The half-life of an AI tool

Every category of software has a half-life — the median time it takes before tools in that category get replaced by something better. Word processors have a long half-life; the dominant product in 2026 is recognizably descended from products in 1996. Spreadsheets, similarly. CRMs, project trackers, even most mobile apps — they evolve, but the rate of replacement is measured in years.

AI tools, as a category, are not like that.

The half-life of an AI tool in 2026 is, by my unscientific count, about ten months. I've seen tools — well-designed, well-funded, with real users — go from "hot" to "abandoned" in less than a year. The ones I was using nine months ago, half of them I no longer touch. The ones I'm using today, statistically, half of them won't survive the next nine months either.

This is not a scolding observation. It's a structural one. The category is unstable for reasons that have nothing to do with the makers being lazy. Underlying models change. APIs reprice or sunset. New entrants leapfrog old ones because the leap doesn't require building infrastructure — it requires writing a better prompt. The competitive moat that protected, say, a CRM company in 2010 (years of data integrations, accumulated workflow lock-in) doesn't exist for a tool whose entire value-add lives between user input and a model API call.

What I've started doing, in light of this, is paying close attention to which AI tools have the *traits* of long half-life rather than betting on individual tools surviving.

Tools with content survive longer than tools without. A pure UI-on-top-of-API tool is fragile; the moment the underlying model gets a better default UI from its provider, the wrapper has nothing left. A tool that ships with a curated library of patterns, examples, or knowledge — content the maker has hand-crafted — has a moat that doesn't evaporate when models shift. PromptCraft tries to be in this second category.

Tools with a clear point of view survive longer than tools without. This sounds soft but it's structural: a tool with a clear opinion about how a specific user should use a specific model produces consistent results, builds user habit, and has an identity that survives a model swap. A tool that's "AI for everything" is mostly competing with the model itself, and the model wins.

Tools that are integrations of multiple AI capabilities survive longer than tools that are wrappers on one. If a tool combines image generation + text composition + structured output into a single workflow that solves a real task, the workflow is the moat, not any individual model call. Each component can be swapped without the user noticing.

Tools whose makers actually use them survive longer than tools whose makers don't. This is the most predictive trait, and the easiest to detect. You can usually tell within two minutes of using a tool whether the maker uses it themselves. Tools where the maker uses them have weird, specific affordances — the kind of thing only a real user would think to add. Tools where they don't have a smooth, generic, slightly hollow feel.

The implication for me, as a builder, is to bias hard toward all four traits. Build tools that ship with content. Build tools with a point of view. Build tools that are workflow-shaped, not model-wrapper-shaped. And always, always be a heavy user of the thing yourself.

The implication for me, as a user, is to be more comfortable with tool churn than I used to be. The tool I love today probably won't be my favorite in twelve months. That's not a betrayal; it's the category. The right move isn't to lock myself into one tool, but to keep my data and workflows portable enough that swapping is cheap.

I think this period of fast-half-life will eventually moderate. As models converge on a stable plateau of capability, the differentiation will move to interface and curation, both of which compound over time. Tools that survive this turbulent period — the ones that are still around in 2028 — will probably do unusually well, because the trust they accumulated when everything was changing fast will pay off when things slow down.

Until then: design for the churn, expect the churn, and don't take it personally when your favorite tool sunsets. It's not a referendum on the maker. It's a consequence of building anything in a category where the ground keeps moving.