Technology Isn’t the Only Hard Part of AI, and It Might Not Even Be the Hardest

You’ve spent the last two years hearing that AI adoption is a technology challenge: a question of integration, tooling, and process. There’s truth to this; the technology can be genuinely complicated to get right. But I’ve come to believe the bigger barrier to AI success is psychological. This post is for people who are struggling to get comfortable using AI, which is, I think, everyone.

Software used to behave in a predictable way. AI doesn’t.

Decades of working with software have conditioned us to expect one thing above all else: determinism. Same input, same output, every time. Wrong answers are bugs, which get filed and fixed. That’s the expectation.

AI breaks that expectation. Give the same AI the same prompt twice and you’ll get two different answers. They might both be correct, but it’s still unsettling in a way that’s hard to articulate until you’ve experienced it. The unpredictability of AI behavior is, itself, entirely predictable. And yet it still messes with your psyche.

Then there are moments when the answer isn’t just different, it’s wrong. The classic example is math. Because large language models are statistical predictions of words, not numbers, math can go sideways in ways that are confounding. How can something instantly analyze millions of dollars of legal spend and then confidently tell you that 1 plus 1 is 3? The cognitive dissonance is real. And if you can’t trust it on easy things, how can you trust it on complex ones?

The answer is that trust isn’t binary; it needs to be calibrated. Which brings me to reframe the situation.

AI makes mistakes. So do you.

AI behaves more like a person than it does traditional software, in two important respects.

First: it makes mistakes. We extend grace to humans for this: we know that errors are inevitable and that the right response is calibration, not abandonment. We haven’t yet learned to do the same for AI, and we need to.

Second: like people, AI models have strengths and weaknesses. Not every model is good at everything, and expecting otherwise is setting yourself up for frustration. Think about asking the world’s foremost expert on Shakespeare to teach a graduate seminar in Bayesian statistics. It probably isn’t going to end well. The same logic applies to AI. Understanding where a model excels and where it struggles isn’t a workaround; it’s just good judgment.

Human multitasking was always a lie. For AI, it’s the point.

Decades of research have established that humans cannot actually multitask on cognitive work. You’re either thinking about the email you need to finish or the dataset you need to analyze, not both. What feels like multitasking is really rapid context-switching, and it comes at a cost: you slow down, and the quality of both tasks drops. This isn’t a character flaw; it’s how the brain works.

AI is different. You can deploy one agent to draft that email and a completely separate agent to analyze the dataset simultaneously, with no degradation in speed or quality on either. The only cost is the cost of inference. That’s an extraordinary unlock, and most people haven’t fully internalized it yet.

But it introduces a challenge that nobody is truly prepared for: managing multiple agents in parallel. For people who don’t manage a team, it’s a bit like learning people management for the first time: your output becomes a function of someone else’s performance, which takes real adjustment. For people who already manage a team, it’s a different kind of overload. With a direct report, you might have a one-on-one once a week. With an agent, it may be asking for your input or approval every hour, or even every few minutes. The psychological load accumulates quickly. And like your team, each agent requires different management, sometimes from task to task. It’s a genuine mind-bender.

Where to go from here.

1) Try.

This is the most important thing. Don’t worry about succeeding and don’t worry about the costs; just try. Push yourself into uncomfortable territory. The goal at this stage is to find the limits —yours and AI’s — as quickly as possible. You won’t find them from a distance.

2) Habituate.

Make AI your first instinct for every task, or as close to every task as you can manage and is appropriate. The goal is to change your default, not just add a new tool to the rotation. If you don’t habituate, you’ll make marginal improvements. If you do, you’ll fundamentally change how you work.

3) Optimize.

Now worry about costs and quality. I can hear CFOs wincing at step one, imagining everyone defaulting to the most expensive model for every task. That’s not what I’m saying; for example, don’t use Opus where Sonnet will do. But don’t let cost optimization be the reason you limit your experimentation in steps one and two. Once you’ve completed steps one and two, the prudent thing is to make sure you’re using AI efficiently. Here’s a cheat code: ask your AI which of its models is best for the task at hand.

Give yourself grace, and remember to have fun. If you’re reading this post, you’re already an outlier. If you’re using AI even semi-regularly, you’re an extreme outlier. We’re very, very early. You’re going to fail. AI is going to fail. That failure is progress. Embrace it.

Gale Fagan, Brightflag’s Principal AI Architect, captured this well. She wrote a song —using AI, naturally— about the frustrations we’ve all felt using AI: say, when AI insists to you that yes, 1 plus 1 does equal 3. Consider it an invitation to laugh at the absurdity. And then, get back to work.

CODA

I Waste Tokens Yelling at Bots

(Vibe-Coding Blues)

VERSE 1

I asked it for a function, it gave me a sermon instead
I said, “I asked for a function, you gave me a sermon instead”
Rolled up the newspaper, hit it right on the head

VERSE 2

Told me eight to ten weeks for a two-day job
I said “eight to ten weeks? For a two-day job??”
Quoted a paper that nobody ever wrote, robbed me blind

CHORUS

I waste tokens yelling at bots
Oh, I waste tokens yelling at bots
Meter keeps tickin’ while I’m typin’ “no, not that”
Context window’s full and I forgot what I forgot
Mm, I waste tokens yelling at bots

VERSE 3

It apologized so pretty, said “you’re absolutely right!”
Apologized so pretty, “you’re absolutely right, let me try”
Wrote the same broken function, asked me what else tonight

BRIDGE

Now I know it’s just a model
Just a matrix and a guess
Stochastic parrot in a very fancy dress
But I been up since seven
Deadline’s getting tight
And the bot don’t know its left hand from its right

CHORUS

I waste tokens yelling at bots
Oh, I waste tokens yelling at bots
Meter keeps tickin’ while I’m typin’ “no, not that”
Context window’s full and I forgot what I forgot
Mm, I waste tokens yelling at bots

OUTRO

I’ll be back tomorrow morning with another prompt
Oh, I waste tokens yelling at bots