Patrick
Branigan

There’s been a fuck ton of content written about how much and how fast code can be produced by AI. Lines per hour. PRs per day. Features per week. Hell, I was sitting in a pub just this month when I ran into the father of Jesse Pickard – founder and CEO of The Mind Company – and the first thing he mentions when learning about my line of work is “my son was telling me companies out where he is are laying people off if they haven’t spent enough tokens and output enough code.”

It’s absurd how much conversation lives on that side of the ledger.

There’s another side, and I haven’t heard a lick about it in comparison.

AI might generate the code. But AI doesn’t own the consequence of the code.

The model doesn’t get paged at 2AM, refund customers, or sit in a postmortem. It doesn’t look a team in the eye on a Monday. When something goes wrong (which it will, pretty often), the consequence lands on a human, whether it be a reviewer, maintainer, orchestrator, or whoever made the call to ship.

This asymmetry was easy to overlook when generated code was a novelty. But the more code gets produced, the more consequence accumulates somewhere – and that somewhere is likely to be a human.

This changes what’s important to measure when building something. It used to be (still is) enough to ask how much someone (or some model) could produce. The more intriguing question is going to become how much can someone take ownership for? How many decisions can they stand behind? How much of the consequence are they willing to carry? How heavy of a time are they willing to endure?

It’s been well over a month since I last opened Figma to actually create something. I’ve handled every single piece of work at hand with Claude Code. That’s not to say everything has worked out well. Sometimes the experience has been less efficient. Sometimes the process has been bloated. Sometimes the output has been dead wrong. But one thing is true: regardless, it’s made output faster. That’s what people and businesses want. They want everything, faster. It’s the output that everyone is fixated on. But the only real reason I’ve been faster is because I have an interest and a comfort with learning and executing inside a terminal.

Output is now easy. Input is where the world will focus next.

Roughly half a billion people per day intentionally interface with AI tools. 5%-10% of those people are on the supply side (building, deploying or directly supporting AI products). There’s an immense amount of adoption and scale still to be done. However, as we’ve seen with the tumultuous history of VR for example, adoption is directly correlated with customer convenience and confidence (in the product and in the price).

The customer needs to be confident they can do (whatever it is they want to do) and comfortable doing (whatever it is they’re doing). I’m not so certain we’ve cracked this yet. AI products and services seem to be falling into two camps: pushed down your throat through your existing products and services without your permission via quiet integration, or soft and hard gated behind sign ups, paywalls, and other data collection layers up front.

The former makes customers uncomfortable. The latter makes customers lack confidence.

The next swell of AI innovation might just be how we interface with our AI, or rather how our AI begins more naturally interfacing with us. It will become less about what can AI do for me, and more about how comfortable and confident does AI make me feel in whatever I’m doing.