There is a point where a workflow is so well optimised that nobody in it is thinking any more.

I went back to Melbourne Business School for Alumni Day, and one idea has stayed with me. I have heard it discussed more widely lately, which I am glad of, because it is easy to miss when the habit is your own: the move from cognitive deferral to cognitive surrender.

Cognitive surrender

Deferral is a reasonable way to work. I hand a task to an AI, then I sit with what comes back and decide whether it holds. The labour has moved and the judgement has not, which is what keeps the handoff honest.

Surrender is the next step, and it is easy to miss because it feels like progress. The cognitive optimiser tightens the workflow until the pause, where a person used to weigh the result, starts to look like waste. Shorter prompts and fewer checks are each easy to defend, until the moment of decision has gone. What comes back looks finished, and because it was quick the work feels better, even though mostly the waiting got shorter.

What it costs is easier to name than it first appears. Judgement fades because you stop practising it, and a skill left idle gets less reliable. Learning fades the same way: if I no longer stay with a problem long enough to see why an answer is right or wrong, I get better at requesting the work and worse at doing it. Ownership follows the same path, because without a real decision I cannot stand behind what goes out, and the thinking that should have shaped it is not there.

The mindset does not stay in the tool either. Once I am used to accepting an answer because it looks finished, I do the same in a meeting, a design review, or a decision about what to build next, whether or not a model is in the room. All it takes is that I stop expecting a moment where I actually think.

Am I solving the right problem?

This is close to the work I do as a Principal Engineer (Data & AI), building agentic tools on data foundations. Those tools should take work off people, and that is a gain when what they remove is the tedious part. The risk is that they also take away the part where a person was supposed to understand the problem and decide.

The guard I try to keep is plain, which is why it slips once a workflow feels smooth. I still want a person to read the result and make the call, and I still want someone, often me, to ask whether we aimed at the right thing before we get efficient at answering it. A smooth path means the friction came out, which is not the same thing as the thinking having been done.

That is metacognition, reflecting on your own thinking, including the part you were about to hand away. I want the question in front of me to move from "How do I solve this problem?" to "Am I solving the right problem?" An agent can draft the pipeline or the note, but it cannot say we aimed at the wrong thing, and it will not carry what follows. I try to keep that slower step inside the work, rather than treat it as delay.

Better, not faster

What I took home needs more care than a slogan, or it sounds like a wish for everything to go slowly, which is not what I mean. Efficiency means doing the current work with less waste, and I am in favour of it, having spent years taking drag out of systems, so I know slowness is not the same as thought. Transformation is the other aim, changing the work so what we produce is clearer, more reliable, and more honest about what it knows. You can speed up a weak way of working and still be left with the weak work, only sooner. Make the work better, and speed often follows, because less time goes on repairing something that was never right.

So I try to optimise for the quality of the thing, and to let speed follow only if it does. The tools keep getting quicker, and the judgement does not arrive with them.

Use AI to make things better, not faster. Because when things are better, it is naturally faster, if fast is even a metric.