One is easier to think about than two

The opinions expressed in this article are my own and do not necessarily reflect those of my clients or employer. While I think for myself, I do use Claude to do most of my writing.

Here is a thing that happens.

You are a CEO, and you are not sitting on spare money looking for a home for it. You are under pressure to do something in AI, visibly, this year, from a board reading the same headlines you are. And you are funding it by taking budget away from things that were working fine. That’s the real starting condition. Not enthusiasm. Reallocation.

Two options come to you.

One is a per-seat licence for a copilot that lives inside the tools your people already use. It drafts, it summarizes, it preps you for the meeting, it writes the notes afterward, it does the first pass of the research that used to eat a Thursday. Multiply a modest amount of reclaimed time by a few thousand knowledge workers and it’s a serious number.

The other is a custom agent platform aimed squarely at one of your core processes. Two years, five engineers you haven’t hired, your own data plumbing, and a bet on an architecture that may not exist in eighteen months.

Both are good ideas. That’s the first thing worth saying. Neither is the trap.

The actual difficulty

These two things are not opposed to each other. They’re closer to unrelated. One is a car loan and the other is an Uber One subscription — both entirely sensible answers to “how do we get around,” but one is a committed asset with a maintenance bill and a resale value, and the other is a fee you can stop paying in April.

Now try to hold both of them in your head at the same time, properly, for more than about ninety seconds. Different time horizons. Different failure modes. Different units. One pays back in months and can be unwound over a weekend; the other pays back in years, or doesn’t, and by the time you know, the engineers have equity. The honest cognitive experience of doing this is that one of them keeps sliding out of focus while you concentrate on the other.

This is hard. Not hard in the sense of requiring more information — hard in the sense that the human working memory was not built for it. I made a version of this argument in a post about reasoning frameworks: running several moderately unrelated lines of thought in parallel, holding their outputs simultaneously, and noticing where they interact is not a skills problem. It’s a bandwidth problem. Smart, experienced people hit the same wall.

So what happens is what always happens under load. Kahneman and Frederick called it attribute substitution: when a question is hard, the mind quietly answers an easier one and doesn’t flag the swap. How do these two very different things relate, and how much of each should we do becomes which one is better. Which is a lovely question. It has one answer, it fits on a slide, and it ends the meeting.

And then the scorecard shows up to help. Fifteen weighted criteria, two columns, 7.4 versus 6.9.

It’s just that the bottom row is doing something the rest of the sheet isn’t. A licence fee in dollars, a vendor’s odds of existing in 2029 in vibes, the pain of retraining four thousand people in a figure someone offered up in a workshop — multiplied by weights and added together. You can’t really add a fee to a probability to a feeling. Excel will let you, of course. Excel is extremely accommodating. But the half-point gap that comes out the other end isn’t telling you what it appears to be telling you.

There’s a smaller version of this worth knowing about too. Tversky showed in Features of Similarity that similarity isn’t symmetric: we say North Korea is similar to China, rarely that China is similar to North Korea, and flipping the direction moves the answer. Bake-offs almost always ask how the challenger compares to the incumbent, never the reverse. So the incumbent becomes the reference point, its quirks get promoted to “requirements,” and the challenger picks up a penalty everywhere it’s merely different. Nobody intends this. It’s just what happens when the question has a direction and nobody notices.

What good actually looks like

Here’s the part I’ve come to find more interesting than the critique.

The executives who handle this well mostly don’t choose. They allocate.

They don’t come out of the meeting having declared a winner. They come out having decided that the copilot gets broad distribution because the value is thin but very wide, and the agent build gets funded narrowly, in one process, where the economics are concentrated enough to justify the commitment — and that the split gets revisited in two quarters when both bets have produced actual evidence. Not a decision. A mix, with a review date.

That reframing does something useful to my four layers framing from a while back — AI for people, on platforms, in processes, AI-native products. I think I let those read as a ladder, or worse, as rival claims on the same budget. They aren’t. They’re different lines in the mix, with different payback shapes, and a company can hold several at once as long as somebody is willing to say what each one is for.

Which is the skill, I think. Not picking correctly. Holding two moderately unrelated ideas in the same head, on their own terms, long enough to work out how much attention each one deserves — and being comfortable saying “both, in these proportions, and here’s when we look again.”

That’s a harder thing to do than choosing, and a much harder thing to put on a slide. But it’s closer to what the situation actually is.