AI looks cheap when someone else pays the bill

Efficiency is an accounting decision

Technology is called efficient when a chosen output rises relative to a chosen input. That definition is useful and dangerously incomplete. Change the boundary of the ledger and a system can become more efficient without society becoming better off. A hospital can collect more complete documentation, a cloud campus can secure cheaper land, and an agency can reserve enormous purchasing capacity while every offsetting burden is recorded somewhere else.

AI is unusually good at this form of boundary change because it finds, classifies, and acts on information that institutions previously ignored. Sometimes that reveals genuine value. Sometimes it discovers a new route to reimbursement, authority, or extraction. The hard question is not whether the model worked. It is where the gain appeared and where the corresponding bill went.

The $942 million coding dispute is about who defines value

The Blue Cross Blue Shield Association says rising coding intensity added an estimated $942 million to its companies’ spending between 2023 and 2025. Roughly 70 percent, or more than $650 million, was tied to secondary diagnoses that moved hospital stays into higher-reimbursement categories. The association says treatment did not rise with the codes and notes that more than 60 percent of hospital systems use AI coding tools.

That is important evidence, not a causal verdict. Claims data cannot establish that every added diagnosis was clinically unjustified, and an insurer benefits when reimbursement falls. Hospitals can reasonably answer that better software is finally documenting complexity that was always present. The distributional fact remains: AI has made the billing record more valuable even where the payer says the care record looks unchanged. The system may be finding illness, revenue, or both.

A $15 billion hub can still impoverish one household

Google’s planned Visakhapatnam AI hub is presented as a $15 billion investment in computing, energy, fiber, and connectivity. India’s official announcement described a 1-gigawatt project and roughly 600 acres across three sites. The Guardian reports that environmental clearances permit up to 2.51 gigawatts and that residents in Tarluvada say land was taken back, replacement plots and jobs did not arrive, and consultation was inadequate. Google says it engaged deeply, expects no local impact, will use air cooling, and anticipates thousands of jobs.

National growth and local loss can exist in the same project. Compensation may be lawful and still fail to replace a livelihood. Construction employment can be large and temporary. A grid connection can power a globally valuable service while nearby residents bear the heat, transmission, land, and fiscal consequences. The relevant unit is not only dollars invested. It is the household, acre, megawatt, permanent job, and enforceable promise.

Public procurement has its own optical illusion

AIImpactLab’s morning review found three federal AI procurement vehicles with combined maximum ceilings of $880.72 million: a $775.72 million Veterans Affairs ambient-scribe vehicle, a $60 million Air Force logistics vehicle, and a $45 million Air Force legacy-software modernization vehicle. Those numbers show institutional ambition, but they are not money already spent. The two Air Force records identify $7.2989 million in initial or first-task-order obligations; VA task orders were not available in the reviewed record.

A ceiling is permission to buy, not proof of purchase, use, or benefit. Yet maximum values travel farther than task-order details because they make a cleaner headline. The same discipline required in healthcare and infrastructure belongs here: separate capacity from obligation, obligation from deployment, and deployment from independently measured outcome.

The strongest objection is that friction has a cost too

A complete ledger can become a weapon for delay. Hospitals may stop documenting conditions that deserve attention. Communities may lose infrastructure and skilled work. Agencies may be unable to modernize brittle systems because every uncertain spillover becomes a reason to wait. Cost-transfer language can also imply wrongdoing where the real issue is an ordinary, negotiable tradeoff.

That objection is right about the risk of paralysis. It is wrong to treat opacity as the alternative. An impact balance sheet does not decide that every transfer is unacceptable. It makes the transfer contestable before the weaker party discovers it through a premium, a fence, a power constraint, or an incident notice months later.

The impact balance sheet

Every consequential AI deployment should publish four linked entries. First, captured value: whose revenue, time, capacity, or service quality improves, with a measurable baseline. Second, transferred cost: which people, budgets, ecosystems, or institutions absorb new work or risk. Third, funded remedy: what compensation, review, appeal, containment, or restoration is paid for now rather than promised later. Fourth, exit liability: who pays if the model, facility, or contract is withdrawn, fails, or outlives its usefulness.

This is not a universal price tag for human welfare. It is a disclosure discipline. The entries can contain ranges, competing estimates, and explicit uncertainty. What they cannot contain is a claimed saving with no named cost bearer or a promised remedy with no budget and deadline.

• Report contract ceilings and actual obligations separately.
• Publish local jobs as construction, temporary, and permanent roles.
• Test whether a new medical code corresponds to changed care or only changed reimbursement.
• Assign incident-response and decommissioning costs before deployment.

Where the bill arrives

Picture the moment after the launch event. The hospital billing system has found a second diagnosis. A farmer stands outside a new fence. A procurement officer has a vehicle worth hundreds of millions on paper. An engineer opens an incident report because an agent crossed a boundary. None of these scenes proves the technology failed. Each shows the exact place where the original efficiency claim stopped counting.

The next serious debate about AI economics should begin there, not with the demo. Put the beneficiary and the cost bearer in the same frame. Put the ceiling beside the obligation, the diagnosis beside the treatment, the investment beside the delivered local benefit, and the automation beside the cleanup. If the bill disappears when those columns meet, we have found a genuine efficiency. If it merely changes hands, we should say who is paying before we call the system progress.