The two camps killing AI adoption

A couple of months back I had two conversations, a few weeks apart, with two business owners who couldn’t have been more different — and somehow ended up at the same kind of wrong.

The first runs an established software business — old codebase, harder domain than the website lets on. By the time we talked, he’d already spent well into six figures on a refactor onto a modern stack. Senior team. Clean scope. Without AI — to him, AI obviously wasn’t up to it.

The refactor never shipped. They paused it indefinitely. He was open about the money — “we tried, it didn’t work” — and then, without skipping a beat: “we’ll wait this AI thing out the same way.” He’d watched blockchain. NFTs. The metaverse. To him, this was another wave the boring industries get to sit out.

The second runs an established company that just started chasing a new idea. His position is the inverse of the first: he no longer trusts anything that isn’t AI-generated. Copy: AI writes it. Images: AI makes them, no brand book underneath, no longer view of what the visual identity should even be, just whatever the prompt returns that week. Engineering: leaner team, broader AI surface — “the spreadsheet works.”

What he’s not looking at, yet, is what AI does when it’s the only senior judgement in the room. Things that worked six months ago don’t quite work anymore. Nobody on his side can point to when they stopped — no single commit looked wrong on its own. The brand voice drifts a degree further from itself each campaign. None of it shows up in one place. All of it shows up — eventually — at the same time.

Both of them walked out of those conversations feeling like they’d made the smart call.

Both of them are losing — for opposite reasons. And the thing they have in common is more interesting than the thing they disagree about.

The thesis, stated plainly

Most owners I talk to in 2026 are doing one of two things with AI. Both are wrong.

• Camp A says it’s hype. They’re waiting it out.
• Camp B says it replaces humans. They’re firing into it.

The actual winning move sits in the middle. Almost nobody’s playing it — because the middle doesn’t make a good slide, it doesn’t generate a follow-on consulting engagement, and it doesn’t get retweeted.

That’s the essay. Two thousand more words to defend why.

Camp A — “AI is hype, wait it out”

Who’s in this camp. Owners who’ve been through a hype cycle or two. Often early fifties. Established in their industry. The kind of pattern recognition that comes from watching three “next big thing”s land and not land. They were promised “AI-powered” something in 2019 and got a regex. They watched blockchain become a punchline. They feel vindicated, and on most days they have a point.

Why they sound right. They have proof. Blockchain was hype. The metaverse was hype. Most consumer-grade LLM output, examined for thirty seconds, is visibly wrong. The TechCrunch coverage around AI is mostly empty. When an owner in Camp A says “we tried it and it hallucinates,” they are reporting accurately. I’ve watched it happen in production. It’s not subtle when it does.

Where they’re actually wrong. Every prior hype cycle was waiting for something — infrastructure, regulation, behaviour change. Web3 was waiting for users. The metaverse was waiting for hardware. Self-driving cars are still waiting for liability law.

This one isn’t waiting. It’s already shipping work, today, for money, in shops Camp A competes with. The proof isn’t in the press release; it’s in the competitor’s speed of delivering real work, their support response time, their feature velocity. The “wait it out” position assumes the technology is stable enough to wait on. It isn’t — it’s accelerating. The gap between a shop using AI well in 2026 and a shop just starting in 2028 won’t be measured in features. It’ll be measured in operating margin. And, eventually, in customers.

What it costs them. By the time the wait-it-out crowd starts, they’re paying senior engineers to do work juniors should have handled — because they spent the cycle adding headcount instead of judgement. The companies they thought they were ahead of, by virtue of “not falling for the hype,” are now ahead of them on the only metric that matters: what the customer sees.

Camp B — “AI replaces everyone”

Who’s in this camp. Founders who’ve never managed a team of thirty. Investors with slides titled “agent-first”. Owners who read one famous venture capital article this quarter and decided to be on the right side of it. Sometimes — and this is the painful one — capable operators who’ve talked themselves into Camp B because the alternative is admitting they don’t know what to do.

Why they sound right. AI does remove a class of work that was previously expensive. That’s true. Some demos are real — agents that actually compile, agents that actually defuse a support ticket, agents that actually summarise a meeting accurately enough to act on. The unit economics, when you only look at the bill, look great. The pitch is internally consistent and the spreadsheet adds up.

Where they’re actually wrong. Real organisations are not the sum of their tickets.

Support is half the engineering budget. Every product decision you ever make is eventually re-litigated by a customer who couldn’t figure out what you built. Regulatory exposure compounds quietly until it shows up in a fine. Provisioning fails at 3am and the on-call has to know which upstream provider to call and what to say to get it unstuck. Bad architecture turns into a months-long rewrite that nobody in the spreadsheet predicted.

None of that work is in the spreadsheet, none of it is in the agent’s training data, and none of it is going to be learned by reading your Slack. AI cleanly removes recurrent work. It doesn’t replace judgement, taste, empathy, or the political work of getting an org to agree on what’s worth building.

Strip those layers away and replace them with an agent, and you don’t slow down — you accelerate the wrong thing. You ship the wrong product very fast, at senior velocity, from day one, exactly as advertised. That’s the trap: you get all the speed and none of the wisdom.

What it costs them. Camp B ships hard. Then it ships wrong. Then the only people left in the building when it cracks are the agents that can’t tell anyone why.

Both camps are making the same mistake

Camp A and Camp B disagree about what AI is. But they agree about how to think about it.

Both treat it as a unit decision: in or out, replace or refuse. Both are reacting to a story they read, not to a business they’re running. Both are answering the question that’s in the air rather than the question their own operation is actually asking.

The owners not in either camp — the small group quietly winning the next five years — aren’t doing either. They’re treating AI the way they treat any other operational input: as a tool whose value depends on what you ask of it, who’s holding it, and what part of the work it’s pointed at. Same way they treat a database. Same way they treat a hire.

The question isn’t “do we use AI?” It’s “which parts of our work are recurrent, which parts are judgement, and where exactly does the line sit in our business?”

That’s a boring question. That’s why it works. Boring questions get answered correctly. Loud questions get answered fast.

The middle, as a position — not a hedge

The middle is not centrism. Centrism is “a bit of both.” This is not that.

The middle is a specific position with three claims, in this order:

1. AI removes the boring. The recurrent, the templatable, the work that should have been automated five years ago. Take this seriously — most of the cost savings come from here.
2. Humans hold the judgement. What to build, who it’s for, why now, what to defer, what to refuse. The political work of getting an org to agree on any of that. Take this more seriously — this is where the value sits.
3. You combine them on purpose. AI amplifies experienced operators. Neither side replaces the other; the interaction is the product. Get the interaction wrong and the other two claims don’t matter.

This is not a moderate position. It’s a specific one. It says exactly what AI does and exactly what it doesn’t. It refuses to apologise for the second list, and it refuses to inflate the first.

The position shows up in the work, not in the deck. Chillsim — a multi-provider eSIM platform with Hub orchestrator, reseller platform, admin UI, plus a retail backend migrated onto the new orchestrator — built end-to-end in 333 billable hours. One person. AI as the sidekick. It runs in production today with tens of thousands of users on top. The platform belongs to the client I built it for; I architected it, shipped it, and continue to maintain it. The EvoBytes accounting system that handles our own books is AI-native end-to-end — same delivery model, this one for us. The MCP server that lets this very essay carry context across sessions will be open-sourced shortly. None of that is theory. The proof is on the Receipts page — don’t take the position from the essay; take it from the work.

Why twenty years in operations changes the reading

I’m forty-six. I spent twenty years plus running operations — customer service, then customer-service management, then VP of operations at a telecom — before I came back to development. That sequence, operations then AI-native dev, is uncommon. It changes how I read this cycle.

Three things twenty years in ops gives you that the AI conversation otherwise doesn’t have.

1. Pattern recognition on hype cycles.

When VoIP arrived, every telecom operator had a “VoIP strategy” powerpoint slide sitting in front of them. Most of those slides were wrong about the timing, wrong about the threat, and right about exactly one thing: something was changing.

If you were building a retail VoIP business in 1998, great job — you did it early, survived the chaos, and earned the rewards. If you started your VoIP company in 2015, you had terrible timing: the market was already saturated, margins were crushed, and the easy wins were long gone.

The mistake was using the loudness as the signal — either to dive in or to dismiss. The signal was in the operating margins of the shops that quietly adopted early; that’s where you read whether a change is real. By the time the press declared the cycle, the margin had already moved.

What I learned watching that play out is that the loud “this changes everything” narrative is almost always wrong on the timing. And almost always right that something is changing, somewhere boring.

2. Pattern recognition on what doesn’t get automated.

Running an eSIM platform taught me that the boring infrastructure is the product. eSIM provisioning, regulatory reporting, security, fraud detection, GDPR compliance, order flows — none of these get featured anywhere. All of them kill the business if you get them wrong. When an upstream provider quietly dropped a whole product line at 2am, and our best-selling SKUs disappeared from inventory before anyone noticed — that’s not a problem an agent solves. That’s a problem the human operator solves, because the human operator was the one who’d insisted on the inventory alert a couple of months earlier, after a near-miss nobody else thought was important.

The people who survive the next round of AI adoption are the ones who already know which decisions can be delegated and which can’t, because they’ve made those decisions before, under pressure, with consequences. Camp B’s pitch is that the agent will learn this. It won’t. Not from your Slack, not from your codebase, not from your support tickets. It’s tacit knowledge, and tacit knowledge doesn’t show up in training data.

3. Pattern recognition on the reactive trap.

Most “AI strategy” work in 2026 is reactive. The question — “what’s our AI strategy?” — is being answered by people who haven’t done a real strategy review in two years. They’re answering the question that’s in the air rather than the question their own business is asking. That’s the reactive trap.

The operator who’s lived through three strategy cycles knows the difference between answering the question that’s in the air and answering the question your business is actually asking. The first feels like progress. The second creates compounding advantage. They almost never produce the same plan.

The combination matters. None of these three on its own is rare. The three together — recognising the cycle, recognising what survives it, recognising the reactive trap — is what separates strategy that compounds from strategy that just keeps you in the conversation.

What changes when you’re in the middle

Monday morning, this is what you do differently.

• You scope work by what it asks of the org, not by what tool ships it. The architectural question always precedes the tool question. “Should we use AI for this?” is the wrong frame. “What does this work actually require of us?” is the right one.
• You spend the AI-native speed on the thinking layer. Brief writing. Scope debates. The political work of getting the right decision into the building. AI shortens execution; it doesn’t shorten consideration. Spend the saved hours on the thing AI can’t do.
• You hire people who hold judgement, not people who hold a ticket queue. The queue is going. The judgement isn’t. Hire for the surviving role.
• You stop running “AI strategy” as a separate workstream. AI is a tool. You don’t run a “spreadsheet strategy” review. You don’t run a “phones strategy” review. The work is the strategy; the tool serves the work.

The bet

The owners winning the next five years won’t look revolutionary. They’ll look boring. Their teams will be smaller and more senior. Their decisions will be slower and more considered. Their build velocity will be higher and their roadmap shorter. They’ll be — not coincidentally — the ones who already knew how to run a business before AI showed up.

The boring middle wins because the work is in the middle. Always was. The edge is human judgement amplified by AI, not replaced by it.