vas on X: "https://t.co/haApURmzAN" / X<br>Post

Log inSign up

Post

vas

@vasuman

AI Adoption is a Myth<br>You’re already in the top 1% of AI users. It might not feel like it because you see nerds on X spinning up agent fleets on a whim, but trust me, you’re in the top 1%. You’re likely using a model every day, you have opinions about how best to use it, and you maybe even tinker with agents or have set up Hermes on your own machine.<br>Yes, there’s a gap between you and the folks on the frontier at the hottest AI startups. The crazy kids running 20 terminals simultaneously, or a knowledge base that rewrites itself after every run with evals that self-optimize constantly.<br>The bad news is you’re never catching up to those people. The good news is that you don’t need to. That gap, the gap in front of you, is far smaller than the gap behind you.<br>Between you and the median employee at your company is a chasm. They have maybe opened ChatGPT 4 times in 2 years. They used 3.5-turbo and decided AI was stupid, not realizing that GPT 5.6 Sol is solving math problems that have been unsolved for decades. Every AI strategy article or post assumes that this person, this median employee, is going to eventually catch up.<br>I’m telling you, they won’t.<br>For context, I’m the CEO of @varickagents. We work with the largest companies on the planet to implement AI agents and strategy across their organizations, so we know a thing or two about the reality of AI at enterprise.<br>The best AI tool in the world did not make them faster<br>I recently spoke to a leader of a non-technical enterprise’s operations org (thousands of people). Without AI, he said, his team was moving at a decent rate. No major complaints there.<br>Then they rolled out Claude Cowork, and what do you know? The team kept moving at the same rate. He asked me why.<br>I’ve seen the same split across every organization. Doesn’t even matter if it’s 50 people or 5000, it’s the same split every time: a barbell. 5-10% of the org were power users. They were using Cowork every day, leveraging skill files, connectors to Outlook, and more. These were the evangelists, the ones who pushed for Claude Cowork in the first place because they tinker with AI outside the job, and see first-hand how impactful it can be.<br>Of the other 90%, 20% would use it a couple of times a day, pretty poorly. They get some value out of it, but a fraction of what the frontier folk get. And the remaining 70%, they didn’t use it at all.<br>The dashboards in his org would indicate that the rollout counts as adoption. But to him, nothing got faster. Both of these were true simultaneously.<br>Using it and using it well are different skills<br>Even among those who do use AI, the skill gap is enormous. Users of AI that have no idea what they’re doing are actually far worse off after AI than they were before. Allow me to illustrate:<br>It is trivial to install Claude, point it at a repository, send a 4 word prompt, and watch something happen (could be right, could be wrong).<br>Using it well is a different skill. Knowing when to clear context. Knowing that the thing you just did twice should become a skill file the model reads every time, instead of a prompt you retype. Knowing which 15% of the automation project requires a model for judgement vs which 85% just requires deterministic code. And most importantly: reading a diff properly before you accept it.<br>If you give the same ticket to two engineers, this divergence becomes apparent in the first 5 minutes. The first one will simply paste the text from Jira into Claude and hit submit. The fix will touch six files, and this engineer will skim through this, see the tests still pass, and then merge the PR. Three weeks later, when production starts acting up, someone will realize this PR changed a config value for no reason.<br>The second engineer starts the same, they paste the text from Jira, but then they’ll flag where in the repository the code lives, what folders and files to touch vs to leave alone, and they already have skill files that ensure every PR submitted by Claude is minimalist instead of verbose, and is adequately tested before submission. They then read the diff, catch a stray change, fix it with a quick prompt, and then merge a PR half the size of the first engineer's.<br>At least half of any organization is never going to get to that second version. Using AI well is a craft. It takes enormous iteration to get really good at using AI, and the inertia of someone who isn't on AI is insurmountable for many. Turning a slop-cannon into a refined power-user is just as hard.<br>A perfect rollout still produces a barbell<br>The obvious objection is that “well these are companies doing it badly. I would do it much better.” No. Even if you roll it out perfectly, you'll get a barbell no matter what.<br>Another executive we spoke with had just finished putting enterprise licenses in front of their several-thousand-person organization. A commitment to spend at least $10M throughout the year. Roughly 10%...