Much of the current conversation about AI is framed around increasing autonomy—systems that reason, plan, and execute with less human involvement. The aspiration is understandable: reduce friction, accelerate delivery, minimise human bottlenecks.
But in my experience, that framing misses where AI actually delivers the most value.
I spend most days working inside what would be described as a human-centric AI workflow, even if I didn’t originally label it that way. It’s not a theoretical position. It’s a practiced method that already delivers results.
What Human-Centric AI Looks Like Day to Day
On a typical day, I’m working with four or five AI agents simultaneously.
I talk to ChatGPT about ideas, plans, and structure. I talk to Claude Web Chat about the same ideas—often asking it to critique or extend what ChatGPT has produced. I then present a distilled version—the why, the who for, and the big idea—to Claude Code agents inside VS Code, where the real implementation begins.
Before any code is written, I invest time in context: what problem we’re solving, why it matters, who it’s for, and what “success” actually looks like. That context becomes a specification—not just a list of tasks, but an expression of intent.
From there, the agents work. And they work fast.
I keep the agent’s reasoning visible, watching how it interprets instructions, forms assumptions, and makes decisions. Around 60% of the time, it’s remarkably accurate. The rest of the time, I intervene—correcting assumptions, redirecting effort, or clarifying intent.
Last month, I was reading an agent handoff document—the exchange notes when one Claude Code session ends and another begins. At the bottom was a link to a JSONL file in my directory.
I’d been working with AI agents for months on a knowledge extraction problem: how to capture and codify the insights that emerge from AI conversations. We’d explored dozens of technical approaches.
It turns out that the JSONL file was the solution. It contained exactly the structured knowledge we’d been trying to extract—sitting in my directory the whole time, generated automatically by the agents themselves.
When I pointed it out, they immediately understood what it was and what it meant. But in months of technical exploration, none of them had connected the dots: the thing we were trying to build already existed as a byproduct of their own process.
That’s the kind of happy accident that remains a human skill—seeing the hidden potential in something the AI treats as routine.
Crucially though, working with AI agents, nothing important happens automatically. If I want documentation, I must ask for it. If I want automated tests, I must trigger them. If I want a code review, or a feature-benefit analysis, or marketing commentary, I must explicitly request it.
The agents don’t forget to do these things—they simply don’t know they matter unless I tell them.
That’s not a limitation to work around. That’s the design that makes the collaboration productive.
AI as Capability, Not Authority
The agents I work with have access to more accumulated knowledge than I ever could. They reason across languages, frameworks, patterns, and history at extraordinary speed.
But they don’t make creative leaps between unrelated domains. They don’t recognise when a mistake reveals a hidden opportunity. And they don’t hold a durable, holistic sense of the end goal over time.
Those responsibilities remain human.
This is why treating AI as a source of truth is a category error. AI is a source of knowledge, options, and acceleration—not judgment, intent, or meaning.
In practice, the most productive setup isn’t autonomous AI. It’s supervised, conversational, interruptible AI—systems that think quickly but remain steerable.
Human-in-the-Loop Is Not a Safety Feature—It’s the Engine
Human-in-the-loop is often discussed as a control mechanism: a final approval step to prevent harm. That framing dramatically undersells its value.
In real work, the human is not a passive overseer waiting to click “approve.” The human is setting direction, shaping abstractions, questioning assumptions, and deciding when “good enough” is actually good enough.
The AI provides momentum. The human provides meaning.
This pattern holds everywhere I use AI: software development, creative work, research, strategy. Dialogue first, execution second, reflection throughout.
There is genuine satisfaction in this process—in watching ideas sharpen through conversation, in seeing knowledge assembled faster than you could manage alone, and in recognising when human judgment makes the decisive difference.
That sense of intellectual partnership matters. It’s evidence that intelligence is being augmented, not displaced.
The Bicycle, Again
The old Steve Jobs metaphor still works.
A bicycle doesn’t decide where you’re going. It doesn’t replace walking. It multiplies your effort once you choose a direction.
Human-Centric AI works the same way. The human supplies intent, values, creativity, and responsibility. The AI supplies memory, pattern recognition, synthesis, and speed.
Together, they produce something neither could achieve alone.
Where This Leads
The question isn’t whether AI should become more capable—it will. The question is are humans active or passive in the process and where our role begins.
My experience suggests the answer is: earlier than most people assume. Not just at the approval gate, but in the formation of intent, the shaping of abstractions, and the recognition of when “correct” isn’t the same as “right.”
Human-Centric AI isn’t a constraint on AI capability. It’s a recognition of where human judgment makes the decisive difference. This isn’t a future concept. It’s already here, wherever people treat AI not as an oracle, but as a powerful tool for thinking better.
