We Stopped Using the Word "AI" for a Month. Here's What We Learned.

Earlier this year we ran an experiment on ourselves. For 30 days, nobody on the team would use the word "AI": not in writing, not in meetings, not in product specs. Every time we caught ourselves reaching for it, we had to substitute whatever we actually meant.

It was harder than we expected, and more illuminating than we deserved.

The substitutions, by frequency

Here's what we ended up saying instead, ranked by how often:

1. "A model". Usually when we meant a specific LLM doing one specific thing. Way more often than we'd guessed.
2. "A script". For things that were genuinely just regex or a Python loop, which we'd been calling "AI" because the marketing pages did.
3. "A retrieval system". For things doing vector search and pretending to be intelligence.
4. "Whatever's behind that button in the product". For things we genuinely didn't know.
5. "Software". Surprisingly often. A lot of "AI features" were just software with confidence.

Things we stopped saying

"We need an AI strategy" became "we need to figure out which model goes where and why." Considerably less inspirational, considerably more actionable. The strategy meeting got 40 minutes shorter.

"AI will change everything." Couldn't even substitute this one. It just dissolved. Without the word "AI" doing the heavy lifting, the sentence had nothing in it. Which suggests the sentence had nothing in it to begin with.

"Powered by AI." Every time someone tried to write this in a product spec, they had to explain what was actually powering it. Sometimes the answer was "a fine-tuned 7B model running on Together." Sometimes the answer was "five if-statements." The fine-tuned model usually deserved the marketing copy. The if-statements never did.

What changed in our work

Three things, in order of impact:

1. Our specs got better. When you can't say "AI does X," you have to say what specifically does X, with what input, with what fallback when it fails. The act of writing that out caught at least four design problems we would have shipped through.

2. Our estimates got more honest. "We'll add AI to the dashboard" is a quarter of work. "We'll fine-tune a small model on these 4,000 labeled tickets and serve it through a queue with a fallback to a frontier model when confidence is low" is also a quarter of work, but you can argue with it.

3. We noticed how much of "AI" is vibes. Maybe a third of the times someone reached for the word, what they actually meant was "something that sounds futuristic." That's a useful thing to notice in your own thinking.

What we're doing now

We stopped the experiment after 30 days, but we kept most of the substitutions. "AI" sneaks back into casual conversation; we don't fight it. But in writing, in specs, in product decisions, we try to say what we mean.

It turns out a lot of the work in software is figuring out what you actually mean. Words are a load-bearing part of that. A word that does too much, eventually, does nothing.