The Drip
Gallup’s latest look at Gen Z and AI at work shows a shift in sentiment: the share who described themselves as excited fell from 36% to 22%, while the share who described themselves as angry rose from 22% to 31%. That is a useful counterweight to the workplace mandate most people keep hearing: figure out AI, use AI, move faster. The mandate is racing ahead of people's lived experience of what these tools can actually do.
Inside The Bottle
Ask a model to pick a number between zero and one hundred and, for a while, the answer kept coming back 46. Different sessions, even different tools: 46. (Or maybe it was 42?) Either way, the concept holds.
It is a simple example of a much bigger workplace problem. A leader can ask AI for a sales plan, an efficiency plan, or a new process and get something that looks polished and useful. But without the specificity of the company, the customer, and the people doing the work, everybody is being served some version of 46.
The value shows up when human context and judgment enter the equation. As Justin put it, AI works best as the assistant, not the leader. Our knowledge of the problem, our opinion about what matters, and our ability to recognize whether an answer fits are what make the output useful.
That gets harder inside a team. Each person develops their own context, preferences, files, and agent behavior. One person's assistant knows something another person's does not. A 50-person company can quickly end up with a lot of generally similar work that is just different enough to create chaos.
At the same time, the barrier to making things has collapsed. More apps, documents, analyses, and workflows are appearing because they can. The constraint has moved. It is no longer "can we make this?" It is "is this the right thing, and can the people who need it absorb it?"
That is why usage is not the same as adoption. Tokens consumed are a measure of token consumption. Artifact counts are a measure of artifact creation. Neither tells you whether people trust the system, whether a decision got better, or whether the work helped.
A useful progression is crawl, walk, run: start in chat while the idea is still being shaped. Move into a harness when the idea has earned the work. Reach for an API only when the process is understood well enough to run with limited human involvement. The goal is not to remove the human everywhere. It is to be deliberate about where the human adds the value.
Lab Notes
| ■ | Justin's note: When a session keeps missing what I am after, that is usually a cue that I need to do more thinking instead of asking it to try again. |
| ■ | Kellan's note: Start in chat, not in the harness. The capability itself can pull you into building something before you have decided it should exist. |
What Stopped Our Scroll
| ■ | Don’t be a meat proxy: “Meat proxy” is a sharp name for the person who pastes AI output into a thread without evaluating it. The fix is simple: read it, understand it, validate it, then respond in your own words. |
|
