Moving countries and adopting AI agents follow the same adjustment curve

Six months ago I was trying to explain to a friend why it feels physically uncomfortable to let an AI agent rewrite a chunk of code without reviewing every line. He asked – was it easy for you to move to a new country knowing three words of the local language?

I sat with that for a second. Turns out it's the same process. It even has a name.

A model that's 70 years old

In 1954, anthropologist Kalervo Oberg described four stages of culture shock, and the framework has barely changed since. The shape of it:

1. Honeymoon. Everything is new and thrilling. You feel like you can handle anything.
2. Irritation. The novelty wears off, attention shifts to the differences – and they start to grate.
3. Adjustment. The differences stop being a problem. You develop a working model for handling them.
4. Acceptance. You're not just coping anymore – some things about the new environment start to beat the old one.

I've lived through this four times in four years, literally, not as a metaphor. And when I started working heavily with AI coding agents instead of writing everything by hand about six months ago, I noticed the exact same loop – just compressed into weeks instead of months.

Week one: euphoria. The agent writes its own tests, fixes its own bugs, feels like magic. Week two: irritation. The agent breaks a config because it missed context that would've been obvious to any human. Week three: a working model – I know which tasks I can hand off and which ones I can't. After that: calm use, no illusions, no paranoia either.

The numbers behind why the irritation stage is normal

This isn't just a personal theory. Several independent 2026 reports converge on roughly the same figure: about 88% of AI agents that perform well in demos break down once deployed into real workflows (CIO Research, with a matching breakdown from ClarityArc). McKinsey frames it from the other side – only 20% of pilots ever scale to production.

Among developers the picture mirrors it. 84% use AI coding tools, but only 29% trust the output without checking it (Second Talent's 2026 review). 61% agree with the statement "AI often produces code that looks correct but isn't." A familiar feeling if you've ever accepted an agent's diff without reading it first.

There's even a name for what I went through in week two – AI tool fatigue. BCG research found that people running four or more AI tools at once actually lose productivity – the cost of switching between them eats the gains each one provides. Someone put it almost word for word what I was thinking in my own second week, posting on X in March 2026: "Am I the only one getting AI fatigue? Every day a new tool, every day a new feature."

Irritation isn't a sign you're doing something wrong. It's a documented second stage. It passes for everyone who makes it to the third.

A connection I didn't expect to find

While pulling research for this piece, I ran into a number that unexpectedly tied the whole thing together. According to MBO Partners' 2025 data, 89% of digital nomads already use AI in their work. This isn't two separate communities going through a similar adjustment pattern – it's largely the same people.

There are roughly 40-43 million digital nomads globally right now (MBO Partners and Nomad List), and the number keeps climbing a few percent a year. 66-69 countries offer dedicated remote-work visas. This isn't a fringe group anymore – it's a statistically significant slice of the working world, and nearly all of them are running both adjustments at once: to a new country, and to a new AI tool in their daily work.

What actually lowers the anxiety in stage two

Tested on myself across four years of moving and roughly a year of working closely with agents. The same things work in both cases:

Small, checkable steps instead of one big leap. Don't move somewhere forever right away – go for three months first, see how it feels. Don't hand an agent full production access on day one – give it a test branch first.

One reliable anchor who's already done it. In every country I lived in, I ended up with one local contact who could explain the things that were obvious to locals and completely opaque to me. Same with agents – one workflow you've run twenty times until it's automatic. It becomes your anchor point while you're still figuring out the rest.

Learning from failure, not from theory. I never read guides on "how to adapt to a new culture" – I remembered exactly where I tripped and avoided that spot next time. Same with agents: I didn't study prompt engineering from blog posts, I logged which task phrasing produced a bad result and adjusted from there.

Where the analogy breaks

There's a difference I don't want to paper over. A country doesn't rewrite its entry rules every week. AI models do.

Spain raised the minimum income threshold for its digital nomad visa in 2026 to €2,849 a month (up from €2,763), tying it to the national minimum wage – and stepped-up enforcement this year reportedly caught people who'd been living on the visa for years without any real remote income to show for it. Rules change, but once a year, predictably, through an official process.

The AI model you've finally learned to work with can turn into something else after a single update, with zero warning. That's less like moving abroad and more like living somewhere that rewrites its constitution every quarter – except nobody sends you a notice.

That's the one reason I couldn't fully carry my relocation habits over into working with AI. A country stabilizes over six months to a year. Here, there's no such thing as stability to begin with – and expecting Oberg's fourth stage (the one where some things about the new setup start beating the old one) to apply to any specific tool feels, at this point, unrealistic.

What I do differently now

After making this comparison, I stopped expecting AI tools to earn the kind of full trust that eventually shows up after a move. I keep them permanently parked at "month three in a new country" – comfortable enough to be productive, but every major operation still gets a second look. Not because I distrust the progress, but because the ground shifts faster than you can put down roots.

That sounds like a compromise without a happy ending. Maybe that's just the only workable mode for anything moving this fast – whether it's countries with shifting visa rules or models that ship a new version every quarter.

TL;DR

How many stages does culture shock have, and who identified them? Four – honeymoon, irritation, adjustment, acceptance. The framework comes from anthropologist Kalervo Oberg in 1954, and it's still the standard reference point in cross-cultural adjustment psychology.

Why do AI agents fail so often once companies deploy them? Several independent 2026 reports put the figure at around 88% of AI agents that succeed in demos failing once deployed into real workflows. The bottleneck usually isn't model capability – it's the absence of clear success criteria and real system integration during the jump from pilot to production.

What is AI tool fatigue? A documented effect from BCG research: people running four or more AI tools in parallel lose part of their productivity to the overhead of switching context between different interfaces and different tool behaviors.

How many digital nomads use AI in their work? 89%, according to MBO Partners' 2025 data. With the global digital nomad population sitting at 40-43 million, that means the adjustment to a new country and the adjustment to new AI tools are, for most of these people, happening at the same time.

Sources: Kalervo Oberg, "Culture Shock: Adjustment to New Cultural Environments" (1954); CIO Research 2025-2026, ClarityArc pilot-failure analysis; McKinsey State of AI 2025; Second Talent AI developer productivity report 2026; BCG multi-tool productivity research; MBO Partners 2025 State of Independence Report; Immigrant Invest Digital Nomad Visa Index 2026.