Autopilots for x.plural

You've seen the pitch a thousand times. Autopilot for accounting. Autopilot for recruiting. Autopilot for sales ops, legal review, support triage, invoice reconciliation.

It's a good pitch. It's also a filter.

To build autopilot for X, somebody has to raise money for X, hire for X, sell into X, and defend a category called X against three other companies doing the same. Which means X has to be large enough to hold a company inside it.

Almost none of your X's are.

The review you run every second Tuesday with a method only you use. The intake form that has three exceptions nobody can explain. The pile of things you check before sending anything to that one client. These are real, repeated, and expensive, and they will never have a seed round, because the total addressable market is you.

So the honest description of the market today is: autopilot for whatever happens to also be a market. Everything else, you still do by hand.

The word is better than the pitch

Autopilot is a well-chosen word, and I don't think most people using it mean it.

Real autopilot doesn't fly the plane instead of you. It holds a course you set, inside an envelope somebody defined in advance, and it gives control back the moment conditions leave that envelope. The pilot stays in the loop, watches the modes, and takes over. The system earns trust precisely because it is narrow and because its limits are legible.

That's the shape worth copying. Not the agent does your job, but this specific stretch is held, you set it, you can see it, you can take it back instantly.

Karpathy, who ran the actual Autopilot, has the line for it: demo is works.any(), product is works.all(). His autonomy slider is the same idea from the other end — you don't ship full autonomy, you ship a dial and you earn the right to turn it.

Plural

Here's the part that took me a while. Narrow isn't a compromise on the way to general. Narrow is where the trust comes from.

A narrow autopilot has an envelope you can say out loud in one sentence. You know when it applies and when it doesn't, so you know when to look. A general assistant has no envelope, which means every single use is a fresh gamble and you never quite stop watching.

So the right unit isn't one big assistant that does everything. It's a lot of small ones that each do a thing.

x.plural.

Twenty autopilots, each narrow, each yours, each with a limit you could state to a coworker. The second-Tuesday one. The intake one. The one that checks the three things you always forget.

None of them is a company. All of them are an afternoon.

The problem with a fleet

Twenty autopilots that don't know about each other isn't a win. It's twenty contractors who've never met, each needing the same briefing, each keeping a private copy of the truth.

Two things fix that, and neither of them is a smarter model.

Shared ground. Same files, same data, same history, same decisions. One autopilot's output has to be able to become another's input without you acting as the courier.

Shared airspace. If everything can reach everything, everything leaks. So the system has to govern where a piece of data may travel, not merely who may read it. This contract can reach the summarizer and not the outbound mail.

That's the operating system. Not a better agent. A place where a fleet can operate.

Handoff is the whole product

The lesson from aviation isn't that automation fails. It's that automation fails at the handoff.

The classic failure is mode confusion — the human doesn't know what the system currently believes it's doing, and finds out about the gap late, under load, at the worst possible moment. Almost every automation disaster reads like this.

So the design work worth doing is at the seams:

• what is this thing allowed to touch
• what is it doing right now
• what did it actually do, proven, not claimed
• how do I take the controls
• how do I get back

Get those right and expanding autonomy becomes a decision instead of a leap of faith. You widen the envelope where the thing has earned it, in the places it has earned it, and nowhere else. That's a dial you can actually turn, because you can also turn it back.

x is you

The last part is the one I like most.

When the cost of building falls far enough, the minimum viable market for a piece of software shrinks to one person, one team, one afternoon. Not because software should be disposable, but because software can finally be the shape of the work instead of the shape of a category.

Nobody is going to fund the thing you do every second Tuesday. But you could be flying it by Thursday.

The genre was never wrong. It was just rate-limited by company formation.

x.plural.

P.S. Good test for a new autopilot: can you state its envelope in one sentence? If you can't, it isn't an autopilot. It's a hope with permissions.