The boring case for AI

When the small tasks get cheaper, you start doing different ones. The unglamorous uses of AI are reshaping how I work.

In the movie Margin Call, there is a scene where Stanley Tucci’s character recounts how the bridge he once helped build removed a 35-mile stretch for commuters. He then starts adding up the minutes and hours of driving that this bridge saved and it turns into 1,531 years saved in commuting just by building one bridge.

Something small, if done enough times by enough people, can have a big impact. If we can remove some trivial or boring tasks, like driving a car, or reformatting a CV, we might find time for more interesting ones. I think this is one of the best use cases for AI tools like Claude Code, and one that deserves more attention. When people write about using AI for research, they usually focus on the dramatic stuff: drafting text, polishing slides, writing code, brainstorming ideas. I have written about some of these myself here and here. The small tasks that have nothing to do with research are at least as important, but they receive much less attention. This post has a few examples.

The small tasks

I am organizing a conference on housing and real estate, which involves a surprising amount of logistics: budgets, referee assignments, conflict-of-interest checks, agenda construction, messy CSV files, and eventually dozens of small organizational problems no one anticipates at the beginning.

I have used Claude Code in VS Code for all this. This is extremely useful, because Claude can read the folder context and knows a lot about the conference. For example, when I asked Claude to make a budget it already knew that one keynote speaker was flying from the US, and adjusted the budget accordingly.

Another example is about translating one document into a different format. For example, some grant agency wants a CV in a slightly different format. Just reformatting documents can take a lot of time, but now Claude reads my CV in LaTeX and the conversion is almost effortless. I also had to write a data management plan for a grant, which I know little about. I could draft the plan with a lot of project context, just because Claude Code can read the folder and knows the type of data, what the setting is, etc.

When tasks get cheaper, you do more of them

When these tasks get cheaper, you do not just save time. Instead, we start doing different ones. Economists will recognize this as Jevons paradox in miniature: when the cost of doing something falls, you do more of it, not less. Steam engines made coal use more efficient, and total coal consumption went up. Word processors made revision cheap, and papers got more drafts, not the same number of drafts written faster. The same logic applies to small research tasks.

For example, the cost of making a website is now very low. I am not a web developer and I do not even know enough HTML to be dangerous, so until recently I would not even try to make a website. But now I can. The research-relevant version of this idea is a small interactive companion to a paper. In a recent paper on housing returns across the income distribution, I built a small interactive companion where readers can change portfolio shares and watch the return gap move.

Five years ago that calibration would have been an appendix figure. Now it lives as a website where the reader can interact with the results of the paper in their own time. Every empirical economist could have built something like this five years ago, in the sense that the technology existed. Almost none of us did. The cost was not worth the benefit. The old equilibrium was a paper, a slide deck, and maybe a Twitter thread. The new equilibrium can include a small interactive companion, which I would not have thought to try one year ago.

I also wrote a guide to using Claude Code in VS Code and posted it as a website. I described what I wanted in plain English, let Claude write the HTML, and added the Paint illustrations myself (the old Paint skills will never depreciate). I had Claude remake my own website using the information on the old website and my CV.

There is a second-order version of all this that shows up in how I work. I now write in markdown instead of Word, and I redid my personal website in plain HTML on GitHub Pages instead of leaving it on Google Sites. Neither choice is about markdown or hand-written HTML being better in the abstract — they are about keeping my work in formats that Claude Code can read and edit. The first-order effect is that Claude does some of my old tasks faster. The second-order effect is that I have started reshaping my workflow to put more tasks within its reach.

The new equilibrium

I think this is the practical case for AI in academic life that gets undersold. The boring uses sidestep the objection that the dramatic ones invite. Nobody is going to argue that academia is better when I spend an afternoon reformatting my CV, or copying information between systems that almost-but-not-quite talk to each other. The conference logistics, the data management plan, the website I keep meaning to update: these are tasks I was happy to outsource. The high-profile uses of AI (writing the paper or running the analysis) provoke real concerns about quality, originality, and what is left for the human, and I share some of them. The boring uses do not.

The point is not that AI makes work disappear. It is that AI lowers the cost of tasks that were interesting but not feasible to do on a limited time-budget. Once those costs fall, the set of things we can do changes. While none of these were impossible before, I did not have the skills or the resources to do them. Now I do.