10 Questions for AI Leaders: Neil King

Neil King's wife suggested they open a rubber duck museum. He said yes - and it shaped everything he believes about AI, too.

The HubSpot engineer who builds self-serve campaign tools thinks the judgment, the instinct, and the decision about what's worth putting on the shelf should always stay human. Read on for insights into how AI is being built by the man helping to build it. And lots on Rubber Ducks.

Mindstream: When you think about AI's current trajectory, what excites you the most? Does anything concern you?

Neil King: What excites me is how fast the distance is closing between wanting to do something and being able to do it. Three years ago I couldn't write code. I took an AI engineering role anyway, and my first project was a Claude-to-Looker connector for our marketing org.

Now I'm building Campaign Workbench, which takes a marketer from an idea to a live AI personalized workflow without an engineer in the middle. The marketers using it aren't becoming engineers - they just stopped needing one to get started.

The same thing is true at a much smaller scale. My wife and I own The Rubber Duck Museum, a small shop in Tsawwassen, BC, and I run the payroll, scheduling and books from Point Roberts, Washington. There are two of us. 

AI is why the back office doesn't eat the whole week, and why that time goes to the floor instead, finding unique ducks and making the shop something people remember. Two people can run something that should take ten, and the part we kept is the part we actually love.

My concern is that we let AI do our thinking, not just our work. Misuse by people with bad intent is real and stopped being science fiction a while ago.

But the one I feel more comes from my entertainment background. Creative work is about to get optimized toward whatever tests well, and the most memorable work we have is exactly what would fail that test. 

Pulp Fiction is told out of order, momentum broken on purpose. I build these systems, so I know that if you fed that script to a model tuned on good writing, it would flag every one of those choices and clean them right out.

That structure is the entire reason the film is a landmark. The danger we stop trusting the human instincts that matter most.

Let AI do the dirty work. The judgment, the creativity, and the instinct stay human.

The litmus test I use when I build is whether the thing helps someone do what they're already good at, better.

Take hours of report pulling and analysis off a marketer's plate and they get that time back for the creative work and the strategic thinking they were hired to do in the first place. If what I'm building doesn't do that, I'm building the wrong thing.

Same principle, running backwards. I automate whatever is standing between someone and the work they're actually good at. Cross-border import invoices are in the way. Holding several workstreams in my head is in the way. Neither of those is a skill, they're the toll I pay to use the skill, so both went to AI without much thought.

What I won't automate is easier to describe. At the museum, we'll use AI to research a duck, verify where it came from, chase down a detail.

We never use it to decide how we present that duck or how we tell its story. That's the whole product. Someone walks into a small shop and what makes it worth the visit is that two actual people looked at the thing and decided what was interesting about it. The research is fair game. The deciding isn't.

Being able to say precisely what you want. That sounds too simple to be a skill, but it's the whole game. Almost everyone starts out asking "read this and tell me what you think," myself very much included.

You get back something generic and conclude the tool is mediocre. It isn't. It just got asked a mediocre question. The gap between that and "here's the document, here's who it's for, here's what we've already tried, tell me where the argument is weak" is enormous, and none of it is technical.

It's knowing what you actually want before you ask for it, which most of us were never required to do out loud.

Yes. Economically, financially, and in how governments are reacting, expectations are running well ahead of reality. That doesn't mean the technology isn't real. It means a lot of what's built on top of it isn't.

You can see it without reading a market analysis. Store signage, small business sites, social posts, video. The AI nobody proofread is obvious now, and so is the AI that exists purely to get a click.

What's missing is always the same thing: nobody looked at the output and made a call about it. The companies still standing after the correction will be the ones who kept a person in the loop when it would have been cheaper not to.

Almost entirely. Two years ago I was rewriting the same prompt fifteen times to nudge an output closer to what I wanted. That was the job.

Now I come in and check how many campaigns launched through the workbench overnight. If there's a bug, I drop the ticket into OpenCode and the model comes back with a spec: here's the problem, here's the fix, here's how I'll implement it.

I say go, and an hour later we've deployed, and the marketer who was stuck isn't. We also built a self-support system that catches errors, files the ticket for the user, and writes it in a format another model can act on, so a chunk of that loop runs without me in it at all.

I spend far less time on minute code changes and much more time on whether a thing holds up at scale, whether it's even the right thing to build, and how it fits with everything else running alongside it. I was hired to implement AI, so I've never sat around waiting to see what it can't do.

My wife said, "We should open a rubber duck museum." I said yes.

That was the whole decision. It started as a small display in the back of a little shop in Point Roberts, which is a quiet place where a rubber duck museum is not an obviously good business idea.

Now it's a real business in another country, and it's shaped most of what I believe about work. Every strong opinion I have about AI comes from running a shop where the entire value is that two people decided what was worth putting on the shelf.

I'm not saying whimsical ideas always pay off. Plenty don't. But the decisions that reshape your life almost never announce themselves. That one took about four seconds.

Two things - the first is that the gap between the story and the reality is wider than anyone admits.

We use these tools daily and they're genuinely powerful, but the conversation is almost entirely about what AI is about to do rather than what it's doing. The people closest to it are the least breathless about it, because we're the ones who watched it fail this morning. Nobody says that loudly, because skepticism reads as being behind.

The second is the job we trained for is already changing underneath us, and it's happened more than once in a short span.

That's not a complaint, and I'm not predicting anyone gets replaced. But the thing people feel and don't name is that the skills you spent a decade sharpening might not be the ones that matter next, and there's no way to know which will be.

I came in without a traditional engineering background, so I've had less to unlearn. For people who built a career on craft they're proud of, that's harder to sit with than we admit.

Agents. Watching one go off, do its own research, make its own calls and come back to report was the moment the ceiling moved. It behaved less like a tool and more like a manager with specialists under it, and I started attempting things I wouldn't have proposed six months earlier.

But the real rethink came from a less glamorous direction. I built Campaign Workbench the way an engineer would, then put marketers in front of it and watched them get lost on the words. System prompt. User prompt. Language that's invisible to me and meaningless to them.

The capability was there, and it didn't matter, because the door was the wrong shape. It took a long walk one morning, thinking about how the best products hand something genuinely complex to a person who has never seen it before and that person just knows what to do.

That changed everything: how the tool worked, what it called things, what it asked for. Same model, completely different product. The breakthrough that mattered wasn't a model release. It was realizing the hard part had moved from capability to translation.

In five years, AI will make people feel unrecognizable about their work.

Not in a bad way. If you handed someone their own job description from five years ago, they'd have trouble finding themselves in it. None of these shifts announce themselves while they're happening.

There's no memo. You look up one day and your hands are doing something different, and what you were good at last year isn't quite what's being asked of you now.

That lands as restless before it lands as exciting, and I'd rather say so than pretend otherwise. But every shift I've been through has moved the work toward the parts that need a person. Less time producing the thing, more time deciding whether the thing is any good. If that's where this ends up, unrecognizable is a fair trade.

Neil King is an AI engineer at HubSpot, where he builds self-serve AI tools that let marketers go from idea to live AI-personalized campaign without an engineer in the middle. See what HubSpot is building right now by watching the brand-new Spotlight.

Neil also co-owns The Rubber Duck Museum with his wife, a small shop in Tsawwassen, BC, which he runs from Point Roberts, Washington.