Management Office, or: How We Learned to Stop Worrying and Love Metrics

AI didn’t change our mindset about metrics. It made checking one cost five minutes instead of a week. That’s how we stopped worrying and learned to love them.

Management Office, or: How We Learned to Stop Worrying and Love Metrics
FIG. 01 — A DASHBOARD WITH TWENTY METRICS

“We want to add a couple of metrics to the dash.”

“Got it. Building a case against someone?”

That was about the level of trust between us and our IT teams whenever the conversation turned to numbers on a dashboard.

The Management Office would show up with dashboards and think it was handing over a tool. People on the other side of the screen saw something else. Fair enough: it was someone on the MO team who picked the periods, the formulas, the filters, the slices.

Even with 20 metrics on the screen and no one saying “number X means decision Y,” it still looked like somebody’s perspective. “Right, here comes exactly what they want us to see.”

Every year we had a new explanation for it. Our people weren’t data-driven enough. Or managers were bad at selling decisions. Or there were trust issues. Or nobody around here was collaborative enough.

FIG. 01B — THE LABELS WE USED

There was another problem: dashboards, and metrics in general, were expensive and slow to build.

To test one simple idea you had to:

That’s maybe 20% of the hassle. I won’t list the other 80%. Most of it sounds hilarious now.

Short version: the question took 30 seconds, the answer took hours or days.

FIG. 02 — WHAT IT TOOK TO ANSWER ONE QUESTION

Three days of work, one sentence

You’d spend three days building a dashboard, and then some bad guy (that’s what he was, in the moment) would say: “I think you might’ve got something wrong here.” The reasonable response would be: “Fine, show me what looks off.” In practice, though, you’d already grown attached to the results. It wasn’t just a dashboard with some formulas. Someone dismissed three days of your life. Great job. Now trash it all.

The same thing was happening on the other side. The person wasn’t seeing a chart. They were seeing a chart brought in by a manager, which already made it a bit of a statement. Sure, a manager without a point of view isn’t much of a manager, but you can see where it comes from. An ordinary argument about numbers could easily turn into a small political earthquake.

FIG. 03 — A THREE-DAY BUILD, PLOTTED AS A SEISMOGRAM

Then AI happened

We had scripts pulling data out of Jira, GitLab, HR systems, and other sources into a database. We rewrote them a bit, tidied up the schemas, and deployed internal MCP servers for Postgres and Superset (there are MCP servers for Jira and plenty of other things too, but that’s beside the point).

FIG. 04 — THE DATA PATH, FROM SOURCES TO QUERY

Now, when we need to check something in the data, we open Claude Code or Codex, ask in plain English, and get an instant answer. If we need the metric later or want to hand it over as a chart instead of a terminal window, we just ask the agent to push it into Superset.

Five minutes for the whole thing, three of them spent eating a sandwich or reading Slack messages. And it turned out I wasn’t the only one who found it ridiculously convenient. So as the MO, we stopped selling dashboards and started selling the ability to build your own dashboards in five minutes and make your own decisions.

FIG. 05 — TERMINAL QUERY TO SUPERSET CHART

A developer doesn’t trust your numbers? Five minutes and they do. A team lead wants a different slice? By the time their tea has steeped, they’ve already got one. Managers have it easier too. There’s nobody to argue with. Turns out part of our “cultural problem” wasn’t all that cultural.

AI didn’t change anybody’s mindset, but it did fix a real problem: checking used to be expensive. When a check takes a week, the argument turns political. When it takes five minutes, there’s barely an argument.

Goodhart’s law is the same story. We haven’t beaten it: make one number the target and people will still find a way to game it. But now it’s cheap to check ten more alongside it. Arguments get settled without the shouting matches.

FIG. 06 — ONE TARGET, TEN NUMBERS NEXT TO IT

Fifteen years ago, if you’d told us we’d one day skip the scripts and the SQL, type whatever we wanted into a console in plain English, and get a usable answer, we wouldn’t have laughed. We’d have asked if you were feeling okay. Today it’s just how we work, and it’s what finally made us stop worrying and love our metrics.