Learning at work is work

Gianpiero Petriglieri has a piece in MIT Sloan Management Review this month, “Learning for a Living”, with the subtitle “Learning at work is work, and we must make space for it.” It did the rounds on Hacker News last week, and a lot of the comments I read were people saying their company talks about learning a lot and gives them no time for it. One line from the article I’ve already quoted to two people: “Nothing truly novel, nothing that matters, is ever learned with ease.”

I agree with him, and I’ve been the manager he’s writing about. For a few years our policy on learning was that we “encouraged” it. What that meant was that people learned on their own evenings, or they learned in the middle of a sprint by quietly padding an estimate, which is learning disguised as work and makes every plan a bit of a lie. I spent most of the last couple of years doing a data science certificate on evenings and weekends, so I can’t claim I didn’t know what that costs.

What we do now is very mechanical, on purpose. Every other Friday afternoon, four hours, is learning time. Over a two-week, 80-hour stretch that’s 5%, which sounds small and turns out to be about the most a team our size can protect without the roadmap noticing. It goes on the calendar like a meeting. It shows up in sprint planning as reduced capacity, the same way on-call does, so nobody is borrowing against it. People pick what they work on, with one rule: you write two or three paragraphs in a shared doc afterwards about what you tried. I don’t grade them. The point is that the next person who wants to try the same thing starts from somewhere.

The first version failed. We tried “10% time, whenever you like”, and almost nobody used it, because there was never a week where it felt OK to leave the team’s work for your own. Making it the same afternoon for everyone fixed that, because then nobody is the one person who’s not answering Slack.

Some of what’s come out of it has been directly useful: a regex audit after the Cloudflare outage, a couple of people getting comfortable with Terraform, someone rewriting a flaky test harness for fun. Most of it hasn’t been, at least not on any timescale I can measure, and I’ve decided I’m fine with that. If I only funded the learning I could predict the payoff of, it would just be work with a different label on it.