In urban planning, there is a concept called the desire path. It is the trail worn into the grass where people actually walk — not where the sidewalk was laid, but where the foot traffic goes when nobody is watching. Landscapers have two options when they find one: fence it off, or pave it.
Most organisations are currently fencing off their desire paths.
Over eighty percent of workers now use AI tools that their employer has not approved. They are not doing this out of rebellion. They are doing it because the tools work, the approved alternatives either do not exist or take six months to provision, and the quarterly targets did not get lighter while the procurement team deliberated. The workers found a shorter route. They are already walking it.
The standard response is a policy. Ban the tools, restrict the access, send the memo. This is the fence. It has the same effect on behaviour as fencing off a desire path in a park: people walk around it, or they step over it, or they wear a new path three metres to the left. The trail does not disappear. It just becomes harder to see.
The map nobody drew
The deeper problem is not that employees are using unapproved tools. The problem is that their usage patterns contain the most honest map of where AI actually helps — and almost nobody is reading it.
When you interview teams about how they use AI informally, you find something striking. The use cases are not the ones on the strategic roadmap. They are not the ambitious transformation projects the steering committee approved. They are mundane. Summarising a forty-page contract so somebody can find the clause they need. Drafting the first version of a status report nobody wants to write. Cleaning up a folder structure that has been accumulating since 2019.
These are not glamorous use cases. They are the ones that people actually reach for when nobody is telling them what to reach for.
Roughly seven in ten workers say their employer lacks clear AI guidelines or policies. So they make their own decisions. They find the tool that fits the task in front of them. They piece together their own toolkits from whatever works — not because they are indecisive, but because no single approved option covers the ground they are actually standing on.
What the path reveals
Every shadow AI interaction is a signal. It says: this task is painful enough that someone sought help without being told to, and trusted an external tool more than the internal process. That is not a security incident waiting to happen. It is a requirements document writing itself.
The organisations that struggle most with AI adoption are the ones that start from the strategic roadmap and push downward. They identify the high-value transformation use case, build the business case, run the pilot, measure the ROI, and wonder why adoption stalls at seven percent. Fewer than forty percent of organisations have managed to scale AI past the pilot stage. The pilots work. The scaling does not.
The ones that move faster start from the desire path and work upward. They watch where their people are already walking. They ask: what tasks are people solving with unapproved tools, and what would it take to do that same thing safely, inside our systems, with proper data handling? The use case is already validated by the most reliable signal in the enterprise — someone did it voluntarily.
Paving, not planning
This is not an argument against governance. Quite the opposite. PwC calls bottom-up adoption without governance an understandable mistake — one that produces impressive adoption numbers but seldom leads to meaningful transformation. The desire path, left unpaved, is a liability — data leaking into public models, inconsistent outputs, no audit trail, no way to tell what your organisation actually knows versus what it borrowed from a chatbot.
But the answer is not to replace the desire path with a five-year digital transformation roadmap. The answer is to pave the path that already exists. Start with the use case your people have already chosen. Give it guardrails. Make the safe version easier than the shadow version. Then move to the next path.
The teams that adopt AI fastest are not the ones with the most ambitious strategy. They are the ones whose formal tools are easier to use than the informal ones. That is the bar. If your approved AI tool requires a training course, a ticket, and a three-week onboarding, and the unapproved one requires a browser tab, you have already lost the adoption race.
The question that changes the conversation
Most AI governance conversations start with "what should we build?" The better question is: "what are people already doing?"
The desire path is not a deviation from the plan. It is the plan, told honestly. The organisations that read it will move faster, adopt more durably, and waste less time building tools for problems their people never actually had.
The ones that fence it off will keep wondering why nobody walks on the sidewalk.