I am not a policy expert.
Editorial update, October 11, 2026: The original title was "Governments Are Preparing for the Wrong AI Timeline." This remains my February policy concern. The ILO's May 2025 occupational-exposure methodology estimates potential task exposure, not realized job losses; it describes transformation as more likely than complete replacement for most occupations. This essay supplies neither a representative sample of government positions nor a causal employment series establishing my timeline.
I build businesses without venture capital, and I have spent two years watching AI change my own work. That is a narrow vantage point, not a measurement of the whole economy.
So take this as a field report, not a lecture.
The thing I keep running into is a gap between the changes I see in my work and the timelines I hear discussed. I worry that policy could move too slowly. I cannot establish what most governments believe from those conversations.
What the room hears, and what builders already know
The reassuring version I have heard in these conversations goes like this.
AI will create more jobs than it destroys. Reskilling will bridge the gap. We have ten to fifteen years to prepare. Regulation will slow it down enough.
I do not think those assumptions are safe planning defaults. That is a concern about uncertainty, not proof that every one is wrong.
What I would add from my own building experience is less comfortable.
AI can change tasks and staffing needs, but a task becoming automatable does not establish that a role has disappeared. I worry that capability could advance faster than reskilling or policy. The timing is uncertain, and domestic compute is one way to reduce a dependency, not the only ingredient of an AI strategy.
The displacement is already here, quietly
The mistake is waiting for a headline.
My concern is that changes in hiring and job design can be harder to see than announced layoffs. That makes them worth measuring; it does not establish how many unannounced AI-driven cuts have happened.
I do not have data here that isolates AI as the cause of those decisions.
One scenario I worry about is wider restructuring and lower wage income in affected work. It could matter for tax receipts even while some firms produce more. That is a scenario to test against data, not an observed sequence established by this essay.
The questions are whether that scenario occurs, which work changes, and on what timescale.
The economy stops measuring the right thing
Here is the part I think gets missed entirely.
Imagine ten people doing work that previously needed a hundred. Output could hold steady while paid hours fall. That illustrates why output and employment should be measured separately; it is not a measured tenfold staffing change or a prediction that revenue and wages must move in one direction.
You cannot manage what you are not measuring. A lot of the current dashboards will read green right up until they do not.
Compute is the resource, and almost nobody owns it
The other thing.
I worry about depending on a small number of infrastructure and model suppliers. Domestic capacity can offer another option, while hardware, software and services can still introduce outside dependencies.
For a nation, this has an energy-policy dimension as well as a technology-policy one. Relying on suppliers is a strategic choice with trade-offs, rather than proof that no AI strategy exists.
Sovereign compute can reduce some dependencies. It does not, by itself, deliver a competitive model, adoption or economic independence.
What I would actually tell a policymaker
Do not budget for a decade. Budget for the next few years, and assume the curve keeps bending.
Watch roles, not just unemployment. The early signal is in what stops being hired, not in what shows up as a layoff.
Treat compute as strategic infrastructure, the way you would treat power or ports.
And include people who are actually building, as well as workers and researchers studying the effects. Builders can be wrong too. Their observations are inputs to test, not an eighteen-month forecasting advantage I have measured.
I am not selling anything here. I just think the people making these decisions deserve better inputs than the ones they are getting.