Deep Dive · Xiaohu Reads

Turning Every Employee into an "AI Power User" Is a Dead End. Making AI Invisible in the Background Is the Real Answer.

Across dozens of large companies, the post-rollout distribution is always the same: 5-10% use AI daily, 20% use it poorly, and 70% never touch it. Training won't change that shape. What will is making AI something you never have to summon.
The 1-Minute Read
  • Across dozens of large companies, the distribution is always a barbell: a tiny group of heavy users and a huge mass of people who never touch it. Dashboards still record this as "adoption success."
  • One executive had just signed an 8-figure annual license. 10% of the seats burned 90% of the tokens.
  • His conclusion: training only surfaces the people who were already ahead. It doesn't transform everyone else. What works is pushing AI into the background and asking no one to change how they work.
⚑The author, Vasuman, is CEO of Varick Agents, an enterprise AI consulting firm. The second half of this piece—building agents into existing enterprise systems—is exactly the service his company sells, and the original post ends with a pitch. His first-hand observations are genuinely valuable, but the conclusions align with his business interests. Keep that in mind as you read. The sources and sample sizes for the two public surveys cited here were verified and filled in by us.
Opening

You're Already in the Top 1%. Your Problem Is Behind You.

Vasuman's company, Varick Agents, helps large enterprises (over $500M in revenue) embed AI agents into finance, sales, procurement, operations, and HR. Across dozens of large companies, the post-rollout distribution looks the same, and it all points to one blunt conclusion: "AI adoption rate" is a fake metric.

If you're reading this, you're almost certainly in the top 1% of AI users. You might not feel that way—X is full of people running 20 terminal windows at once and an agentic knowledge base that rewrites itself after every session. But you can't catch that crowd, and you don't need to. The gap ahead of you is far smaller than the chasm behind you.

Behind you is the median employee at a large company. Over the past two years, they've probably opened ChatGPT a total of four times, still on the 3.5-turbo model, and concluded that "AI is dumb"—right as GPT-5.6 Sol Ultra just proved a graph theory conjecture that had stood for 50 years. Every article about AI strategy assumes this median employee will eventually catch up. They won't.

Frontier 20 terminals open You Daily user A short gap Median employee Opened it 4 times in 2 years That's a chasm—and every model release widens it
The coordinates for this entire piece. Illustration by Xiaohu Reads.
Original figure: The chasm
FIGURE 1 - The chasm. Source: Vasuman / Varick Agents
Observation

The Tool Went In. Team Speed Didn't Move.

Here's what that chasm looks like inside a real company.

An operations lead at a non-tech company manages a few thousand people. Before AI, team speed was fine—nothing to complain about. Then the company rolled out Claude Cowork.

Team speed stayed the same. The lead had one question: why?

Because usage inside the organization splits the same way every time, whether the team is 50 people or 5,000. It's always a barbell:

5-10%
Heavy users: daily, writing skill files, connecting Outlook connectors
20%
Use it a few times a day, but poorly—capturing only a sliver of what the frontier users get
70%
Never touch it
What "barbell" means here

A distribution split into two extremes with hardly anyone in between: a small group using it very deeply, and a large mass not using it at all. Like a barbell at the gym—heavy plates on both ends, and a bar in the middle so thin it barely registers.

The 5-10% heavy users are the ones who pushed the company to buy Cowork in the first place. They were already tinkering with AI on their own time and had seen what it could do.

On the ledger, this distribution reads as one sentence: Nothing got faster, but you definitely burned a few million dollars.

Key point

The organization's dashboard records this rollout as "adoption success." But in the lead's eyes, nothing got faster. Both of those are simultaneously true.

Original figure: The barbell distribution of an operations org
Distribution of an operations org. The bottom-right line reads "Observed speedup: none." Source: Vasuman / Varick Agents
A note on the numbers: The 5-10% / 20% / 70% split comes from his own client engagements, not from public survey data.
Mechanism

Using AI and Using It Well Are Two Different Skills

But why does this happen? Even among the people who do use AI, the skill gap is terrifying:

AI users who don't know what they're doing are worse off with AI than without it. Vasuman