The Work That Doesn’t Scale

There is a kind of work whose value becomes difficult to see when we measure it by how far it travels.

A conversation may matter enormously to one person and produce nothing that can be counted afterward. An editor can spend an hour helping a writer understand why something isn't working and have no additional artifact to show for the hour. A reporter can spend years becoming familiar with a community before that accumulated context becomes important to a particular story.

None of this means the work is somehow purer because it remained small. It means reach and value aren't measuring the same thing.

I've become interested in that distinction because so much of modern work is organized around multiplication. If something is useful, we naturally ask how to make more of it, distribute it farther, automate it or reduce the amount of labor required for each additional person served.

Often that's exactly the right question.

Printing allowed one person's writing to reach people they would never meet. Radio extended a voice across a region. Software lets the same capability be used thousands or millions of times without rebuilding it for each person. A newspaper would be a strange institution if it regarded reaching more readers as evidence that something had gone wrong.

Scale can be extraordinarily useful.

It just doesn't scale everything.

Suppose I write an article and ten thousand people read it. The artifact scaled beautifully. Whatever relationship each reader forms with the article still happens individually. One person recognizes something useful in it, another disagrees, another misunderstands it, another skims the first two paragraphs and leaves, and most may never think about it again.

Increasing distribution multiplies encounters.

It doesn't make the encounters identical.

This becomes easier to see in work that depends heavily on context. A teacher can reuse a lesson, but the moment when a particular student finally understands something may depend on noticing exactly where that student is stuck. A physician can rely on knowledge developed across millions of cases while still needing to determine what is happening with the person in the room. A reporter can publish one story to an entire community while building the relationships that made the reporting possible through hundreds of conversations that never became stories themselves.

The scalable part and the non-scalable part can belong to the same work.

That changes the question. Instead of deciding that something either scales or doesn't, we can ask what scales cheaply, what scales with additional cost and what changes character when we try to multiply it.

Sometimes the answer is surprising.

Software can make a process available to thousands of people, but understanding an unusual failure may still require somebody to look closely at one case. A community directory can publish thousands of records through the same interface, while keeping those records trustworthy may depend on individual organizations noticing changes and telling us about them. A newspaper can distribute an edition widely, but the local knowledge contained in it accumulates through encounters that can't simply be duplicated by increasing server capacity.

The infrastructure scales differently from the relationships that feed it.

This is where optimization can become misleading. Once we've found something valuable, we tend to look for the repeated operation and remove whatever appears to prevent repetition. Sometimes that reveals unnecessary friction. Other times the apparent friction is where important information enters.

A conversation takes too long because the situation is genuinely ambiguous.

An editor asks another question because the first answer changed their understanding.

A customer doesn't fit the workflow because the workflow doesn't represent something about the customer.

A reporter keeps talking after obtaining the quote because the quote isn't the only reason for the conversation.

Removing those things may increase throughput while decreasing our ability to notice what differs from one case to another.

That doesn't mean we should preserve inefficiency merely because it feels human. Plenty of repetitive work can disappear without damaging anything important. If software can remember an advertiser's approved price, nobody becomes wiser by reconstructing it from email every month. If a form can collect information correctly the first time, requiring a phone call isn't relationship-building. It's friction.

The difficult part is distinguishing unnecessary repetition from necessary attention.

I think that's where some discussions of scale go wrong. We treat the human involvement as either sacred or wasteful before asking what the person is actually contributing.

If they're carrying information the system could preserve, perhaps we should preserve it.

If they're making the same mechanical decision a thousand times, perhaps we should automate it.

If they're noticing differences the system doesn't know how to represent, removing them may remove the mechanism by which reality gets another turn.

Scale isn't the enemy in any of those cases. It's a change in architecture.

We decide which parts can be repeated, which can be shared and which still need somewhere for particulars to enter.

This also changes how I think about work that receives little visible response. The older temptation is to say that some work matters precisely because it remains quiet, unrecognized or resistant to amplification. There's something attractive about that idea because it protects work from the judgment of metrics.

It can also protect bad work from evidence.

If nobody reads something, perhaps it was ahead of its time. Perhaps it reached exactly the one person who needed it. Perhaps its effects are subtle and will take years to appear.

Perhaps nobody read it because it wasn't very good.

All of those possibilities need to remain available.

Persistence isn't proof of importance any more than popularity is proof of value. The absence of measurable response can mean our measurement is poor, that the effect is delayed, that the work operates through something we're not counting, or simply that the work isn't doing what we hoped.

We still need feedback.

The challenge is choosing feedback appropriate to the work.

A newspaper article may reasonably care about readership while also asking whether the reporting remained accurate and useful after the initial attention disappeared. A community directory might care less about time on page than whether people can find correct information. A relationship can't be evaluated by the number of conversations it produces, but that doesn't mean the people inside it have no way to notice whether it is healthy.

Not everything needs the same instrument.

Once that distinction is visible, smallness no longer needs to become a virtue. Some things should remain small because the cost of preserving context rises rapidly with the number of participants. Other things begin small only because we haven't yet discovered how to preserve what matters while expanding them.

Occasionally technology changes that boundary.

A handwritten letter once required another act of writing for every recipient. Printing separated composition from reproduction. Recorded music separated performance from listening. Software separated some forms of service from the labor previously required to perform them each time.

Each technology allowed something to scale that previously couldn't.

Something else remained local to the encounter.

That boundary will keep moving.

AI is moving it again. Work that recently appeared to require individual human attention can increasingly be reproduced, adapted or generated at very low marginal cost. That doesn't tell us whether the resulting interaction preserves what mattered about the original one. Sometimes it will. Sometimes we'll discover that we were paying humans to perform repetition rather than judgment. In other cases, removing the human may reveal that the relationship itself was carrying information we hadn't learned to measure.

We won't know by declaring either outcome in advance.

We'll have to watch what changes.

This is why I no longer think the interesting category is “work that doesn't scale.” Almost anything can be distributed farther, copied more cheaply or represented through another medium. The important question is what happens to the work as we do it.

What survives multiplication?

What becomes cheaper without becoming worse?

What requires another person because we haven't automated it yet, and what requires another person because that particular encounter is part of the information?

Those are different questions.

The goal isn't to protect meaningful work from scale. It's to avoid scaling away the part that made the work meaningful.

Sometimes one conversation becomes an article that reaches thousands of people. Sometimes an article brings one person into a conversation that could never have been designed for thousands.

Both movements matter.

The artifact can travel farther than the relationship that produced it, and the relationship can produce more than any artifact manages to carry.

Knowing the difference tells us what we can multiply.

It also tells us what we still have to show up for.