I called my first customer at 5am to come get my dad.

I didn’t call the main line. I scrolled to his cell number—the one he’d given me nearly four years earlier to the day, back when he was the first guy willing to try our funeral home software—and pressed Call. He picked up on the second ring. He was at my childhood home within the hour. Then they drove thirty minutes back, and he carried out the rest of his day. Families sometimes look at a funeral bill and wonder why “Transfer of Deceased” is such an expensive line item. This is why. The price reflects the job, and the job is harder than the price suggests.

I spent five and a half years building software for that job, and I’d learned it from nearly every seat in the room. That morning, I was in the last one.

”Do things that don’t scale” is not a phase

There are businesses where the unit economics look absurd from the outside and make perfect sense once you see what’s happening inside. Funeral homes are one of them.

Paul Graham’s line has become so ambient we barely hear it. You do unscalable things early so you can learn, and then you scale them. The scaffolding comes down when the building is up.

The funeral industry does not take down the scaffolding. The scaffolding is the building. It is still, overwhelmingly, family-run in the literal sense, because the things that make it good don’t compress into a playbook. My customer answered his phone at 5am because that is what the job is as the baseline, not as a heroic exception. His father did it before him.

And here’s what’s changed: in a world where AI is driving nearly every commodity interaction, the value of high-quality human-to-human service is skyrocketing. The things that can’t be automated are becoming the things worth the most. The unscalable parts are the product.

I watched my mother use my software

Four years ago we built a digital register book: the tablet funeral homes set out at a service so guests can sign in and leave a memory. Last week I watched my mother use it, on the day of my dad’s service. What I saw was something different than in the hundreds of customer calls and funeral home visits.

It wasn’t that my mother was in her seventies or the UI was unfriendly. It was that there were four of us standing around the tablet, each with our own opinions, each backseat-driving every tap, each representing a different relationship to my dad and therefore a different view of the right answer. And the detail we were stuck on, the thing four grieving people debated in real-time, minutes before I was to deliver a eulogy, was how to enter a street address. Specifically, how to enter the apartment number. Does it go on its own line? After a comma? Do I tap? Do I press Tab? In the address field or somewhere else? Four people, four answers, one tablet, one address we were trying to type into a form.

The software had been designed to be used by a user. We were not a single user. We were a committee of grieving people negotiating in real-time, and the software had no idea what to do with us.

That’s the lesson I didn’t have before. Software for moments like this doesn’t serve an individual. It serves a group: dynamic, overlapping, each member bringing their own context, and it has to hold all of that without forcing anyone into the passenger seat. Most enterprise software assumes a single operator with a clear goal. Most life does not look like that.

I also don’t think the software will ever be “good enough” in an absolute sense, and I’ve stopped thinking that’s the interesting question. My family loved the end result of that service. The software was not the main event or even an afterthought. It was the last detail they remembered, if they remembered it at all. But they did remember that all the register book addresses were formatted and printed on mailing labels for them. That is exactly what good software here should aspire to: to be invisible behind the thing actually worth remembering.

The review, and the industry showing up

I wrote my customer a review afterwards. It was longer than reviews usually are because I knew exactly what it had cost him to be that good, which parts of his week had been hard in ways we wouldn’t see. A normal happy customer says “they were wonderful.” I could say why. And the funeral profession showed up, with directors, suppliers, and folks at other software companies sending messages & gifts in the practiced and quiet way they reach out to their own. I don’t recommend the path that qualifies you to notice either of those things, but I’m forever grateful for both.

Different industry, same shape

I run engineering at an AI-native insurance company now, and I’ve realized the exact same dynamic exists here. I’d be lying if I said the parallel hadn’t been sitting on my chest for months.

A claim is a committee

The day someone files a claim is, for them, a version of my 5am phone call. A building burned. A truck rolled. A worker was hurt. The thing standing between them and getting whole again is a process that, to put it mildly, is not designed to feel human. It has a messy middle: weeks or months between “something bad happened” and “here is a check,” experienced mostly as silence punctuated by forms. Even a simple slip-and-fall is not immune to this level of complexity.

And the claimant is not alone in that file. A claim is a committee: insured, broker, adjuster, expert, carrier, sometimes counsel, each with their own relationship to what happened, each backseat-driving every decision. Software built for a user breaks in exactly the place it broke around that tablet. The job is to hold the group.

The question I keep asking

How do you scale what a great funeral director does? You can’t hire a million of them for insurance claims. But you can’t self-serve-portal your way out of it either: the claimant is in distress and the judgment required is real. The model where the software serves the enterprise, and the human in distress is treated as incidental, is the dominant one in claims too. It’s the one I want to build the opposite of.

What AI is actually for

AI is the only way the math works. Not as a replacement for the adjuster or the director—AI as the thing that absorbs the scaffolding. It simplifies the complexity. The intake. The triage. The document chasing. The hundredth form. The thousandth summary. So the humans-in-the-loop can spend their attention on the parts that matter: the holding, the judgment, the phone call.

Used well, AI doesn’t dilute the human moments. It concentrates them. It drives the cost of “do things that don’t scale” down by orders of magnitude, which means you can afford to do more of them, for more people, in more places.

And because of that, the 5am phone call still gets picked up on the second ring. The adjuster has time to actually talk to the claimant. The director has time to actually sit with the family.

There’s a second, underrated thing AI does here: it captures the invisible work. Funeral homes and claims shops have run on tacit knowledge for generations. It lives in the heads of the best operators and walks out the door when they retire. AI is the first technology that can catch that knowledge as it happens to turn it into something the whole organization gets better at over time. It’s not a replacement for judgment. It’s a memory for it.

My first customer answered his phone at 5am. Four years after I built it, my mother used my software in a room with three other people who each had their own idea of how it should go, and it didn’t quite know what to do with us. And in the end it didn’t even matter, because the thing they’ll remember isn’t the software.

The businesses that matter most are the ones where scale is the enemy of the thing the customer actually came for. The job of software in those businesses—and of AI in those businesses, now—is not to replace the unscalable parts. It’s to protect them. To make it easier to answer the phone at 5am, forever.