Gates wrote the transition plan pitch for regulators. Employers still owe theirs.

Read the two proposals as line items rather than as ideas. A levy on tokens and robots ^[4] is a tax on the exact inputs a company buys when it automates, and Gates is explicit that its purpose is partly to slow the substitution ^[5] and partly to fund retraining and the safety net ^[6]. A human-reserved category ^[2] is a floor under headcount in named sectors, set by someone other than the operator. Neither costs anything to a firm that never planned to automate. Both cost a great deal to a firm whose 2027 plan assumes the current tax treatment, in which a hire carries payroll tax and a robot is written off as a business expense ^[7].

That asymmetry is the mechanism, and Gates says the tax code is currently pointing employers at the machine ^[7]. Which means the automation decisions being approved this quarter are being priced against a code that at least one prominent voice is publicly lobbying to change. A capital plan built on immediate expensing is a plan with an unhedged policy assumption in it.

The timing does not favour deliberation. Gates puts the change in law, customer service, medicine, software and manufacturing at roughly a decade rather than generations ^[10], and dexterous robots reaching construction and hospitality tasks by the end of the decade ^[3]. The trigger he identifies for employers is narrower than general capability: the point at which output is close to error-free, because a human checker stops paying for itself ^[11]. That is a procurement threshold, not a research milestone, and it is the sort of thing a vendor will tell a buyer has already been crossed.

There is also a distributional detail worth pausing on. Gates points to early evidence that young workers in AI-exposed roles have already seen employment fall while older colleagues have not ^[18]. For an employer, that reads as a hiring freeze at the entry level rather than a redundancy programme, which is precisely the form of workforce reduction that generates no severance line, no notification, and no internal record that anyone has to explain later. It is also the form least visible to the department that would eventually be asked to design retraining.

The Guardian notes this is Gates's first lengthy piece on AI in three years ^[17], and that his staff are trying to arrange a November meeting with Xi Jinping on AI risk ^[13]. He concedes the institutional build will take years ^[14]. Employers do not get those years. The gap between when a token tax could plausibly be drafted and when the cheap-robot timeline bites is where a redeployment budget either exists or does not. Gates's own line on sequencing is the sharpest thing in the essay: waiting until people are already displaced or underemployed will be too late ^[15]. He wrote it at governments. It reads the same at a compensation committee.