AI Adoption is a Myth

If you’re reading an article about AI on a weekday, 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. 

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. 

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. 

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.

I’m telling you, they won’t.

For context, I’m the CEO of Varick Agents. 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. 

FIGURE 1 - The chasm

The best AI tool in the world did not make them faster

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. 

Then they rolled out Claude Cowork, and what do you know? The team kept moving at the same rate. He asked me why.

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. 

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. 

FIGURE 2 - A tenth of them are good

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. 

Using it and using it well are different skills

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:

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).

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.

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. 

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.

FIGURE 3 - Same tool, same seat, different outcome

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.

A perfect rollout still produces a barbell

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. 

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% of people burn 90% of the tokens. If you think about it, this is basically an extension of Figure 2.

To put that into perspective: if your entire org WAS great with AI, your spend would not be $10M, it would be $100M. All of a sudden your best-case scenario becomes your worst. 

FIGURE 4 - 10% of your seats burn 90% of the tokens

Adoption is binary, skill is a spectrum

McKinsey says 78% of organizations use AI in at least one business function, yet 80%+ of those same organizations report no significant EBIT impact. MIT NANDA's GenAI Divide report says 5% of integrated pilots produce over $1M in value, and the other 95% have nothing to show for it. 

If you track adoption as a metric, you’re very likely asking (at most) an abstraction of a yes-or-no question. “Did this person log in this month?” “Did this person log in this year?” “Did this person send at least 5 prompts per day?” Etcetera. If you want to differentiate between "has never opened it" and "pastes emails in for reformatting" and "has three agents in production touching the general ledger," you need a better system of measurement. This is not the case for 99.9% of enterprise AI adoption trackers. This is stupid, for obvious reasons (see: barbell, above).

FIGURE 5 - What the adoption number hides

In all fairness, the number represents something real, and people are not lying about adoption. This is far easier to measure than the real thing you care about: how skillful is each person at using AI, and what is their ROI per token used? 

Hence why AI adoption is a myth. If you actually tracked this metric, you would realize there is no hope for the vast majority of people to catch up to AI nativity in the enterprise. Does that mean they have no value? Of course not. It just means that your implementation and rollout plan needs to be adjusted accordingly. 

Every incentive keeps the chasm open

Every AI vendor’s roadmap pitches increased capability for agents, but says nothing about the teams’ ability to use it effectively. So the only people who use it effectively are frontier users, who are the loudest, so they get their say. They’re also the most flattering to build for, so it’s a clear win-win. Every feature release raises the skill floor to use the product well, and this only widens the chasm between the frontier and the majority in an organization. 

I keep seeing enterprises seduced by the idea of ‘train your employees how to prompt!’. This is 10% of the game. The other 90% is “train your employees to understand which workflows should never touch a model vs which workflows should be entirely automated”. The reason you can’t sell this as easily as the prompt-training is because the answer is different at every single company, and it takes a few weeks on-site to work out. 

And the AI ninjas inside your company have no incentive to fix it either. Their leverage IS the chasm. They’re doing their work 70% faster, then get to coast. Or they do the work of 3 people. But the moment everyone’s on that wave, they lose their edge. Why would they fix it without additional incentive?

So what’s the solution:

Well, you have to train anyway. You won’t even know who’s in the top slice vs who’s not without some training. There are definitely some people at your company who are interested in AI but lack the time and leverage to go deep. You should train to expose everyone’s ability and comfort level. 

For your top slice, your power users, give them somewhere to publish, like a shared database, so that every skill they make gets posted, ranked, and installed by others. This is the only mechanism I’ve seen that turns one person’s breakthrough into something sharable. Here, ranking is the incentive: as power users will trade their edge for status in an org. The gap still remains though: you can blast this to everyone and still 50% of people will never use a skill. 

For everybody else, the work has to get done without them changing how they work, which means the AI goes into the background. Why do you need AI to be prompted by a human anyways? Just figure out what the most repetitive processes are, build agents in your existing systems of record that employees are used to (Salesforce, NetSuite, Dynamics, etc), and only pull in humans where you need a second glance. 

For example: you have dozens of AP analysts who move invoices all day. 90% of this is automatable. You can’t expect them to spin up agents or prompt their way through it reliably without breaking in production. So build the agents that run every single day, and now these analysts are just approving, rejecting, or editing the work the agents do. 

One thing you need to remember: people are not in the market for a tool that helps them get the work done. They just want the work done. 

FIGURE 6 - Two paths, sized to who is actually on them

So stop reporting ‘adoption’ to your board. Report what share of the work today is manual, vs hybrid, vs fully automated. This is the only reporting and conversation that matters. 

TLDR

• You’re likely in the top 1% and while there’s still a gap in front of you, there’s a chasm behind you. The median employee at your company is miles behind you, and the gap actually increases with every release. 

• The average Claude Cowork or similar rollout goes like this: 5-10% become power users, 20% use it poorly, 70% hardly use it, and nothing gets faster but you definitely burn millions of dollars. 

• Using a tool and using it well are separate skills, and at least half of any organization never uses it well. 

• A great rollout does not fix this. You can put AI in front of thousands, but 10% of your org will still burn 90% of your tokens. 

• Adoption metrics collapse entire spectrums of distribution into a binary yes and no, which is how you get figures like 78% adoption while 80%+ report no significant impact.

• You need to train the top slice and publish what they build. For everyone else, put the AI in the background of the systems they already work in.

Get help doing this:

This is what we do at Varick. We audit where the work actually happens inside an organization, then build agents into your existing systems of record, so the work gets done without asking five thousand people to become good at a craft they did not sign up for.

We work with companies doing over $500M in revenue, across finance, sales, procurement, operations, HR, and more. Check us out at varickagents.com.