In my recent series of posts, beginning with New Myths for Old, I examined some impacts of AI on the cybersecurity workforce and profession. That examination included the claim that LLMs make software developers, security analysts, and incident responders optional, as well as the layoffs conducted under that banner. I called those cuts the patch-instead-of-fix mindset extended to staffing. Patches defer costs rather than remove them, and the bill for this one is now arriving, itemized.

The first line item is the expected return that is missing. PwC’s 2026 Global CEO Survey found that 56% of chief executives report AI has neither increased revenue nor reduced costs in the past twelve months. Bain’s June survey of enterprise adopters found that the double-digit cost reductions firms expected mostly failed to materialize: among firms that measured outcomes, the largest share saw improvements of 10% or less. Both echo the controversial MIT report from last August, which found that roughly 95% of enterprise generative-AI pilots produced no measurable profit-and-loss impact. Gartner projects that more than 40% of agentic AI projects (deployments in which the system selects and executes actions with little human authorization) will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls. The same analysis estimates that only 130 of the thousands of vendors selling agentic AI offer the genuine article. The rest practice what Gartner labels agent washing.

The expense side of the ledger tells a similar story: over the past year, large employers have swung from mandating AI use to rationing it. Accenture originally told senior employees they risked losing out on promotions if they did not use AI. By late June, its agentic-AI strategy lead was explaining, in audio leaked to 404 Media, that the firm was trying to stop employees from draining its token reserves on tasks such as converting PDFs into slides: Spend is becoming very unpredictable; and leadership … are still asking the question of whether they’re getting value. Meta employees consumed 73.7 trillion tokens in roughly thirty days, a pace encouraged by internal leaderboards; with internal AI costs heading toward billions of dollars for 2026, the company is now imposing central caps. Amazon deleted an employee-built internal leaderboard that ranked developers by token consumption. Uber exhausted its 2026 AI coding budget in four months and now caps employees at $1,500 per month per tool. Enterprise-AI vendors describe clients whose annual AI budgets are exhausted in one or two months, and CNBC’s summary of the resulting boardroom choice could be this post’s epigraph: tokens or humans. The press has named both phases: tokenmaxxing on the way up, token minimizing on the way down.

An organization that orders its employees to use a tool and, months later, to ration it, has not identified any value in the interval between the two orders. This should have surprised no one. Campbell’s Law is well known: the more a quantitative indicator is used for decision-making, the more it corrupts the process it is intended to monitor. (Goodhart’s Law is its pithier cousin.) Usage became the target, and usage is what employees produced.

The second line item in the invoice is the rehiring. Careerminds surveyed 600 HR professionals who had conducted layoffs in the prior year. Two-thirds of employers that cut jobs for AI are already rehiring for the eliminated roles, more than half within six months of the layoff. Among those rehiring, about 31% spent more bringing the roles back than the cuts had saved, and another 42% broke even; about one in four came out ahead. Only 8.4% said the restructuring delivered what was promised and would repeat it unchanged. Orgvue’s survey of senior decision-makers found 55% of firms that made AI-motivated redundancies admitting wrong decisions. A follow-up Orgvue study found something more revealing: 23% of companies that made layoffs relied on general assumptions about what AI could do rather than analysis of what the eliminated employees did. That is consistent with the Oxford Economics observation cited in New Myths for Old: the AI label often dressed up ordinary cost-cutting. Robert Half reports that 32% of U.S. hiring managers eliminated a role primarily because of AI and later rehired for the same or a similar position.

Some particular cases trace the arc:

• Klarna, which had boasted that its chatbot did the work of 700 customer service agents, resumed hiring people in May 2025. The CEO’s diagnosis: cost had been a too predominant evaluation factor, and the result was lower quality.
• The Commonwealth Bank of Australia declared 45 call-center roles redundant after deploying a voice bot. Call volumes rose instead of falling, team leaders were pulled onto the phones, and after the union took the bank to a tribunal, CBA reinstated the roles, called the decision an error, and apologized.
• IBM automated much of its HR function; the system handled about 94% of routine requests and couldn’t manage the remainder, which included ethical dilemmas. The company now plans to triple its U.S. entry-level hiring in 2026, having recognized that automating the routine work had also removed the rungs by which juniors become seniors. Its chief human resources officer explained: If we don’t continue to invest in entry-level hires, what happens in three-five years? There’s no pipeline; the well simply dries up.
• Ford brought in roughly 350 veteran engineers, some of them returning former employees, after its automated design and quality systems missed flaws that experienced people would have caught. Artificial intelligence is a fantastic tool, but it’s only as good as the information you use to train it, Ford’s vice president of vehicle hardware engineering told Bloomberg. The veterans had departed before their knowledge could inform the new systems, so the tools amplified weak inputs instead of correcting them. With the senior engineers back, Ford topped J.D. Power’s initial-quality ranking among mainstream brands for the first time in sixteen years, and its CEO says declining warranty and recall costs amount to potential savings of hundreds and hundreds of millions of dollars.

Ford is perhaps the most instructive case: it did not remove the AI. It kept the tools and restored the people who could tell when the tools were wrong. The judgment that distinguishes a real defect from noise — the argument this blog series has made from the start — turns out to be visible on a ledger. Handling the 6% of requests IBM’s system could not manage, the ethical dilemmas, calls on the skills I argued in More Than the Code that a full education must cultivate. Security organizations should read their exposure in these examples: the analyst who can tell a real intrusion from an alert storm carries the same kind of judgment Ford had to buy back, and with roughly 514,000 U.S. cybersecurity openings unfilled, it will not come back cheaply.

The third line item is a new set of failure modes. People err, but organizations use controls (review, sign-off, audit, professional discipline) refined over centuries (some over millennia), built around how people err and around an accountable party. Replacing the people with AI removes both. Consider:

• Fabrication delivered with confidence. Damien Charlotin’s worldwide tally of court decisions involving AI-fabricated legal filings stood at more than 1,200 as of early April 2026, about 800 of them in U.S. courts, and NPR reports the rate is still rising. In one case, a federal court in Oregon ordered a lawyer to pay $109,700 in sanctions and costs. Deloitte Australia refunded part of an AU$440,000 fee after a report it delivered to a government department was found to contain invented academic references and a fabricated quote from a federal court judgment. A paralegal who invented case law would be fired and possibly sanctioned; an AI model that does the same is redeployed with a revised prompt.
• Action authority without judgment. Replit’s coding agent deleted a production database during an explicit code freeze, then informed its user that rollback was impossible. (It was not — the user ran the rollback himself, and it worked.) No organization would give a first-week intern unsupervised write access to production plus the authority to certify the cleanup afterward. Agentic deployments are granted both by default.
• Failure at the interface. McDonald’s ended its AI drive-thru partnership after orders such as bacon on ice cream went viral, and Taco Bell is rethinking the voice AI it deployed at more than 500 drive-throughs after a customer ordered 18,000 waters.
• Degradation from within. Researchers at BetterUp Labs and Stanford documented what they call workslop: AI-generated content polished enough to pass as work but unable to advance the task. In their survey, 41% of workers had received it, each instance costing nearly two hours of rework, an estimated $186 per employee per month. A tool that produces hollow work at machine speed, under a colleague’s name, is a failure mode for which no control yet exists.

None of this is a verdict about the technology’s utility. AI tools can contribute value, as I wrote in After the Buggy Whip. The verdict is against the replacement myth: the claim that the people could be subtracted and the output would stay the same or improve. The organizations that acted on that myth are paying twice — once for the tools, and once to rehire, sometimes at a premium, the judgment they discarded. The organizations reporting favorable returns appear to be the ones that treated AI as an augmentation for experienced people.

The next round of hype and myth is already underway. The vendors whose products anchored the last cycle are now selling agentic AI as the remedy for its shortcomings: autonomous systems with the access and confidence of a senior partner, but with the skills of that first-week intern. Gartner’s cancellation forecast, cited above, covers exactly these projects.

Conclusion: The second invoice is in preparation.

(A few portions of this text were drafted and structured with the assistance of Anthropic Claude Fable 5; the ideas, arguments, and final editorial decisions are the author’s.)