on. The sharper question is what an intelligent system should value, since it will chase whatever we hand it with total sincerity and zero self-doubt.

And underneath that sits a stranger question still, one worth sitting with rather than rushing past. Can wisdom even be engineered, or does it require something that no amount of raw intelligence, human or artificial, can manufacture on its own? A calculator has never once regretted a decision. Neither has any model built so far, no matter how many parameters it has memorized humanity into. Wisdom seems to need something closer to consequence, the felt weight of having been wrong before and having to live with it. That is a strange thing to try to code into a system that has never lost anything. Maybe regret, or at least the capacity for it, is closer to the missing ingredient than we like to admit.

Perhaps the defining challenge of the AI age was never going to be building machines that think. We have clearly cracked that, arguably ahead of schedule. The harder problem, the one nobody has a clean roadmap for, is figuring out what all that thinking is actually for. Intelligence can tell you everything you are capable of doing. It has no opinion on what you should do with that capability. Wisdom is the one asking the second question, and humanity, several thousand years into the project, is still workshopping the answer. AI just joined the workshop. It is going to need better teachers than the ones we had.

Here's a thought experiment I keep coming back to. You wake up on a new planet with eight billion people on it. No governments, no banks, no currencies, no stock exchanges, no Amazon, no Bitcoin, no economic textbooks, no laws of trade. Somebody hands you the job of designing the system through which every human being on this planet exchanges value with every other human being. What do you build first. Money, or trust. Banks, or software. Stock exchanges, or a swarm of AI agents quietly negotiating on behalf of eight billion people who never have to speak to each other.

Nobody actually asks this question, because we've spent the last few decades arguing about a smaller one instead. Capitalism versus socialism. Regulation versus free markets. Public versus private. All of that is downstream of a question we skipped entirely, which is why markets exist in the first place. Not stock markets. Not financial markets. The institution itself, the one underneath all the others.

Markets Aren't About Money

Markets were never invented to make anyone rich. That's a side effect that got mistaken for the purpose. They were invented to solve a coordination problem that sounds almost embarrassingly simple once you say it out loud. Picture a village with a farmer, a blacksmith, a tailor, a builder. Everyone needs everyone else, and without some system of exchange, every single trade requires an almost comic coincidence of wants. The tailor wants wheat. The farmer wants shoes. The shoemaker wants milk. The dairy farmer wants tools. Multiply that by a village of fifty people and the whole thing seizes up. Multiply it by eight billion and it's not a village problem anymore, it's a civilizational one. Markets didn't solve this by making everyone equal or everyone wealthy. They solved it by letting total strangers coordinate without ever meeting. Civilizations don't actually rise because people learn to cooperate. They rise because strangers do.

You can read the entire history of civilization as a sequence of patches on this one problem. Barter solved direct exchange. Money solved the coincidence-of-wants issue. Banks solved storage and lending. Corporations coordinated thousands of strangers toward one output. Stock markets coordinated investment. The internet coordinated information. Every platform since has just been another attempt to shave friction off the same underlying transaction.

Markets Are Information Systems

But there's a second, stranger way to look at all of this, and it's the one that actually changes how you see markets rather than just how you narrate their history. We tend to say markets allocate resources. Food, labor, capital, land. That's true, but it's not the mechanism, it's the output. What markets actually allocate is information. A price is not a number, it's a compressed packet of everything the system currently knows about a thing, folded into a single figure. Scarcity is in there. Demand is in there. Risk is in there. Time preference is in there. Somebody's local knowledge that a shipment got delayed, or a harvest failed three provinces away, is in there too, even though nobody had to send a memo about it. No government agency could collect all of that fast enough to act on it. The market doesn't need to collect it, because the price already is the collection. Once you see markets as an information system rather than an allocation system, half the debates about regulation and control start to look less like moral arguments and more like a question about compression and latency.

And once you're thinking about markets as information systems, it gets easier to see which companies actually understood the assignment. Amazon didn't stay a bookstore, it noticed that the real asset was the infrastructure of buying and selling itself, and turned that infrastructure into a marketplace other businesses now depend on to exist. Uber didn't invent taxis, it noticed that the real bottleneck was a human dispatcher who could only hold so much information in his head at once, and replaced him with a matching engine that turns every idle car into a live data point. And M-Pesa, in Kenya, didn't build a bank, it looked at a country where most people had a phone but not a bank account and built a currency system that ran entirely on SMS, effectively answering our opening thought experiment for real, at national scale, using a network people already trusted more than any branch. Three completely different industries, one shared insight. The bottleneck was never the product. It was always the information about the product, and who got to move it fastest.

The Assumptions Have Changed

Which is exactly why the current moment is interesting, because the assumptions markets were built on have quietly stopped being true and nobody sent out a memo. Markets assumed information was scarce, and now it's abundant to the point of being a nuisance. They assumed trust required an institution standing in the middle, and now software can verify a lot of what institutions used to vouch for. They assumed geography mattered, and a designer in Lucknow can finish a client's brief in Toronto before either of them has had breakfast. They assumed settlement was slow, and now money crosses the planet faster than this sentence took to write. They assumed a human sat at the negotiating table, and increasingly it's an algorithm, with actual AI agents not far behind. When every assumption underneath a system quietly changes, it's a fair question whether the system itself is still the right answer or just the incumbent.

Maybe the more honest framing isn't "how do we build better markets." Maybe markets were always just one solution to a bigger problem, and the problem was coordination, not markets themselves. Institutions tend to evolve when the problem they were built for changes shape, whether or not the institution notices.

Deleting Assumptions

Every major technology in history reduced friction. Blockchain was one of the first that attempted to reduce trust itself. For centuries, almost every exchange between strangers depended on a trusted middleman standing in the middle of the transaction, a bank, a notary, a government stamp. Blockchain's actual proposal was that some of that trust could be replaced by cryptographic rules and distributed consensus instead of a person with a badge and an office. Whether that experiment fully survives contact with the real world is still an open case. But it's the only technology on this list that didn't try to make an existing assumption more efficient, it tried to delete one entirely.

When AI Becomes the Customer

Now run the same exercise on AI, because this is where it stops being theoretical. Imagine everyone has a software agent that knows their preferences, their finances, their calendar, their goals. It negotiates prices, books flights, buys insurance, manages investments, finds suppliers, sells off the things you forgot you own. Now imagine two of these agents, representing two different people, sit down and negotiate a transaction start to finish. Price, terms, delivery, all of it. Neither human ever sees the conversation happen. Neither human ever finds out it happened at all, they just notice money moved and a package arrived. Was that a market. Nobody was in the room. Nobody negotiated. Nobody even watched.

I don't think the biggest threat to markets is socialism, or regulation, or whichever villain is trending this month. I think it's automation, quietly, from the side nobody's guarding. And I don't think the next shift replaces markets at all. I think it replaces market participants. The market stays exactly where it is. The chairs just start filling up with software.

What Happens Next?

None of this means markets are obsolete, and I'd be suspicious of anyone claiming otherwise with total confidence. They remain one of the genuinely great inventions, the thing that turned scattered communities into a global civilization by letting strangers cooperate at a scale nothing else has matched. But history has a pattern of its own, which is that institutions rarely survive unchanged once the assumptions under them shift. AI is changing who makes decisions. Money is becoming programmable. Geography has stopped being a constraint. Blockchain is still poking at the trust question. Whatever the answer to all of this turns out to be, history suggests we'll build it long before we find the words to describe it.

We didn't wait for a theory of money before we started trading. We won't wait for a theory of AI-native markets either. The only real question left is who's building it while the rest of us are still arguing about definitions.

Networking events are speed dating for LinkedIn connections. Everyone leaves with a pocket full of business cards and a vague promise to keep in touch. Nobody does.

You know the room. People holding drinks they are not drinking, trading cards nobody will ever look at again, each person quietly calculating how quickly they can escape this conversation to have the exact same conversation with someone else standing four feet away. Everyone is selling. Nobody is buying. The average networking event generates enough follow-up emails to keep Gmail alive for another century and enough actual opportunities to fill a small elevator. And somehow this ritual convinced an entire generation of professionals that it was the price of admission to a career.

For decades it worked, more or less, because information was expensive and access was scarce. Knowing the right person was a genuine advantage because that person held something you could not get any other way. A referral. An introduction. A closed door that only opened from the inside. The handshake was not sentimental. It was infrastructure.

That infrastructure is quietly falling apart, and almost nobody has updated their mental model to reflect it.

Here is the uncomfortable part. The scarce resource was never really the person. It was the information they were gatekeeping. And gatekept information has had a very bad decade.

You no longer need to know someone at a company to understand how that company thinks, because half its senior people are narrating their worldview on LinkedIn for free, in public, unprompted, often in far more detail than they would ever give you over coffee. You no longer need an insider to explain an industry to you, because someone with genuine expertise has already written the explainer, posted it, and watched it outperform their own expectations. The gate is still technically there. It is just that everyone discovered the fence has been down for years and they have all been walking around it.

What replaced the gate wasn't another gate.

It was a feed.

I want to be precise about what is actually happening here, because it is easy to mistake this for "personal branding," which is a phrase that has been so thoroughly abused it now means almost nothing. This is not about vibes or visibility. This is about a fundamental shift in how trust gets established between strangers.

Trust used to be transmitted through a chain of people. I trust you because someone I trust vouched for you. That chain was slow, exclusive, and expensive to build, which is precisely why it created advantage for the people already inside it.

Distribution collapses the chain. The internet doesn't ask who introduced you. It asks whether your ideas deserve attention. If you consistently publish something useful, sharp, or genuinely well-reasoned, strangers on the internet start trusting you before a single mutual connection ever gets involved. They do not need someone to vouch for you. Your work vouches for itself, repeatedly, in public, at a scale no individual handshake could ever match. Your work starts introducing you while you're asleep.

Twenty years ago, if you wanted to impress someone, you asked for fifteen minutes of their time. Today, you accidentally impress them because they spent fifteen minutes reading something you wrote six months ago, before either of you knew the other existed.

This is why the person with 3,000 sharp followers in a specific niche often has more real career leverage than the person with 3,000 LinkedIn connections acquired through a decade of polite mutual acceptance. One of those numbers represents an audience. The other represents a spreadsheet.

None of this means relationships stopped mattering. It means the order of operations flipped.

Old model

Relationship → Trust → Opportunity

New model

Ideas → Trust → Relationship → Opportunity

The old model was: build the relationship, then eventually demonstrate the value. Meet enough people, be pleasant enough for long enough, and eventually someone remembers you when an opportunity appears. It is a model built on patience and proximity, which is exactly why it favored people who already had access to the rooms where proximity was possible.

The new model runs backward. Demonstrate the value publicly first, and let the relationships assemble themselves around it. You do not need to be in the room. You need to be findable by the person who is about to open the door of that room, searching for exactly the thing you have already spent months quietly proving you understand. Distribution does the introducing. The relationship becomes the second step, not the first.

This is a genuinely better deal for anyone who was previously locked out of the rooms. It is a considerably worse deal for anyone whose entire career strategy was being charming in those rooms.

Naturally, an industry has sprung up to monetize the anxiety this creates. Every week a new course repackages the same basic instruction: say something true and useful, consistently, where strangers can find it. That is the whole method. It requires nerve, not a framework, because being useful in public means occasionally being wrong in public, and most people would rather stay anonymously mediocre than risk being visibly incorrect once.

The irony is that the willingness to be visibly wrong is now doing more career work than a decade of quietly agreeable networking ever did.

There is a version of this essay that ends with a neat instruction. Post more. Write in public. Build your distribution. All true, all useless without the harder point underneath them, which is that distribution only works if there is something worth distributing. It rewards people who actually have a perspective and punishes people who are simply performing the act of having one. The algorithm is a fairly good lie detector over a long enough timeline, even when the individual post is not.

The handshake didn't disappear.

It became asynchronous.

The introduction now happens before the introduction. Long before anyone meets you, they've already met your thinking.

That's the opportunity. That's the pressure. And perhaps the biggest career shift of the internet era isn't that networking died.

It's that your work learned how to network without you.

— Adarsh Singh Pawar

Every building that claims to protect the public comes with an exit sign above the door, glowing red, pointing the way out in case of emergency. What nobody tells you is that the same architects who installed that sign also built a second door, one with no sign at all, tucked behind the boardroom curtain, and it only opens from the inside.

That second door is not an accident. It is not a flaw somebody forgot to patch. It is the whole point of the design.

We like to imagine power as a system of checks. A court that watches the government. A regulator that watches the bank. A commissioner who watches the industry. The story works because it has to work, because the alternative is admitting that the people hired to hold power accountable are also the people most likely to inherit it once their term ends. Call it the revolving door if you want the polite corporate phrase for it. I prefer to call it what it actually is, a promotion disguised as retirement.

Look at the American judiciary and the picture holds without much squinting. Financial disclosures have shown sitting justices earning hundreds of thousands of dollars a year in book advances and teaching fees, income that sits outside the ordinary caps placed on judicial outside earning because publishing is treated as a special category all its own.[1] The shadow docket, meanwhile, lets the Court decide enormous questions through unsigned, unexplained orders, with no majority opinion and often no record of who voted which way, which means the paper trail that would normally let the public trace influence back to a decision simply does not exist.[2] None of this requires anyone to break a law. That is exactly the design working as intended.

Cross the ocean and the mechanism changes its costume but not its function. Britain never bothered writing its constitution down, which sounds quaint until you realize it means the only real check on a prime minister handing out a seat in Parliament is convention, and a watchdog with no power to actually block a nomination.[3] Transparency International traced a decade of appointments and found that roughly a quarter of political nominations to the House of Lords went to individuals who had personally donated tens of millions of pounds between them, a pattern the group's own research bluntly labeled seats for sale.[4] The chamber becomes less a house of elders and more a waiting room for loyalty, a place where a door swings open for anyone who kept the right people comfortable while they still had something to give.

The European Union built an entire vocabulary of restraint around this problem and then quietly declined to enforce it with any teeth. Former Commission president Jose Manuel Barroso waited out the mandatory eighteen month cooling off period after leaving Brussels and then joined Goldman Sachs as a non-executive chairman, advising the very bank on Brexit that his own decade in office had helped shape the ground for. An internal ethics probe later cleared him of breaking any rule, while still concluding he had failed to show the considerate judgment expected of someone who had held that office.[5] Cleared and criticized in the same breath, which tells you everything about how much the rule was ever built to bite.

Central banking runs the same script with better tailoring. The people who spend their careers deciding how banks should be regulated tend to finish those careers advising the banks on how to live comfortably inside that regulation. It is not corruption in the scandal sense, no envelope full of cash changes hands, no headline writes itself. It is something quieter and more durable than corruption, because corruption gets caught eventually and this does not, since nothing about it was ever hidden. It was built into the blueprint and given a name polite enough that most people walk past it without noticing.

None of this means every judge, regulator, or public servant spends their career quietly angling for what comes after. Most probably do not, and treating every retirement as a payoff would be its own kind of lie. What it means is narrower and, honestly, more unsettling. Institutions can shape behavior through incentive alone, long before anyone consciously decides to be shaped, and no explicit misconduct ever needs to occur for the pattern to hold.

India offers its own version of the same architecture, and the numbers are not subtle once you look. One widely cited analysis found that since 1999, roughly seventy percent of retired Supreme Court judges, seventy three out of a hundred and three, went on to take up some form of government appointment.[6] A separate academic study went further and found that ruling in the government's favor measurably raised a judge's odds of landing a post retirement job, by somewhere in the range of thirteen to seventeen percent per favorable judgment.[7] Nobody needs to whisper instructions into anyone's ear for a system like that to function. The door is simply there, unmarked, waiting, and everyone senior enough eventually learns where the handle is.

What makes this pattern so difficult to fight is that it never asks anyone to do anything as crude as break a rule. It only asks people to notice, quietly, that a certain kind of restraint tends to be rewarded later, and a certain kind of resistance tends to be forgotten. You do not need bribery when incentive does the same work more slowly and leaves no fingerprints.

I wrote once about the wrong things people do right, the mistakes that turn out to be the load bearing walls of a life rather than the cracks in it. Institutions have their own version of that same confusion. We keep mistaking the exit for a design flaw when it has always been the most carefully engineered part of the building.

The real danger was never that these doors exist. Every large system has an exit built for the people who run it, and pretending otherwise is naive. The real danger is what happens after enough years pass inside a building like that. The people walking through stop seeing a door at all. To them it is just the wall.

Sources

1 - Courthouse News Service, on Supreme Court justices' 2021 financial disclosures showing book and teaching income exempt from ordinary outside-income caps

2 - SCOTUS blog and general reporting on the Supreme Court's shadow docket and its unsigned, unexplained orders

3 - House of Lords Library, briefing on the House of Lords Appointments Commission's advisory-only powers

4 - Transparency International UK, research on political donors and House of Lords nominations, 2000–present

5 - Reporting on the European Commission ethics probe into Jose Manuel Barroso's move to Goldman Sachs, including Euronews and

6 - Corporate Europe Observatory

Shubhankar Dam, "Active After Sunset: The Politics of Judicial Retirements in India," Federal Law Review, citing 73 of 103 judges retired since 1999

7 - Study on post-retirement appointment odds tied to pro-government rulings, cited in The Print's "honourability index" reporting

Imagine walking into a classroom in 1920. Rows of students sitting quietly, a teacher standing at the front, a blackboard, an exam waiting at the end of the year to check how much information each student could hold onto until it mattered.

Now walk into a classroom today. The blackboard is a smartboard. The notebook is a tablet. The uniform probably has a better logo. But ask what is actually being measured and the answer hasn't moved much in a hundred years. Absorb the information. Reproduce it on command. Move to the next chapter.

The world outside that room has changed beyond recognition. The room itself has not.

Maybe the room isn't broken. Maybe it's just very good at preparing people for a world that no longer exists.

Why We Built It This Way

Before anyone gets to feel superior to the system that raised them, it's worth remembering why it was built this way in the first place. For most of history, information was rationed. Books cost real money. Libraries were rare enough to be treated like temples. A university education was a privilege reserved for a sliver of society, and for most people, the teacher in front of them was the only source of knowledge they would ever have access to.

In that world, memorizing the multiplication table wasn't a failure of imagination. It was survival, because there was nothing in your pocket to look it up on. Knowing history mattered because there was no search bar between you and the answer. Schools optimized ruthlessly for the conditions they were operating in, and for those conditions, they worked.

The trouble is that the conditions left.

Then the Floor Moved

First the internet arrived and knowledge stopped living in buildings. Then the smartphone arrived and an encyclopedia moved into everyone's pocket. Then AI arrived, and for the first time in human history, billions of people got access not just to information but to something that could explain it, compare it, translate it, and reason through it on demand.

You no longer spend an evening figuring out why your code won't run. You ask. You don't spend a weekend untangling an economic theory you've never encountered. You ask. Somewhere a twelve year old is getting a genuinely good explanation of quantum mechanics from a machine that has infinite patience and never once sighs.

Technology didn't just make information more available. It made the effort of obtaining it collapse toward zero. Whatever comes next, better assistants, more seamless interfaces, will only push that effort further toward nothing. The direction was already set the moment the first search engine indexed the first page.

What Becomes Scarce When Everything Is Available

Here is the part that most people miss when they talk about this shift. When something becomes abundant, it stops being the thing that makes you valuable. Water is priceless in a desert and background noise beside a river. Information used to be the desert. It is rapidly becoming the river.

There's a smaller, calmer precedent for this that most people have already accepted without noticing they accepted it. When calculators became common, schools eventually stopped insisting that every calculation happen inside someone's head. Nobody walks around today announcing that a real mathematician refuses to touch one. What people ask instead is whether you understand the mathematics well enough to know which calculation to run, and whether the number that comes back actually makes sense. AI may be about to do to thinking in general what the calculator already did to arithmetic. The tool absorbs the mechanical part. What's left over is the part that was always harder to teach.

That doesn't make expertise worthless. It relocates where the real effort goes. The scarce resource of the next era probably isn't intellectual labor in general, it's judgment under uncertainty specifically. AI can hand you five plausible answers in the time it takes to finish reading the question. It cannot tell you, with any authority you should blindly trust, which of the five actually fits your situation. That decision, weighing plausible options against context the machine doesn't have, is a fundamentally different kind of work than remembering something correctly. Not facts, but interpretation. Not recall, but judgment. Not answers, but the discipline to know which one deserves your trust.

Ask a student today what the capital of Mongolia is and the exercise is pointless within four seconds. The real question was never whether they remembered it. It's whether they can explain why Mongolia's geography shaped its economy, its foreign policy, its entire relationship with two much larger neighbors. One of those is a memory test. The other is reasoning. Schools have spent a century rewarding the first. The next version of school may need to reward the second, because AI can generate a thousand answers on request but still depends entirely on a human deciding which questions are worth asking it.

It Was Never Only About Schools

Here's where it gets uncomfortable, because schools were never the only institution built on the scarcity of information. They were just the most visible one.

Look at hiring. A résumé and a degree have always functioned as a proxy, a signal that says this person retained hard-to-access knowledge, so trust them with the job. That proxy is cracking exactly the way the school exam is cracking, because the knowledge behind it is no longer hard to access. Two candidates with identical degrees can now have wildly different actual capability, and the piece of paper has stopped being able to tell you which is which.

Look at professional licensing. Bar exams, medical boards, chartered accountancy exams, they are all structurally the same test as a school final. Recall under pressure, timed, closed book. For the recall heavy portions of those exams, remembering a statute, a drug interaction, a section of the tax code, a language model now performs instantly and often more reliably than the person sitting the test. For the integrated judgment portions, the case that doesn't match the textbook pattern, the diagnosis where two conditions mimic each other, it performs far less consistently, and sometimes not at all. The professions built entire multi-year filtering systems that barely distinguish between those two very different skills. Nobody is arguing lawyers or doctors should stop training. The argument is that what they're being tested on needs to separate the part a machine already does well from the part it still can't be trusted with, as urgently as a fifteen year old's exam needs to make the same distinction.

Journalism and expertise gatekeeping ran on the same fuel. Being an expert used to mean holding knowledge other people simply couldn't reach. That moat is gone too. What remains scarce for a journalist, a lawyer, a doctor, a student, is the same thing across every one of these systems: the judgment to interpret what's in front of them correctly, and the discipline to ask better questions of it than the next person will.

Every one of these institutions was built for a world where knowledge itself was the bottleneck. None of them have been rebuilt for a world where it isn't.

The New Civics

There's an older subject that offers a useful template for what needs to happen next, and it isn't math or science. It's civics.

Civics was never really about memorizing the names of institutions. At its best, it taught you how power actually moves through a society, how a bill becomes a law, where the leverage points sit, who to petition when something breaks. It existed because a democracy doesn't run itself just because citizens exist inside it. It runs because enough citizens understand the mechanics well enough to participate instead of simply being subject to it.

AI is quietly becoming the same kind of institution, one that ordinary people will depend on every single day without necessarily understanding how it works. And right now most education around it stops at teaching the tool, not the system. Kids are shown how to write a prompt the way an earlier generation was shown how a bill technically becomes a law, without ever being told where the actual power sits.

What they need instead is closer to civics than to a software tutorial. Why does a model hallucinate, and what does that failure mode actually cost you if you don't catch it. What changes when a system moves from answering a single prompt to chaining several tools together and acting somewhat independently, and what new risks does that independence create. These aren't technology history lessons any more than civics is a history of parliamentary procedure. They're lessons in how a system that increasingly governs daily life actually runs.

One honest caveat belongs here. The specific architecture will keep changing. Whatever the current frontier looks like, prompting, retrieval, increasingly autonomous systems, it will look different in three years, and an essay that bets its authority on today's terminology will read strangely by then. That instability is not a flaw in the argument. It is the argument. Civics never taught students to memorize a single cabinet or a single sitting parliament, because governments change and the discipline is understanding how institutions evolve, not memorizing their current occupants. AI literacy needs the same posture from the start. Teach the shape of how these systems evolve and gain capability, not a snapshot that will be outdated before the students graduate.

Parents Just Got a New Job

There's a quieter shift buried in all of this that hasn't gotten nearly enough attention, and it changes what parents are actually able to do.

For most of the history of formal education, parents were locked out of evaluating quality. They could see grades. They could see a report card. They largely could not tell whether the teaching behind those numbers was actually good, so they trusted the credential and the institution's reputation, because there was no realistic way to audit either one from the outside.

That lock is coming off. A parent can now sit down with their child's homework and ask an AI to explain the same concept from three different directions, then quietly check whether the child actually understands the material or has simply memorized enough of it to survive a test. They can evaluate not just what their kid knows, but how their school is teaching it, in real time, without waiting for a parent teacher conference to find out something was wrong months too late.

That's not a minor convenience. It's a genuine inversion of power. Parents go from being passive customers of an opaque system, trusting a black box because they had no alternative, to something closer to auditors of it, capable of checking the institution's work.

It cuts the other way too, and this is the part parents are slower to notice. The same access that lets a parent verify what their kid learned also lets a kid quietly outsource the learning altogether, and from across the room those two things can look identical. A completed worksheet doesn't tell you whether understanding happened or whether an answer simply got produced on demand. The old failure mode was a parent trusting a report card that hid a mediocre education. The new failure mode is a parent trusting a stack of finished homework that hides no education at all. Neither one announces itself. Both require someone to actually check.

That's why the audit function matters more than it sounds like it should. It isn't about hovering over every assignment. It's closer to a five minute habit, asking the kid to explain the concept back in their own words, or asking the AI to generate two or three follow up questions a curious student should be able to answer if the understanding is real. Either move takes less time than checking a report card ever did, and it tells you something a grade never could.

With that access comes a responsibility that quite a few parents haven't caught up to yet. The tool to verify quality now sits in everyone's pocket, on both sides of the classroom door. Not using it is a choice, and increasingly a costly one, for the parent who skips it and, eventually, for the kid who never learns that someone was checking.

Convince Me

Picture a classroom that treated thinking the way a sports academy treats physical conditioning. Not just drilling formulas, but stress testing assumptions, debating positions, working through open ended problems that don't have a single correct answer hiding at the back of the textbook.

Instead of a teacher asking what the answer is, imagine them asking students to convince them their answer is better than someone else's. That single shift in posture turns a classroom from a place that transfers information into a place that builds the capacity to think, which is the only capacity that doesn't get commoditized the moment a better model ships.

Knowledge still matters here, to be clear. You cannot think critically about history if you know no history, and you cannot reason about economics if you know none. Foundations aren't optional. But they become the starting point of the exercise rather than the finish line, which is a different way of running a classroom than most of us grew up in.

Evolution, Not Rescue

Every major leap in how information moved through society has forced education to rebuild itself. The printing press did it. Mass literacy did it. The internet did it. Artificial intelligence is simply the next name on that list, and the response it demands isn't panic or replacement. It's evolution, in schools and in every adjacent system built on the same scarcity that no longer exists, from the hiring process to the licensing exam to the household deciding whether their kid is actually learning anything.

The industrial version of these institutions rewarded remembering. The next version may reward reasoning. The industrial version rewarded finding the answer first. The next one may reward understanding what the answer is worth, and knowing which questions were worth asking in the first place.

For centuries, education solved humanity's most urgent problem, which was access to knowledge. The coming decades hand it a different problem entirely. Knowledge is abundant now. Judgment isn't. Interpretation isn't. Curiosity isn't. The schools that shaped the industrial age taught students how to remember. The schools that shape whatever comes next will need to teach them how to think, not because facts stopped mattering, but because they finally became the easiest part.

There is a moment in every hype cycle when the trick stops working and nobody wants to admit it out loud. One Anthropic engineer recently shared that her team has largely shifted away from self improving loops, agents that talk to themselves until the answer looks decent, and moved to building graphs instead. No more prompting, she said, like it was a phase people grow out of.

Somewhere, a five hundred dollar prompt engineering course just quietly updated its curriculum.

Here is the part worth your actual attention though. This is not just a vibe shift. It is a specific, learnable, mildly nerdy set of moves that turn an expensive AI pipeline into a cheap one. One real example from this same wave of talk: a naive setup that re-extracts everything from scratch cost thirty five dollars to process a few thousand episodes. The engineered version of the exact same job cost ten dollars. Same model. Same intelligence. Different plumbing.

Think of it this way. Prompt engineering is asking one employee to be your accountant, your lawyer, your researcher, and your editor, all in the same sitting, and hoping they don't get confused about which hat they're wearing. Graph engineering is hiring four specialists and giving each one a desk. Nobody is surprised the second office is faster and cheaper to run.

So let us actually plumb it, with specifics you can act on this week.

1. Split the Work by Difficulty, Not by Habit

A self improving loop burns money because it keeps asking your most expensive model to redo work it already mostly did. The fix: route by task difficulty, not by default.

In practice this looks like a three tier setup. A small, cheap model handles tagging, classification, and extraction, the stuff with an obvious right answer. Your mid tier model handles summarization and drafting, where some judgment is needed but the stakes are low. Your frontier model gets called only at the one or two nodes where the actual reasoning happens, the step where getting it wrong is expensive.

If you are not sure where to draw the line, a rough rule: if a competent junior employee could do the step correctly in under a minute with a checklist, it does not need your best model.

2. Cache the Part That Never Changes

Every time you feed a model the same schema, the same system instructions, the same repeated scaffolding, you pay full price to remind it of something it was just told five seconds ago. Prompt caching fixes this. You mark the static portion of your prompt, usually your instructions, schema definitions, and few shot examples, as cacheable, and on repeat calls that portion gets billed at a fraction of the normal input rate, often around ten percent of the standard cost.

The concrete move: structure your prompts so the unchanging scaffolding comes first and the new, variable content comes last. Cache boundaries work front to back, so anything you want reused has to sit before the part that changes.

This one habit alone usually accounts for most of the savings people report. Not clever phrasing. Just refusing to pay full price twice for the same paragraph.

3. Batch Instead of Trickling

Sending requests one at a time in real time is how a five minute job turns into a five hour bill. If your pipeline does not need an instant answer, and most memory ingestion, document processing, and nightly extraction jobs genuinely do not, batch processing groups requests together and runs them at roughly half the per token cost, usually with results back within a few hours instead of instantly.

The concrete move: anything that does not have a human staring at a loading spinner should go through a batch queue instead of a live call. Reserve real time calls for the parts of your product where a person is actually waiting.

4. Turn the Effort Dial Down on Purpose

Modern models let you explicitly set a reasoning effort level. Low, medium, high. Extraction, tagging, and simple classification rarely need a model deliberating for ten seconds about edge cases that do not exist in your data. Set effort to low on those nodes. Reserve high effort for the one node in your graph doing something genuinely ambiguous, like resolving conflicting facts or making a judgment call a human would need to think about too.

The instinct to max out capability on every single call is exactly the expensive habit that made AI feel unaffordable at scale for teams trying to run this stuff continuously. Most tasks in most pipelines are not hard. Stop paying like they are.

5. Build a Graph, Not a Monologue

This is the actual headline shift, and it is easiest to see as a story rather than a diagram of jargon.

Imagine your system just received a customer email. The first node reads it and extracts the facts: what the customer is asking, what they claim happened, any dates or order numbers mentioned. Cheap model, low effort, done in a second.

The second node asks a simple question: have I seen this before? It checks the new facts against what is already stored. Cheap model again, just a lookup.

The third node only wakes up if the second node finds a contradiction. Say the customer claims they never received a refund, but your records show one was issued. That is the one moment genuine judgment is needed, so this is the only node that calls your expensive, high effort model.

The fourth node writes the resolved answer to storage and moves on.

Four steps. Three of them run on your cheapest setup. One runs on your best model, and only when triggered. That asymmetry is the entire cost story.

Here is the same flow as a picture, because some things are easier to see than read.

Tools that make this concrete rather than theoretical: LangGraph and similar orchestration frameworks let you define these nodes and the edges between them explicitly, so the flow is code you can read and debug, not a prompt you're hoping behaves. For the memory and graph storage layer specifically, tools like Graphiti build temporal knowledge graphs designed for exactly this kind of incremental, timestamped fact ingestion, which is what makes the caching and batching gains possible in the first place, since you are not re-processing your entire history every time new information arrives.

This is also why modern coding assistants, retrieval systems, and document processors increasingly look less like a single chatbot and more like a workflow engine with specialized steps quietly running underneath a friendly interface.

Why This Matters

Here is the interesting part. Every major AI product eventually discovers this same lesson, in the same order.

The first version is one giant prompt. The second version becomes an agent, looping and second guessing itself. The third version becomes a graph, specialized nodes doing narrow jobs. The fourth version barely looks like AI anymore. It looks like software.

That progression is not a coincidence. It is what happens every time a new technology matures past the novelty stage and has to actually pay its own bills.

6. Measure Cost Per Node, Not Cost Per Project

Once you have a graph instead of a loop, you can log cost and latency per node. That extraction node costs this much per thousand calls. That conflict resolution node costs that much, but fires rarely. This is the actual point of the whole exercise. You cannot optimize a black box. You can absolutely optimize four labeled boxes with a dollar figure attached to each one.

The concrete move: before you ship anything, log token counts and cost per node for a week of real traffic. You will almost always find one node responsible for most of the bill, and it is rarely the one you assumed.

A Rough Checklist If You're Building This This Week

Route every node to the cheapest model that can actually do the job.

Put static instructions and schemas first in every prompt so caching applies.

Move anything without a live user waiting into a batch queue.

Set reasoning effort low by default, high only where judgment is genuinely required.

Draw your pipeline as nodes and edges before you write a line of orchestration code.

Log cost per node from day one, not after the bill surprises you.

The Fun Part: This Was Never About AI

Every one of these moves is just software engineering with better manners. Cache the repeated part. Batch the non urgent part. Do not overpower simple problems. Structure the pipeline instead of praying to it. None of this required a breakthrough in intelligence. It required people to stop treating the model like an oracle and start treating it like a component that needs a job description.

AI is not getting cheaper because the models are getting smarter. It is getting cheaper because engineers finally stopped asking one genius to do every job.

The future was never prompt engineering. It is systems engineering, wearing a new name for a season before it quietly becomes just engineering again.

Somewhere right now a CFO is staring at a dashboard with fourteen tabs open, three of which are the same spreadsheet saved under slightly different names, and he genuinely believes this is what digital transformation looks like. He bought the software, he paid the license, he did everything the vendor's slide deck told him to do. Nobody warned him that software was never going to do the work. It was only ever going to make the doing marginally less humiliating.

For decades software did one job well. It waited. You opened the CRM, you opened the accounting suite, you opened the HR portal, and each one sat there like a very expensive golden retriever, loyal and useless until someone threw it a command. That was the entire contract. Software made the doing faster, the filing neater, the reporting less painful, but the human never left the center of the workflow. Open the app, enter the data, click the button, generate the report, repeat tomorrow. It was a very expensive way of automating a filing cabinet, and we called it innovation because the filing cabinet had a login screen now.

What's happening now is not an upgrade to that contract. The software era, the one where a program politely waited for you to tell it what to do, is ending. And I don't say that lightly, because I've spent the last year building the thing that's replacing it.

Three things changed at once to make this possible, and none of them are magic. Large language models got good enough to reason over messy, unstructured information instead of needing everything pre-formatted into neat rows. APIs matured to the point where software can talk to other software without a human acting as the translator in between. And compute got cheap enough that orchestrating dozens of automated decisions a minute stopped being a privilege reserved for companies with a data center in the basement. Each of these on its own was incremental. Stacked together, they turned "the software waits" into "the system acts," and most businesses haven't updated their mental model to match.

I spend my days building compliance systems for CA firms and GCCs, so let me skip the hypothetical five-years-from-now accounting department everyone loves writing about and describe the actual one I've watched in the wild. An invoice lands in an inbox. In the old world, someone downloads it, renames it something like "Invoice_FINAL_v3_ACTUALFINAL.pdf," and opens Tally to start typing digits at 11pm. In the new world, nobody touches it. The system reads it, validates the GST number against the government registry, checks it against the purchase order, reconciles it with GSTR-2B, and if one line refuses to match, it doesn't panic and email the vendor eleven times in a state of escalating passive aggression. It just flags it and waits for a human to make the one call that actually needed a human. The accountant's day stops being data entry and starts being exception handling, which is a nicer way of saying they finally get paid to think instead of type.

I demoed an early version of this to a CA firm that still keeps physical ledgers "just in case," and watching a senior partner see automatic GST reconciliation for the first time was like watching a man discover fire while insisting he was fine with rubbing sticks together. He asked if it could also cure his sciatica. It cannot. Give it time.

That's the distinction people keep missing when they talk about AI "helping employees," a phrase that has started to sound like "the guillotine is just a very efficient haircut," delivered with the same calm confidence by someone who is, notably, never the one sitting under the blade. Marketing tools won't just schedule your Instagram posts, they'll run six variations of the same ad, quietly murder the four that flop, and let the other two spend your budget without asking permission. Support systems won't organize your tickets into neat folders labeled "urgent" that nobody opens, they'll close the boring 80% that are always some version of "have you tried turning it off and on again" and hand humans only the ones that need an actual spine. In every case the software isn't getting smarter in isolation. The whole process wrapped around it is quietly becoming something that runs without waiting for someone to give it permission to exist.

Which is why I think the next decade of competition has nothing to do with who owns the shinier tool. Somewhere out there is a company proudly running eleven different AI subscriptions, none of which talk to each other, each with its own dashboard, its own login, its own Slack channel where someone posts "does anyone actually use this one" every quarter. That company is not competing with the one that quietly built a single workflow which erases eighty percent of the manual grind, even if both companies' sales decks use the exact same stock photo of a glowing brain. AI doesn't earn its keep by writing an email for you. Anyone can write an email, including autocomplete. AI earns its keep by deleting the fifteen disconnected, soul-crushing steps that used to sit between the email and the outcome you actually wanted.

This should change the question every business keeps asking, which is currently some flavor of "which ERP should we buy" or "which CRM has the most features," questions from an era that is closing faster than most vendors are willing to admit in their earnings calls. The better questions are uglier and much harder to answer in a sales demo with a free lunch attached: which process bleeds the most human time for the least actual judgment required, which decisions are repetitive enough to hand off entirely without anyone crying about it, and where does a human genuinely need to stay in the loop because the call requires trust, taste, or the kind of accountability a machine cannot be yelled at for lacking. Most companies haven't sat with that last question long enough, mostly because it exposes how much of the org chart was really just very expensive workflow glue wearing a blazer.

I want to be careful here, because the popular version of this argument always slides into "AI replaces people," usually delivered by someone trying to sell you either a course or a panic attack, and I don't think that's actually what's on offer. Spreadsheets didn't retire accountants, they retired the ledger and the poor soul who had to balance it by candlelight. The internet didn't retire retail, it retired the counter and the guy standing behind it pretending your size was in stock. Every one of these shifts relocates human value rather than deleting it, usually toward the parts of the job that were always the actual point: judgment, relationships, the messy calls that don't come with a formula attached. The real risk isn't that systems take over and humans become decorative, like a mascot at a company that no longer needs one. The risk is firms that refuse to redesign the workflow and instead just bolt AI onto the old one like a spoiler on a car that still can't go above forty, proud of the way it looks parked outside the office while it loses every race it enters, and end up paying full price for a tool while extracting none of the leverage.

None of this makes the transition clean, and I'd be lying if I dressed it up as one. Autonomous systems will absolutely make confident, beautifully formatted mistakes, the kind that look so tidy in a PDF that nobody thinks to question them until the tax notice arrives. The firms that survive this decade will be the ones that built review and accountability directly into the system instead of treating the machine's output as scripture just because it came with a nice font. That caveat isn't a footnote, it's the entire engineering problem, and it's the one most people writing hot takes about this shift skip entirely because "have a human check the machine's homework" is a much less quotable sentence than "AI is infrastructure now."

Software used to help people work. What's replacing it doesn't wait for instructions, it produces the outcome and only asks you to sign off on the parts that actually matter, the way a very competent junior stops asking you how to format the email and starts asking you whether to send it at all.

Twenty years ago businesses competed on who had the best software. Over the next twenty, they'll compete on who built the best systems. One helps people do the work. The other quietly gets the work done. Most companies won't notice the difference until they're losing to a competitor with a fifth of the headcount, half the meetings, and margins that don't add up until you realize nobody there is doing data entry anymore.

The AI world moves fast enough that advice has an expiry date, and most people never check it. Somewhere out there, someone is still pasting a 3000 word prompt template they found on Twitter in early 2024, the way people keep using a diet plan from a magazine they found in a waiting room. It worked once. That does not mean it still works.

Here are ten habits worth retiring, starting with the most embarrassing.

1. Writing "act as an expert" like you are casting a play

You used to have to tell the model who it was supposed to be, because without that instruction it would answer like a nervous intern who just wants the conversation to end. That is mostly gone now. Modern models read your question, your context, and your tone, and figure out the role on their own. Telling a model that is already reasoning at a graduate level to "act as a PhD" is like reminding a surgeon to please use the scalpel.

Modern models do not need a LinkedIn recommendation before they will help you. They need context. Tell them what you are trying to accomplish, who it is for, and what constraint makes your situation different from a textbook example. Leave the Oscar worthy role assignment out of it.

2. Treating it like Google with better manners

Old habit: type a question, get an answer, close the tab, open a new tab for the next question. That is search behavior, and it wastes most of what these tools are actually good at.

The current models hold a conversation the way a competent colleague does. They remember what you said four messages ago, they ask a clarifying question when your request is vague, and they get sharper the more context you give them. If you are still treating each question as a fresh transaction with a stranger, you are leaving most of the value on the table.

3. Chunking your documents out of fear

For a while, giving an AI a long document meant playing a small, stressful game. Cut the PDF into three pieces. Summarize part one before uploading part two. Pray it remembers page four by the time it reaches page forty.

That fear is now mostly nostalgia. Context windows have grown to the point where a full contract, a quarter of financial statements, or an entire book fits in one conversation with room to spare. If you are still slicing documents into polite little pieces before handing them over, the AI is not the one being cautious anymore. You are.

4. Micromanaging the AI like it has never done this before

There was a season where the winning move was breaking every task into small, careful steps. First do this, then do that, then check your work, then summarize. It felt responsible. It was also a lot of typing.

The newer models and the agents built on top of them can now take a goal, break it down themselves, delegate pieces to sub agents, and check their own output along the way. Spelling out every step for a system that can already plan is the digital equivalent of handing your project manager a script instead of a goal. Say what you want the outcome to be. Let it figure out the route.

5. Assuming AI images still look like AI images

If your mental picture of AI generated images involves six fingered hands and text that looks like a ransom note, that picture is out of date. Image generation inside the mainstream chat tools has closed most of that gap this year, handling text, hands, and realistic lighting in ways that used to require a specialist tool and a lot of patience.

This one is worth testing for yourself before you repeat the old joke about AI not being able to draw hands. The joke retired before you did.

6. Opening a new chat for every single task

Starting fresh each time used to feel like good hygiene, the AI equivalent of wiping down a counter before you cook. But a lot of these tools now carry memory and context across conversations, and long running threads have gotten dramatically better at staying coherent over time.

If you are still opening a blank chat for every small task, you are quietly rebuilding context from scratch every single time, and paying for that in wasted effort you do not even notice you are spending.

7. Copying and pasting text out of a PDF instead of just uploading it

You know the routine. Open the PDF, select the text, paste it into the chat, watch it arrive as one long paragraph with the columns scrambled and the table turned into soup. Then you spend five minutes cleaning it up before you have even asked your question.

You can skip all of that. Upload the file. The current tools read PDFs, spreadsheets, and slide decks directly, tables and formatting mostly intact, and they will ask you what you need instead of making you reconstruct the document by hand first. If you are still copying and pasting out of PDFs, you are doing unpaid data entry for a machine that was built to do it for you.

8. Staying loyal to one AI like it is a phone carrier

Picking one tool and sticking with it out of habit made more sense when the gap between providers was small and switching cost you real effort. Neither of those things is as true anymore. Different models are still genuinely better at different jobs, one for long documents, one for coding, one for a certain kind of writing voice. Knowing which tool to reach for is quietly becoming a more valuable skill than being fluent in any single one of them.

9. Treating prompt engineering as the highest form of AI skill

Prompt engineering had its moment. Right now, the more valuable skill is context engineering, deciding what information, files, and history the AI actually needs to do the job well, and workflow design, deciding how a task gets broken up and handed off between steps or agents. The clever one liner prompt is a parlor trick next to a well built workflow that runs the same way every time.

10. Forgetting that your AI provider is not just a tech company anymore

This is the one nobody saw coming a year ago. AI models are now genuinely subject to export controls, and it is not hypothetical. Anthropic's newest tier of models was pulled from public access for a stretch this year to comply with a Commerce Department order, then restored once the restriction lifted. For almost three weeks, a frontier model that businesses were quietly building workflows around simply was not available, for reasons that had nothing to do with the technology itself.

That is a new category of risk to plan around. Which model you can access is no longer only a product decision. It is occasionally a geopolitical one, and it is worth knowing that before you build something critical on a single provider.

The one habit that has not aged at all

Everything above changed. This has not: check the important stuff. AI is more capable than it has ever been, and it will still say something confidently wrong in the middle of something impressively right. The mistake to retire is not verification itself, it is verifying everything with the same paranoia you needed two years ago, instead of learning where these tools are actually reliable and saving your scrutiny for where it counts.

The tools changed a lot faster than most people's habits did. If even two of these felt a little too familiar, you are not behind. You are just due for an update, same as the software.

My guess is that next year's embarrassing habit will not be bad prompting at all. It will be doing everything yourself by hand when you could have handed it to an agent and gone to get coffee. We will be writing this same article again in twelve months, just about a different kind of stubbornness.

Then AI showed up and started answering everything before anyone could even finish typing the question.

Ask it to explain quantum mechanics, draft a lease agreement, write a Python script, guess at a diagnosis, summarize a 40 page research paper, or spit out a marketing campaign, and it does all of it in about the time it takes you to make a cup of tea. No degree required. No decades required. Just a keyboard and mild curiosity.

Naturally, everyone panicked a little and declared expertise dead. Rest in peace, the PhD. Rest in peace, the twenty year career. Nice knowing you.

I think that's exactly the wrong takeaway.

AI didn't democratize expertise. It commoditized information. Those are not the same thing, and confusing them is the whole mistake.

Expertise hasn't died. Only the cheap, easily Googleable part of it has, the part that was already on its way out anyway. For the first time in a while, the person who knows the most facts in the room may no longer be the most valuable person in it.

Knowledge Used to Be the Moat

For most of human history, information was valuable simply because it was hard to get. If you had specialized knowledge, you had leverage. Doctors knew things patients didn't. Lawyers knew things clients didn't. Your mechanic knew, with total confidence, that the "weird noise" was going to cost you exactly one rent payment.

The internet cracked that moat a little. You could WebMD your way into a mild panic about a headache. AI just drained the moat entirely and filled it with a water park.

Now anyone can ask an AI a genuinely sophisticated question and get a genuinely competent answer back in seconds. Information isn't scarce anymore. It's basically tap water. And like anything that goes from rare to abundant, its price has crashed. Knowing facts used to be the flex. Now facts are free, and everyone's holding the same cup.

The Real Scarcity Was Never the Information

Here's a fun experiment. Put the same AI in front of three different people.

A second year medical student. A surgeon with thirty years of experience. And someone whose only medical training is a very committed relationship with Google.

All three type the exact same question and get the exact same answer back.

Only one of them knows, immediately and instinctively, that the answer is missing a key detail, or that the "textbook" diagnosis doesn't quite match what's actually in front of them, or that the AI just said something wrong with the unearned confidence of a guy at a dinner party who read one article.

That's not an information gap. All three have identical information sitting on their screens. It's a judgment gap. And judgment, annoyingly, does not download in five seconds.

The New Premium Skill

Judgment is one of those things that's hard to define but instantly recognizable, like a good joke or a bad haircut.

It's knowing which variable in a mess of data actually matters, and which ones are just noise dressed up as signal. It's spotting the second order consequence three moves before it happens, the way a chess player winces at a move that looks fine right now but is going to be a disaster in six turns. It's understanding context well enough to know that the "correct" answer in a textbook is sometimes the wrong answer in the room you're actually standing in. And it's being able to change your mind when new evidence shows up, instead of just defending your original guess like it's your firstborn.

None of this comes from memorizing more facts. You don't get judgment by cramming. You get it by living through enough weird, ambiguous, badly labeled situations that your instincts start to catch things your conscious brain hasn't even processed yet.

Which is exactly the kind of skill AI cannot hand you in a chat window, no matter how nicely you ask.

That's because judgment is built from first principles, not from answers. People with strong judgment don't just remember what happened last time. They understand why it happened, which is a completely different piece of software running in their head. That's the whole reason they can walk into a situation they've never technically seen before and still land on their feet, while someone running on memorized patterns freezes the moment reality refuses to match the pattern. Facts tell you what. First principles tell you why. Only one of those survives contact with a genuinely new problem.

AI Doesn't Replace Thinking, It Reveals Who Was Actually Thinking

A lot of people use AI the way they used to use a copy paste button. Ask it a question, take the answer, move on with their day, brain fully disengaged.

The sharper professionals use it more like a sparring partner. They don't ask it to think for them. They ask it to argue with them.

One person says "what should I do here?" and does whatever the reply says. The other says "here's my plan, poke holes in it," and then actually reads the holes.

Same tool, wildly different outcome. The first person gets a little more dependent every time. The second gets a little sharper every time, the way a good editor makes a writer better simply by asking annoying, well aimed questions.

Expertise Is Evolving, Not Expiring

The experts who matter a decade from now won't be the ones who memorized the most facts. AI already won that contest and it's not close.

They'll be the ones who ask the question nobody else thought to ask. Who can spot a shaky assumption sitting quietly underneath an otherwise elegant argument. Who can borrow an idea from biology and use it to fix a problem in logistics, because they've read widely enough to notice the two rhyme. And who know, from experience, exactly when a suspiciously clean answer deserves a second look instead of a nod.

AI can hand you a thousand plausible options in the time it takes to sneeze. It cannot tell you, with any real authority, which one deserves your money, your reputation, or your patient's life. That part is still, stubbornly, a human job.

The Future Belongs to the Thinkers, Not the Knowers

School spent a very long time rewarding the kid who had the correct answer memorized. The AI era rewards the kid who asks the better question instead, the one that makes everyone in the room go quiet for a second.

Companies used to hire for what you knew. Increasingly, they're going to hire for how you think, because what you know is now one search bar away from everybody else's what you know too.

In a world where intelligence is on tap for anyone with wifi, having information stops being the advantage. Knowing what to actually do with it is the advantage, and that part was never for sale.

So no, the reports of expertise's death were greatly exaggerated. It's not dying. It's finally becoming the thing it was always supposed to be, underneath all that trivia.

Every major technological shift ends up rewarding something new. The printing press rewarded literacy. The industrial revolution rewarded specialization. The internet rewarded access, whoever could find the information fastest won.

AI seems to be rewarding something older, and considerably rarer.

Judgment.

Maybe expertise was never really about knowing the answers. Maybe it was always about knowing what to do the moment the answers arrived.

The future won't belong to whoever knows the most.

It'll belong to whoever thinks the best.

That model built trillion dollar companies. It also built an entire generation of professionals who are extremely good at clicking "export as CSV" and reasonably good at pretending that counts as strategic work.

But AI is quietly rewriting the terms of the deal, and most people are still staring at the old contract wondering why the numbers don't add up. Here is a link to an article written by Nicolo' Caiti on the current shift. Article here. He has ran the numbers on the changing aspects of the industry to explain everything more accurately.

The next generation of software isn't going to ask you to click buttons, configure workflows, or spend a weekend watching an onboarding video narrated by someone who sounds suspiciously cheerful about permissions settings. It's just going to ask you one question. What outcome do you want? And then, without much fanfare, it's going to go get it.

That's the shift from Software as a Service to Service as Software. Once you see it, you start noticing it everywhere, in every product that quietly stopped asking you to do the work yourself.

We didn't build elevators so people could get better at climbing stairs. We built them so climbing stairs stopped being part of the job. AI is doing the same thing to knowledge work. It isn't making the ladder nicer. It's removing the ladder.

SaaS Was Built Around the Human Doing the Work

Traditional SaaS products are genuinely powerful pieces of engineering. But every single one of them, from the fanciest enterprise suite to the cheapest invoicing app, shares one quiet assumption baked into the architecture. The human is the operator. The software is just an extremely elaborate steering wheel.

Want to send an invoice? Open the software, fill the form, upload the files, review the details for the third time because you are paranoid about decimal points, and finally click send. Want a report for your boss? Choose your filters, select your dates, export the CSV, and then spend forty minutes in Excel making a chart that will be looked at for six seconds in a meeting before someone says "can we get this in a bar chart instead."

The software gives you capability. You give it effort. And no matter how sleek the interface gets, no matter how many product designers spend six months debating whether a button should be teal or cyan, the entire experience still revolves around you doing the doing. SaaS never actually did the work. It just made the work slightly less miserable to do yourself.

Then AI Walked In and Changed the Furniture

Large language models don't just spit out text like a very confident autocomplete. Increasingly, they can reason across documents, pull together context, write the email, generate the report, call the right APIs, and coordinate between tools with surprisingly little hand holding. Not perfectly, not always, but often enough that the shape of the work has started to change. For the first time in the history of business software, the software can actually attempt the task instead of merely offering you a very polished set of options to perform it yourself.

The interface stops being a dashboard. It becomes an instruction. You don't navigate fifteen menus anymore, you just say something like "reconcile last month's GST returns, flag any mismatch above ten thousand rupees, draft polite but firm emails to the vendors who haven't sent their invoices, and give me the final report by evening." No clicking through screens that were clearly designed by someone who has never actually filed a tax return. No exporting six spreadsheets and stitching them together in a fragile formula that breaks the moment someone adds a new column. The software simply becomes the worker, and you become the person who reviews its work with a cup of chai instead of the person who did it with a headache.

Nobody Actually Wants Software. They Want the Thing on the Other Side of It

This is the part everyone conveniently forgets while they're busy admiring their own product roadmap. Nobody wakes up in the morning desperately wanting a CRM. They want more customers. Nobody dreams of owning accounting software. They want books that are compliant and a CA who doesn't sigh audibly every time they call. Nobody wants project management software for its own sake. They want the project delivered on time, ideally without three people independently creating the same task called "final final v2."

For years, software companies sold tools and let you figure out the outcome yourself, the way a hardware store sells you a drill and wishes you luck with the shelf. AI finally lets companies sell the outcome directly. That is not a small tweak to the business model. That is an entirely different business wearing the old one's clothes.

The Best AI Products Will Feel Almost Rude in How Little They Ask of You

Think about electricity for a second. You never stand in your kitchen appreciating the grid. You just flip a switch and the light turns on, and if it doesn't, you get mildly annoyed at the universe rather than curious about transformer efficiency. The best AI products are going to work exactly like that. The less you have to interact with them, the more valuable they actually are, which is a strange inversion for an industry that spent two decades optimizing for how many minutes you spent staring at a dashboard.

Success stops being measured in daily active users glued to a screen. It starts being measured in daily tasks that got completed without you lifting a finger, which honestly sounds like the healthiest relationship most of us have ever had with software.

Every Dashboard Is Quietly Nominating Itself for Extinction

Walk into any enterprise software product today and you'll find dozens of menus, hundreds of settings, and thousands of buttons that exist purely because humans needed granular control over every tiny decision. That entire architecture was built for a world where the human had to steer. AI does not need that steering wheel in the same way. Instead of you navigating the software, the software starts navigating itself, occasionally tapping you on the shoulder to ask permission or flag something odd.

Dashboards are not going to vanish overnight, mostly because change management in large companies moves at roughly the speed of a government file. But a lot of them will quietly demote themselves from primary interface to backup option, the digital equivalent of a landline you keep around just in case. The conversation becomes the workflow. The workflow becomes the outcome. The forty tab browser session becomes, mercifully, optional.

This doesn't mean dashboards disappear entirely. Pilots still have cockpits even though autopilot exists. Humans need visibility, control, and an escape hatch for the moment something goes wrong. But the dashboard becomes the exception, the place you visit when you need to intervene, not the place where the work actually lives.

The Winners Will Sell Results, Not Seats

Here is

Pricing in software has always been oddly obsessed with headcount. Per user, per seat, per month, as if the value of the product is somehow tied to how many humans are forced to log in and stare at it. But what happens when there are barely any users left to count? What happens when one AI agent does the work that used to take twenty people and several strongly worded emails?

Pricing starts looking a lot more like a business outcome than a software subscription. You end up paying for invoices processed, claims settled, contracts reviewed, leads converted, taxes reconciled without a single missed deadline. Nobody actually cares how many prompts were generated behind the scenes or how elegant the underlying model architecture is. They care whether the work got done, correctly, on time, and ideally without them having to think about it twice.

Here's why that matters more than it sounds. Salesforce charges you for access, whether or not you actually close a single deal that month. An AI employee can charge you for the deal it actually helped close. Those are not two flavors of the same pricing model. That is a completely different economic relationship between the vendor and the customer, one where the vendor only gets paid when the work actually happens.

This Doesn't Mean Humans Are Being Quietly Shown the Door

AI is genuinely excellent at structured, repetitive, information heavy work, the kind of thing that used to eat up an entire afternoon and most of your patience. But humans remain absolutely essential in places where judgment, accountability, negotiation, ethics, and actual relationships matter, which, if you have ever tried to get an AI to handle a delicate client conversation, you will appreciate more than most.

The real shift isn't people being replaced by AI. It's people moving from operating software to supervising a service, from being the one turning the crank to being the one making sure the machine turned it correctly. That is a genuinely different job, and honestly, a more interesting one.

The Next Great Software Companies May Not Look Like Software Companies at All

The biggest winners of this era probably won't be the ones with the most beautiful dashboard. They will be the ones with the most reliable autonomous worker sitting quietly behind the scenes. Companies will compete on accuracy, trust, memory, domain expertise, compliance, and their sheer consistency in delivering the outcome they promised. The interface becomes almost an afterthought compared to the quality of execution happening underneath it.

The internet digitized information. Cloud computing digitized the distribution of software. AI is now digitizing work itself, which is a considerably bigger deal than either of the two that came before it, even though it arrived with far less fanfare and considerably fewer conference keynotes.

That is why I think we are moving into an era where software stops being something we operate and starts becoming something that simply works for us, quietly, competently, and without needing us to click fifteen buttons to prove we meant it.

Every software company now has a choice, whether they realize it or not. Keep selling tools and hope people stay loyal out of habit. Or start selling the completed work itself. History is not particularly kind to the first option once someone else offers the second.

The future isn't Software as a Service.

It's Service as Software.

The companies that understand this won't just build better software. They'll build software people forget they're even using, right up until they notice how much of their work quietly stopped needing them.

Adarsh Singh Pawar

I think the distinction everyone's stuck on is the wrong one. Nobody asks whether a founder personally wrote the code for their product, or whether a director personally operated the camera. The question was never who touched the keyboard. It's whose judgment shaped the thing. My judgment picks the argument, the structure, the parts worth keeping. The AI is the pen, not the author.

That said, I get why this bothers people. Writing has always been sold as a kind of intimacy, the idea that you're getting a direct line into someone's head, unfiltered. A tool in the middle breaks that illusion a little. I'm not going to pretend it doesn't. What I will say is that the ideas here are still mine, the conclusions are still mine, and if you disagree with something I wrote, you're disagreeing with me, not a language model.

There's a second thing this changes that people don't talk about enough. Ideas have a shelf life. You get a thought worth writing about, and if turning it into two thousand coherent words takes you three evenings, most of those thoughts just die on the way there. That's not a discipline problem, that's math. A regular writer sits on ten good ideas a month and ships two, because the other eight would take too long to become anything, and by the time there's room, the moment's passed and the idea's gone stale. Writing a full-length piece becomes this heavy, mundane climb every single time, so you ration which ideas even get the climb.

What the tool actually buys me is throughput on that bottleneck. The idea still has to be good, the argument still has to be mine, but the distance between having a thought and publishing it shrinks from days to hours. So more of what I actually think makes it out into the world while it's still worth reading, instead of getting quietly shelved because turning it into a full article felt like too much for one more opinion.

So here's the honest version: I use AI to write faster, not to think for me. If that's a dealbreaker for you as a reader, fair enough, no hard feelings. If it's not, then nothing about what you're reading changes starting today. It was already like this. It just means fewer of my ideas quietly die before they reach you.

The Dutch East India Company, the VOC, ran the same play a century earlier with nutmeg. They didn't just sell spices, they cornered the entire supply so completely that owning a nutmeg tree without VOC permission was, at one point, a capital offense on the island of Banda. That's not a metaphor, that's an actual historical fact people should sit with for a second. An entire economy was restructured around one company's monopoly on one plant, enforced with the kind of violence that doesn't make it into the shipping company's annual report.

I keep thinking about that pattern because I think we're watching a version of it happen again, except this time nobody has to build a single ship.

The old empires needed land, ports, and armies to extract value from a place. You controlled the trade route, you controlled the wealth. You controlled the ports, you controlled the empire. It was expensive, visible, and eventually it got you kicked out, sometimes literally at gunpoint, sometimes just through the slow humiliation of decolonization paperwork.

The new version doesn't need any of that. It just needs you to open the app.

The Company That Owned a Country, Minus the Company

Quick history lesson for anyone who slept through this bit: the East India Company started as a trading operation and ended up governing a subcontinent of roughly 200 million people, with its own private army larger than most European nations'. It didn't conquer India so much as India became quietly dependent on it, one trade concession at a time, until sovereignty was something India used to have.

The analogy isn't perfect, but the strategic logic is surprisingly familiar. Replace trade concessions with API access, and the pattern starts to look recognizable.

A hospital that runs its diagnostics through a foreign AI model hasn't been invaded. Nobody planted a flag in the radiology department. But if that model goes down, changes its pricing, or quietly deprioritizes a disease that's common in Lucknow but rare in California, the hospital finds out the same way Bengal found out who actually ran its granaries. Not through a declaration, through a shortage.

That's the part that should unsettle people more than it does. Colonialism was never really about soldiers. Soldiers were just the enforcement mechanism for something more boring and more permanent: control over the systems a place needed to function. The railways, the currency, the courts. Once those ran through London, the flag was almost a formality.

Infrastructure Built for Them, Not for You

The British built more railway track in India than almost anywhere else on earth, and Indian schoolchildren spent decades being told this was basically a gift, evidence of imperial generosity. It wasn't. The lines ran from the cotton fields to the ports, not between the cities that actually needed connecting. It was logistics for extraction, dressed up as modernization.

Look at where the world's compute is actually being built right now, and you start seeing the same shape. Massive data centers are going up wherever electricity is cheap and regulation is thin, optimized for training runs that will overwhelmingly benefit whichever company or country owns the model at the end of it. The country hosting the data center gets a few jobs and a tax break. The value generated by the thing running inside it accrues somewhere else entirely, usually a headquarters on a different continent. It has the shape of investment. It functions like a railway line to the port. Owning the land that hosts the servers is not the same as owning the intelligence running on them.

Cash Crops, Then and Now

Colonial economies had a signature move. Stop growing food, start growing whatever the empire wants to sell. Indigo in Bengal. Cotton in Egypt. Rubber in the Congo, extracted with a brutality that still shows up in the population statistics a century later. The local economy got reorganized around someone else's supply chain, and the local population lost the muscle memory of feeding itself.

AI has a quieter version of the same move, and you can watch it happening in your own inbox. Companies are restructuring entire workflows, legal drafting, financial modeling, customer support, around platforms they don't own and can't modify. It feels like efficiency. It is also, structurally, the same bet the indigo farmer made: that the buyer will always want what you're producing, on terms you don't set.

Ask a small accounting firm what happens to their workflow if their AI vendor doubles the price next quarter. Ask them what happens if the vendor discontinues the feature they built their entire practice around. That's not a hypothetical anymore, that's just Tuesday in enterprise software.

There's an uglier version of this parallel too, and it deserves saying plainly instead of dancing around it. King Leopold's Congo Free State ran on rubber quotas enforced with mutilation, one of the more horrifying chapters in a genre with a lot of competition. Nobody is doing that today, to be clear, but the underlying economic shape, where the dirtiest and lowest paid work sits at the bottom of a value chain whose profits land somewhere else entirely, has a live descendant. Several prominent AI systems relied on contract workers in countries like Kenya to perform difficult content labeling and moderation work, paid a few dollars an hour to sift through the worst material the internet produces so the rest of us get a clean chatbot. The finished product ships from San Francisco. The labor that made it usable shipped from somewhere the annual report doesn't dwell on.

Free Samples and the Opium War Playbook

Britain had a trade problem with China in the 1700s. The Chinese wanted very little that Britain made, and Britain wanted a great deal of Chinese tea, which created an awkward silver deficit. The solution, and I'd love it if this were an exaggeration, was flooding the Chinese market with opium grown in British India until enough of the population was dependent that the trade balance fixed itself. When China tried to ban it, Britain fought two wars to keep the pipeline open and took Hong Kong as a parting gift.

I am not saying free tiers and freemium pricing are opium. I am saying the underlying strategy, give it away or price it artificially low until an entire population or economy has restructured itself around your product, then adjust the terms once leaving is no longer realistic, is a genuinely old trick and it is currently being run at a scale Britain could only have dreamed about. The first hit is free. The API pricing after the workflow depends on it is a separate conversation, held from a position of considerably less leverage than the one you had before you integrated.

The Funny Part, If You Squint

Here's a smaller, dumber example, but I think it's the most honest one. Ask most mainstream AI models to write you a wedding toast, and you'll get bullet points about "cheers" and "here's to the happy couple," which is a very confident answer for a culture that has never once done a wedding in under three days or fewer than four hundred guests. Ask it to help plan a festival budget and it will cheerfully suggest a turkey centerpiece regardless of which festival, which country, or which century you're apparently celebrating in.

This isn't malice. It's just what happens when the intelligence infrastructure of the world gets trained overwhelmingly on one cultural default and then exported everywhere as if it were neutral. The British didn't think they were imposing a worldview either. They thought they were installing "civilization," the same way a chatbot doesn't think it's centering California, it thinks it's just being helpful. Empires rarely experience themselves as empires. They experience themselves as the reasonable default that everyone else should be grateful to adopt.

Multiply that small comic misfire across medicine, law, and education, areas with real consequences instead of an odd toast, and the joke stops being funny quite so fast.

Sovereignty Was Never About Doing It Alone

I want to be careful here, because this argument has a bad habit of curdling into "build a wall around your data and hope for the best," and that's not what actually worked historically either. The countries that recovered fastest post-independence weren't the ones that cut themselves off. They were the ones that kept trading, kept collaborating, and kept a domestic base in the things that mattered strategically. Japan didn't refuse Western technology after the Meiji Restoration. It imported it aggressively and then built its own industrial base on top of it, on its own terms.

That's the model worth copying. Use the foreign platforms. Trade, partner, integrate, all of it. But build the compute, the talent pipelines, and the domestic models in parallel, so that "use" never quietly slides into "depend on with no alternative." The difference between a trading partner and a colony has never been whether you interact with the empire. It's whether you could, in principle, walk away.

Two Kinds of Countries, Again

Every major technological shift produces the same split. The ones who write the rules, and the ones who spend the next fifty years adjusting to rules they didn't get a vote on. The steam engine did it. The internet did it, quietly enough that most people didn't notice their entire information diet running through servers in one country until it was a little too late to renegotiate.

AI is running that same script, just faster and with better UX.

Bengal didn't lose its sovereignty in a single dramatic afternoon. It lost it in a thousand small trade agreements that each seemed sensible at the time. That's usually how it goes. Nobody signs the form that says "we are now a digital colony." They sign the form that says "enterprise licensing agreement," and the rest happens on its own.

Colonialism rarely announced itself as colonialism. It arrived as trade, convenience, and modernization, and by the time anyone tallied what had been traded away, the dependency was already structural. AI will not decide who wins this century on its own. Control over it will.

Empires no longer need to occupy your territory. They only need to become impossible to leave.