Your Data Team is a Fire Station. Almost Nothing Is an Emergency.

The message came in 15 minutes before closing and it had the word "Quick" in the subject line.

Subject: Quick Request
Body: Board meeting tomorrow at 9. Can we get retention by customer segment, three years back? Should be quick.

I want to be fair to the person who sent it. She had no way of knowing that "retention" had never actually been defined at our company, that the segment field had changed owners twice during my tenure, or that three years back crossed a CRM migration nobody had fully reconciled. From where she sat, she was asking for a number that surely existed somewhere already.

So I did what you do. I closed everything else. I told my wife I'd be late for dinner. I wrote the query, found the gap, wrote a second query, found a worse gap, and then spent two hours in a spreadsheet.

Late, at home, the hunt continued. At 11:40 PM I had a number I believed, but slept on it in case I had become delusional in my tired state. After a quick review and some last minute formatting changes, I sent it eight minutes before the meeting, under a paragraph of caveats nobody was going to read.

Hero moment, right? That's me, Fire-Fighter extraordinaire! Saving the office, one cold report at a time.

The reply came that afternoon. "This was exactly what we needed, thank you." I read it four times. If you have ever been the only person who could produce the number, you know the specific feeling. It's about as close as your average spreadsheet jockey gets to what (I imagine) bungee jumping must feel like.

Two weeks later I was cleaning up my request log and noticed that the certified retention dataset — the one that would have turned that six-hour night into a five-minute answer, permanently — had been sitting in my backlog since March.

It had been de-prioritized nine times. Once for each emergency.

We Say "Firefighting" Like It's a Compliment

Somewhere along the way our industry adopted firefighting as shorthand for the reactive scramble, and we mostly say it with a shrug and a half-smile. It's the humble brag of the data profession. Sorry I missed the sync, I was firefighting.

I've used the phrase for years without once thinking about what actually happens inside a fire station.

So I looked into it, and it turns out we have the metaphor almost exactly backwards. In 2020, U.S. fire departments logged just under 27 million runs. Roughly 64 percent were EMS and rescue. Fires accounted for about 4 percent.

Four percent.

A fire station is not a building full of people fighting fires. It's a building full of people getting ready, interrupted occasionally by a fire. That's the job we actually signed up for. We just described it wrong, and then started living up to the description.

What Happens When the Station Gets Busy

If you wear your utilization like a merit badge, I apologize in advance.

The fire service tracks how busy a company is with Unit Hour Utilization. Its the percentage of time a unit is committed to an incident, from dispatch to back-in-service. There's no single national standard, but the traditional ceiling for a 24-hour fire company sits around 25 to 30 percent.

Thirty percent. That's the number at which a fire department starts talking about adding units.

Now go look at your team's calendar.

The Illinois Fire Chiefs Association put the reasoning plainly: at 30 percent utilization, a 24-hour company no longer has time for inspections, training new personnel, or public education. What dies first is prevention. Prevention is the only elastic item on the schedule. Nobody can defer a structure fire. Everybody can defer a hose test.

And it compounds. Capacity doesn't degrade in a straight line; anyone who has studied queue behavior or traffic patterns knows the shape. As utilization climbs toward full, wait times curve upward and then go nearly vertical. A team at 70 percent has slack that absorbs a surprise. A team at 95 percent has none, so every surprise becomes a delay and every delay becomes its own surprise downstream.

Layer the prevention loss on top of that curve and you get the machine that ate my March backlog. The busier the team, the less prevention. The less prevention, the more emergencies. The more emergencies, the busier the team.

Station Log — The 30 Percent Ceiling
Find your team's real committed-time number using whatever you can actually measure: tickets, hours, calendar blocks. The key here is you are only tracking requests from outside your team. Handling tech debt or upgrading your BI template doesn't go into this number.

Left unprotected, north of 70 percent is a capacity problem smoldering and waiting to burn. Reserve the difference on the calendar with a name on it, because unnamed time is the first thing taken. That block is the entire prevention program, and it will not survive as a good intention.

Real Stations Aren't Busy. They're Ready.

What does a well-run station do with the other 96 percent? It maintains the thing that has to work during the actual emergency.

Apparatus checks every shift. Hoses pressure-tested on a schedule. Gear inspected, cleaned, and pulled from service when it fails. Trucks washed — not because anyone is precious about a clean truck, but because you cannot find a leak on a dirty engine.

We have all of this. We just don't dignify it with a maintenance schedule. The dataset with fourteen measures nobody renamed. The report still pointed at the old warehouse because migrating it was never a convenient Tuesday. The test that's been soft failing for five months and has therefore stopped meaning anything. I'll spare you the sermon about the messy Semantic Model and add only the station framing: the crew doesn't clean the gear because it's dirty. They clean it because the next call comes at 3 AM in the rain, and whatever condition the gear is in at that moment is the condition you chose weeks ago when you didn't protect the maintenance schedule.

Fire departments also do something called pre-incident planning, formalized in NFPA 1620. Crews walk the buildings in their district before anything is on fire. They map the hydrants and the standpipe connections, note which wall has the gas shutoff, where the sprinkler riser lives, whether there's chemical storage behind the loading dock. They write it down and keep it current, for buildings that will statistically never burn.

The purpose isn't to predict the fire. It's to make sure that when the fire happens, nobody is standing in a parking lot at 3 AM trying to figure out where the water is.

We call this documentation, which is a much worse name for it, and we treat it as something you produce after a project (at some point, but not right now) instead of something you carry into one.

Station Log — Pre-Plan the Building
For every source system in your district, write down four things before you need them: where the data lives and who owns it, when it lands and what breaks if it's late, which fields are trustworthy and which are free-text landmines, and what the business calls each thing versus what the schema calls it. The test of a good pre-plan is whether a teammate could answer a question from it at 4:30 on a Friday without calling you — and that teammate is eventually going to be you, with none of today's context loaded.

Fire Prevention Week

Every October, firefighters walk into elementary schools with a truck and a helmet and become instant icons to seven-year-olds everywhere. They teach kids to crawl under smoke and check a door for heat before opening it. Fire Prevention Week has run for more than a century, making it the longest-running public health observance in the country.

It may be the cheapest hour on the department's whole calendar. A kid who knows two rules costs almost nothing. A rescue costs a great deal, and the department knows the exact figure.

Half the "urgent" requests I've ever received were urgent because the requester didn't know the answer already existed, didn't know how to filter the report they already had, or didn't know the number they wanted had a definition and an owner.

That isn't their failure. Nobody ever taught them. We built a platform, handed people a link, and called it enablement.

Run the office hour. Do the twenty-minute walk-through. Record the four-minute clip on how to read the tool-tip. Teaching a department head to use a slicer removes an entire class of future emergency from your queue, permanently, without you ever being in the room. None of it feels like real work while you're doing it, which is precisely why it never makes the calendar unless you defend it.

Hand Out Smoke Alarms

Working smoke alarms cut the risk of dying in a home fire by roughly 60 percent. That's an absurd return on a twenty-dollar device, and it's why departments send crews door to door installing them for free. They aren't waiting to be called. They're going out and reducing the severity of a fire that hasn't started yet.

The mechanism is worth being precise about: the alarm doesn't stop the fire, it changes when you find out. Detection moves the fire from blaze back to smolder, and a smolder is a smaller problem with a cheaper solution.

Every emergency data request has a moment where it is only smoldering. The reconciliation broke on the 3rd. The source system added a status value on the 11th. The refresh started failing silently on the 14th.

In every case, something in the pipeline knew before any human did, and the only reason it became a 5-alarm scramble is that nothing was configured to say so out loud.

Put alarms in the building.

Freshness checks, row count deltas, schema change detection, referential tests that fail loud, threshold alerts on the metrics that actually drive decisions.

There's a failure mode here too, and the fire service documented it long before we did: smoke alarms get disabled because of nuisance activations. People pull the battery out of the thing that saves their life because it goes off every time they make toast.

Your alert channel is the same battery. If it fires forty times a week and thirty-eight are noise, it's already been muted — maybe not in the settings, but definitely in the heads of everyone receiving it.

Station Log — Alarms That Earn Their Battery
Every alert needs a named owner, a documented action, and a threshold someone actually tuned. If nobody can tell you what to do when it fires, it isn't an alert, it's noise. Review the list quarterly and delete anything that has fired repeatedly without producing an action. That one isn't detecting a fire, it's training your team to ignore the sound.

Renting the Hall

There is one more thing stations do that has no obvious operational justification at all. They rent out the hall. Pancake breakfasts, blood drives, Scout meetings. They park the engine at the farmers market or the classic car show and let kids climb on it.

From an efficiency standpoint this makes no sense. From a risk standpoint it's brilliant, because the community learns who these people are before the worst day of their life, and familiarity lowers the barrier to calling early.

Y'all. We're famously bad at this. We sit in our own channel, join the meeting (with camera off) where our agenda item lives, and leave when it's done. "Have to drop, putting out a fire this afternoon."

Then we wonder why nobody loops us in until the deck is already built.

So, sit in the ops stand-up with nothing to present. Answer the dumb question in the channel where it was asked instead of DM'ing the reply. The difference between "we should ask the data team next week" and "wait, let me just ask right now" is usually about four days of smoldering.

The Rare Real Fire

Let's run the drill again, from a different station.

The alert had already fired on Tuesday. A regional retention figure moved outside its band and the check that caught it posted to a channel four people read.

Nobody panicked. At that point it was a curiosity.

The message arrives: Board meeting tomorrow at 9. Retention by segment, three years.

Retention has one definition. It has an owner, an effective date, and a tool-tip explaining how reactivated accounts are handled, because that argument happened in a room eighteen months ago and does not need a sequel. The certified dataset has been published for months and already powers 8 other reports, refreshed daily at 6 AM, and already carries segment as a proper dimension.

That CRM migration issue? It's in the pre-plan, with a two-week seam where the join logic changes and a note on which side of it you can trust.

The build takes forty minutes. Most of that is the visual.

The number, when it lands, is the one that moved on Tuesday, but you already knew. So the email that goes out on the pre-meeting agenda isn't a number buried in caveats. It's a number, the reason it moved, and one sentence on what it means for the decision on the table. There's time for the CFO to read it, think, and ask a follow-up before he walks into the room.

Nobody stays late. Nobody thanks you. The meeting is uneventful.

That's what winning looks like. You get a boring Thursday and an on-time dinner. The hardest part of running a station instead of a scramble isn't building the datasets, it's giving up the identity of the person who saves the day.

But look at what carried the day! The alarm bought two days. The pre-plan meant nobody hunted for the water. The certified dataset meant the model was already built. The definition meant nobody re-litigated it in the room. The fire never got past smoldering, and the entire reason is a stack of unglamorous work done on quiet afternoons that nobody wrote a thank-you note about.

But you can still be proud.

Four percent of Fire Station runs are fires. The rest of the shift is the job.

Spend it accordingly.