Count the AI assistants in your company. Not the licenses your IT team bought. The ones people actually use, including the personal accounts nobody declared.
In most organizations we walk into, the number is roughly equal to the headcount. Everyone has one. Everyone likes it. Adoption, on paper, looks like a success story.
Then ask a second question. In the last month, what did your organization learn from all of that AI use?
The answer is almost always nothing, and it is worth being precise about why.
Fifty assistants, fifty dead ends
Someone in your finance team spent an afternoon getting an assistant to untangle a messy variance report, and by the end of it they had a way of framing the question that works every time. That framing exists in one chat history. Nobody else will ever see it.
Someone in sales worked out how to pull the three objections that actually killed a deal out of six months of call notes. Also one chat history.
Someone in legal found a way to catch clause drift across a contract set. Same story.
Three genuinely valuable pieces of institutional knowledge were created in your company that month, and all three of them are trapped in private conversations that no colleague can search, no manager can review, and no new hire will inherit. When any of those people leave, the knowledge leaves with them, the same way it always has, except now it leaves faster because it was never written down anywhere in the first place.
You did not adopt AI. You created fifty new silos, and unlike the old ones, these do not even have a shared drive somebody could go rummage through.
The tools did exactly what they were sold to do
This is not a failure of the products. It is worth saying plainly, because the reflex is to blame the vendor or the rollout.
A personal AI assistant is built to make one person better at their own tasks, and it is very good at that. Drafting, summarizing, reformatting, explaining, checking. Every one of those is a task, and a task is what one person does.
A workflow is different. A workflow is how the organization does something: it crosses people, it crosses systems, it has handoffs, and it has exceptions that consume most of the time. No personal assistant can improve a workflow, because no personal assistant can see one. It only ever sees the slice in front of the person typing.
So the value from AI lands where the tool is pointed. On individuals. Individually, everyone is faster. Organizationally, nothing has changed, and the numbers at the end of the quarter will say so.
The second thing that happened
There is a subtler cost, and it usually shows up about six months in.
A handful of people turn out to be very good at this. They read the release notes, they have opinions about which model to use for what, and their output improves noticeably. Everyone notices. Those people become the ones you route the hard requests to.
Meanwhile most of the organization is using AI to write emails slightly faster, and a real fraction is quietly not using it at all, either because they tried once and got a confidently wrong answer, or because nobody ever told them what they were allowed to put into it.
You have not uplifted the company. You have created a small AI elite and a large group who now feels behind. That is a management problem long before it is a technology problem, and no platform purchase fixes it.
What to do in the next ninety days
None of the following requires buying anything new. It is worth doing in this order, because each step makes the next one cheaper.
Find out what is actually happening. Not a policy audit. Ask people, without consequences attached, what they use and what they use it for. You will find personal accounts, you will find one team that has quietly built something impressive, and you will find at least one thing being done with company information that should stop this week. All three are useful to know.
Set a floor before you chase a ceiling. The gap between your best AI users and your median ones is where the elite problem lives. Closing it is not glamorous work: it is teaching everyone to write a prompt that gets usable output, to recognize the tasks a model will quietly get wrong, to check something before it reaches a customer, and to know what never goes into a public tool. A company where everyone clears that bar beats a company with five power users and no floor, every time.
Make one team's prompts a shared asset. Pick a single team and a single recurring piece of work. Have them keep the prompts that work in a shared document instead of in their own histories, with a note on what each one is for. This is a small thing and it is the first time AI knowledge in your company will exist outside one person's head. It also tells you quickly whether the culture will carry it.
Write the rules down in one page. What can go into which tool, what has to be disclosed, what has to be reviewed by a human before it leaves the building. One page. People follow rules they can remember.
Then pick one workflow and describe it end to end. Not automate it. Describe it: the trigger, the inputs, the decision points, the exceptions, the handoffs, and where the work sits waiting. Most organizations cannot do this for any of their important processes, which is the real reason the automation never happens. AI can only run a workflow you can describe.
What ninety days will not fix
Being honest about the limit: everything above raises the floor and gets one workflow legible. It does not give your organization a shared memory. Your people will still be starting from scratch on questions somebody already answered, because there is still nowhere for an answer to live where every team and every assistant can reach it.
That part is a real piece of infrastructure and it comes later, after the workflows are worth connecting. Buying it first is how organizations end up with an expensive system full of documents nobody agreed on.
Start with the floor. Then the workflow. The shared memory is the third thing, not the first.
Where to go next: the fastest way to close the gap between your best AI users and everyone else is AI literacy across a broad group, then a manager track for the people leading them. Tell us the group size and we will send a quote and a suggested order.
The platform view: our product team wrote the architecture version of this argument in Beyond Chat Silos.