The analogies for AI are usually too large.
Electricity. The printing press. The internet. The industrial revolution.
They are useful if the point is scale. They are less useful if the question is adoption. At that altitude, the mechanism disappears. Everything becomes inevitable in retrospect. The messy part, which is the part that matters, gets compressed into a sentence: then the technology diffused.
Excel is a better analogy because it is recent enough to remember.
Most organizations did not adopt spreadsheets because a central transformation office designed the perfect operating model. They adopted them because people found a way to get work done faster than the official system allowed.
The first spreadsheet was often not a sanctioned system. It was a workaround. Someone in finance built a model because the report was late. Someone in operations made a tracker because the system of record could not answer the question being asked. Someone on a project team copied last month's workbook and changed the assumptions because that was faster than requesting a new dashboard.
At first, this looked like local productivity. Then it became dependency.
The person with the workbook became the person who understood the business. The spreadsheet became the meeting. The tab names became the operating model. The color coding became governance, even when nobody called it that. A formula written by one analyst became a decision input for people several layers above them.
That is the part worth studying.
Not "spreadsheets made knowledge workers more productive." Of course they did. The more interesting story is how a tool moved from personal productivity to organizational infrastructure without ever fully announcing the transition.
AI is following the same path.
The first useful adoption is not a grand enterprise workflow. It is a person with a recurring problem and enough agency to try a tool before asking permission. They use it to draft, compare, summarize, debug, structure, classify, or rehearse. It saves them time. Then it saves their manager time. Then the output becomes expected.
Only later does the organization discover that some of its work has quietly changed.
This is why adoption numbers are less interesting than dependency patterns. A survey can tell you how many people have used an AI tool. It cannot tell you whether the tool has become part of how decisions are made. It cannot tell you which unofficial workflows are now carrying real organizational load. It cannot tell you whether the person using the tool has become more capable, more dependent, or simply faster at producing the same judgment.
Excel created local experts. AI will too.
In many organizations, the first AI power users will not be the most senior people or the most technical people. They will be the people closest to a recurring bottleneck. The person who has to reconcile messy inputs. The person who prepares the weekly brief. The person who turns meeting fragments into decisions. The person who knows which exception actually matters.
They will not always have the job title that matches their influence.
This is another reason the Excel analogy works. Spreadsheet expertise did not map cleanly onto hierarchy. The most important model in the room was sometimes maintained by someone whose formal authority was limited. Everyone knew not to touch the workbook. Everyone also knew the meeting depended on it.
AI will create the same awkward dependence, but faster.
The unofficial prompt, workflow, saved instruction, retrieval folder, or agent configuration may become the thing the team quietly relies on. It may live in one person's account. It may not be documented. It may not be reviewed. It may be excellent. It may be dangerously stale. Either way, the work will begin to route through it.
That is not a future governance problem. That is the adoption curve.
The standard enterprise response will be to formalize too late. First there will be experiments. Then guidelines. Then approved tools. Then centers of excellence. Then a belated effort to find the shadow systems that already matter. This is roughly what happened with spreadsheets, except spreadsheets were slower and less capable of sounding authoritative when wrong.
The lesson is not that organizations should clamp down on every unofficial use. That would miss the point. Local workarounds are often where the organization discovers what the official system cannot do. If every spreadsheet had needed approval before it existed, a lot of useful business knowledge would never have been written down.
The lesson is that unofficial adoption should be observed earlier.
Where are people already using AI because the official workflow is too slow? Which of those uses are personal drafts, and which are becoming shared infrastructure? Which outputs are being copied into systems of record? Which prompts are being reused by a team? Which users have become local experts? Which old work has actually stopped, and which work has merely gained another step?
Those questions matter more than whether everyone has completed training.
Training is the comfortable part of adoption. It lets the organization pretend the problem is capability. Sometimes it is. More often, the problem is that nobody has mapped the new tool onto authority, review, memory, and retirement of old work.
Excel did not just teach people formulas. It changed who could create a model, who could challenge a number, who could run a scenario, and who had to wait for a system team to produce a report. It redistributed a small amount of analytical agency across the organization.
AI does the same thing with a broader class of work.
It lets more people create first drafts, first analyses, first classifications, first plans, first explanations. That redistribution is useful. It is also destabilizing. The first draft carries assumptions. The first classification becomes a category. The first plan narrows the search space. The first explanation can become the story everyone repeats.
If the organization treats that as mere productivity, it will miss the governance shift.
The durable adoption question is not: when will everyone use AI?
Everyone did not need to become a spreadsheet expert for spreadsheets to change work. Enough people needed to become good enough, in the right places, for the organization to start depending on the outputs.
The better question is: where will the organization become dependent before it becomes aware?
That is where the real adoption curve begins.