AI predictions

Everyone assumed the moat was the model. It was never the model. It was always the data loop, the workflow it slots into, the brand people trust enough to share work with, and the defaults the product picks for you. The model was the easy part.

We also assumed chat was the interface. Chat is rarely the interface. Chat is a debugger for the interface you haven’t designed yet — a way to brute-force functionality before someone sits down and decides what the actual product is supposed to do.

Most AI features fail because they answer a question the user wasn’t asking. The button gets added to the product, generates a confident-looking blob, the user nods politely, and never clicks it again. Adoption looks fine for two weeks, then trails to zero.

Real adoption comes from features that compress an existing job by 80%, not features that invent a new job. If a user wouldn’t have done the task without AI, the feature doesn’t have a foothold yet — it has a curiosity moment.

The test is brutally simple: does removing this feature make the product feel broken, or just less interesting?

Personalized retrieval — the model gets better the more it knows about the specific user, team, or repo. Workflow templates with real defaults — not blank prompts, not 30 settings, just the four common starting points pre-filled. Onboarding flows that learn from the user’s first three actions and adapt the second visit.

Anything that gets sharper with use is compounding. Anything that stays the same regardless of usage is a demo.

Smaller models running locally on capable consumer hardware, mostly invisible — surfaced as faster suggestions, not as a chat button.

Voice as a secondary interface for narrow, hands-busy tasks — driving, cooking, walking. Not voice as a primary interface for office work, which keeps not happening despite a decade of trying.

Better defaults beating better prompts. The product that picks for you wins the user who didn’t want to learn prompt engineering.

Brand trust becoming the second moat after data. Users will pick the AI product they’re willing to share private context with — which is a brand question, not a model question.