Five Stages of AI Adoption

A curious thing happens every time a breakthrough technology shows up. First we surround it with jokes, as if laughter were a perimeter fence. Then we argue with it, as if the quarrel might slow it down. Finally, we begin to build on top of it. That last step is where the real story starts, but it's also the step that gets the least airtime because the earlier emotions are so loud.

I've been watching teams make that journey with AI agents for three years now. During that time the models grew by orders of magnitude, but the emotional choreography stayed weirdly constant: denial, anger, bargaining, depression, synthesis.

1. The Shock of the New

When GPT-2 appeared it could barely keep a topic alive for a paragraph, yet the shock was real because text itself had always felt like human turf. A model that could juggle syntax was novel; a model that could juggle meaning felt borderline profane. Most people, myself included, misread that early stumble as an upper limit rather than a baby step.

By the time GPT-3 arrived the conversation had become ritualized. A demo would circulate, someone would reply "toy," and the Slack channel would move on. What looked like skepticism was mostly insulation: if a tool is trivial, you are safe from the work of changing your workflow. Denial is efficient that way.

2. Five Stages (and Why They Matter)

The Kübler-Ross curve was coined to talk about mortality, yet the same shape shows up whenever we're forced to revise a self-image quickly. Losing a paradigm is not as traumatic as losing a friend, but the nervous system doesn't keep separate wiring for existential and professional threats; it just fires. So the stages arrive.

Denial cracks first, then anger rushes in to fill the vacuum. With AI the anger sounds oddly practical: "The code is ugly", "It needs constant babysitting", "Look, six fingers". These objections are accurate; they're also temporary. Each edit you make trains you, not the model, to ask better questions next time.

Bargaining sneaks in next. "Let it handle boilerplate", we say, drawing a protective fence around the "real" work. Fences slide. The agent handles docstrings, then unit tests, then migration scripts. By the time the fence reaches system design the owner of the fence is too tired to move it again.

Depression is never admitted out loud, but you see it in reduced curiosity. A bright colleague who once tore into every new tool now limits himself to the one prompt he knows works and prays the conversation ends. It looks like laziness; it's actually grief.

Acceptance — if that word means "sure, I'll use it when I must" — is too low a bar. Synthesis is the real destination. That's the moment when an agent stops feeling like a rival and starts feeling like a pair of glasses: you notice the world, not the lenses.

Knowing the pattern buys you time. When you recognize your own anger as a stage, you can stand outside it long enough to ask what useful data the anger is hiding. When you hear a colleague bargaining, you can guess where the story is headed and prepare resources for the next bend in the road. Maps matter even when the roads are emotional.

3. Grief Hiding Inside Progress

A strange cognitive dissonance sits at the heart of technological optimism: you can cheer for progress in the abstract and still mourn its local consequences.

The minute an agent solves a task you once treasured, the victory feels personal only if you're the beneficiary. If you are the task's former custodian, the same victory looks like eviction.

The grief is real. For years you were rewarded for manual dexterity in a specific craft — debugging regex, pushing pixels, writing boilerplate tests — and that dexterity became part of your identity.

The agent shows up and performs the same act in seconds. Of course the rational voice in your head says, "Great, now I'm free to tackle harder problems". But another quieter voice whispers, "If that part of me was replaceable, what else might be?". Pretending the whisper doesn't exist just amplifies it.

Once you recognize the whisper as grief, you can treat it with the normal courtesies humans extend to grief: time, conversation, the permission to feel awkward for a while. Teams that skip this step often circle at bargaining forever, because no one feels safe enough to admit how tired the dance has become.

4. What Work Looks Like After the Dust Settles

In the old workflow you typed, compiled, stepped through, and typed again. In the new workflow you ask a well-aimed question, wait eighteen seconds, evaluate a wall of output, and decide which twenty percent deserves oxygen.

You've moved up one "meta" level: from generating artifacts to curating them.

Curating turns out to be cognitively harder than typing because the solution space expands faster than your attention budget.

Say you asked the agent to propose five database schemas. It politely emits twenty, because quantity is cheap and enthusiasm is infinite.

Now you need criteria. The job description quietly rewrites itself:

"Wanted: human who knows which constraints actually bind the real world."

Another surprise is how quickly tasks fracture. One person used to write the code, another reviewed it, a third wrote docs. An agent can do all three halfway well, so each of those humans must now decide whether to go deeper (specialize) or go higher (orchestrate). The result is a mesh, not a ladder.

5. Meta-Skills Versus Ten Thousand Hours

Because agents are cheap to spin up, what limits progress isn't how fast you can type but how precisely you can think. Prompt literacy is the surface skill, but the deeper asset is problem framing. A vague prompt is a foghorn; it tells the agent "somewhere in that direction." A precise prompt is a compass; it tells the agent "here is north, find me east."

Critical review becomes the heat shield that lets you re-enter the atmosphere without burning up. Past a certain quorum of generated artifacts, the danger flips: you no longer fear scarcity of ideas; you fear abundance of wrong ones. Audit skill — knowing where reality's tripwires hide — is suddenly a growth industry.

Domain storytelling is the third pillar. Agents don't read between the lines unless you write the lines. The faster you can convert tribal knowledge into narrated constraints, the faster the agent's ceiling rises.

Finally comes prioritization. You don't need twenty usable answers; you need one decisive next step. Your leverage now depends on the questions you don't explore, not the ones you do.

6. Helping a Team Cross the Ravine

You can't legislate synthesis, but you can cultivate it. Start with small wins. Give an agent a ticket no one loves, then run a retro on the delta in cycle time. Turn that anecdote into a rumor, because rumors travel faster than memos.

Next, build muscle memory. Open office hours where people can bring messy prompts and walk out with cleaner ones. Cheer for edits, not first drafts; the joy of skipping step zero is contagious once demonstrated. Most importantly, signal that grief is discussable. A single Slack thread where someone admits "I feel obsolete" does more to unblock adoption than ten lunch-and-learns about vector databases.

Progress rarely stalls for lack of information; it stalls for lack of psychological safety. If you clear that path, the technical wins follow at surprising speed.

7. A Frontier, Not a Cliff

Whenever somebody asks "How many jobs will AI eliminate?" I picture an optical illusion: a black silhouette that could be a vase or two faces in profile. Both readings are true; neither is complete. Yes, certain job descriptions will compress. I've already seen tasks that once justified whole backlogs collapse into a single function call. At the same time, possibilities that once looked like science fiction are suddenly doable on a Tuesday.

History says we usually under-count the upside. When electricity automated the factory floor, the pessimists imagined cities full of idle workers and dark tenements. What emerged instead were skyscrapers, jazz, neon, and a thousand new professions nobody had language for in advance. Automation didn't end work; it rearranged it around higher-order leverage.

AI agents feel similar, except the rearrangement targets cognition instead of muscle. If muscle automation multiplied horsepower, cognition automation multiplies curiosity. The limit stops being "What can I build in a quarter?" and becomes "What frontier is worth claiming now that I can iterate ten times faster?"

Where We Land

We won't remember these months for their model numbers. We'll remember them for the day drudgery finally became negotiable, and for how quickly that realization rewired our ambitions. Everything routine is on a timer now; everything imaginative just got a raise. Our best defense against displacement is the same impulse that once drove us to master the old tools: curiosity amplified by craft.

So ask the hard question sooner. Not "Can I trust the agent?" — that debate will age about as well as dial-up modems. Ask instead, "If the dull parts vanished overnight, what question would I chase that I've never had time to ask?".