How this editorial can be challenged
When an AI interaction feels effortless, has the work disappeared or moved to somebody else?
A user-facing shortcut can conceal review, consent and recovery work borne by employees, customers or counterparties; measuring only generation or transaction costs can reward cost-shifting instead of true efficiency.
Early technologies have transition costs, and forcing every uncertain downstream cost into each transaction could smother useful innovation and deny consumers better services.
Proportionate accounting should distinguish learning costs from persistent failures, compare against real alternatives, and enable bounded experiments rather than demand zero risk.
An executive survey cannot establish audited AI returns; airline statements cannot settle the balance of security and distribution incentives; climate-emulator research does not validate every downstream policy decision.
Transparent comparative workflow studies showing durable gains after review, error recovery and affected-party costs, without materially reducing customer control, would justify more expansive deployment.
The person behind the effortless result
There is a particular kind of technological convenience that feels almost magical: ask a question, receive a finished answer, move on. But the person moving on may not be the person responsible for checking the answer, repairing the mistake or explaining it to someone affected. Our argument is not that the convenience is fake. It is that its apparent price can leave out the work that makes it usable.
That omission matters for ordinary working lives. If an assistant saves a manager ten minutes but creates an hour of correction for a colleague, a dashboard measuring only the manager's activity will celebrate the wrong result. This is an illustrative example, not a finding from today's sources. It shows why productivity needs a boundary wide enough to include the people who inherit the consequences.
The budget is a clue, not a verdict
Capgemini's new survey reports that 59% of its 500 large-company executive respondents experienced significantly higher AI expenditure than expected or budgeted. The research also emphasizes poor visibility into indirect costs. The finding warrants attention, but it cannot tell us that all AI investments fail. It records executive reports, not a controlled comparison of audited returns.
An overrun can accompany a valuable expansion of work. It can also reflect an unrealistic forecast, expensive integration or repeated failure. The responsible response is to distinguish these possibilities. Ask what useful task was completed, how well, how often it needed correction, and what the same outcome would have cost through a plausible alternative. A lower model price is one input to that calculation, not its answer.
A flight booking makes the hidden contract visible
United says it is restricting unauthorized third-party agents, citing account security and other concerns. That confronts a real problem: a request to find a convenient trip is not necessarily consent to spend money, change an account or accept every condition associated with a ticket. An assistant that can operate a website has not automatically acquired legitimate authority to do so.
It would nevertheless be too easy to treat every restriction as a public service. Airlines also have an interest in controlling how customers reach them. A customer-friendly system should make authorized delegation practical while bounding what can happen. The design question is who can permit the action, inspect its record and correct it—not whether the interaction contains the letters AI. These are standards we propose, not claims that the named airlines already satisfy them.
Do not turn accounting into a veto
The strongest objection is fair: new technologies require learning, and perfect accounting is impossible. Some human review will be temporary. Some expensive experiments will unlock lasting improvements. A rule that allows deployment only when every possible consequence is priced would protect established methods from useful challengers.
The answer is proportion, not paralysis. A reversible draft needs a different level of control from an irreversible purchase or a decision affecting another person's rights. Track the costs most likely to matter, set a review period and make the comparison explicit. Do not quietly classify recurring correction as a transition cost forever. Equally, do not assume that a difficult first deployment disproves the possibility of a better one.
There is a better kind of efficiency story
The climate-emulator work highlighted by MIT offers a useful counterexample to the assumption that progress must mean more inputs and more infrastructure. Its researchers improved generalization through deliberately designed training scenarios. The study tested simple and intermediate-complexity climate models; extending the approach to full-scale Earth-system models remains an important step ahead.
We should welcome that result on its own terms. It aims to make a research tool more useful per costly training run, with a comparison that readers can examine. It does not eliminate the need for scientific judgment or validate every policy use. The lesson is methodological: define the outcome, test the alternative and show the remaining limits. Honest boundaries can make an efficiency claim stronger.
Look behind the screen
What would change our assessment is evidence of durable gains measured across the whole workflow: fewer errors to repair, less total effort for an acceptable result, clear consent and no unacknowledged transfer of costs to people with less power to refuse them. Such evidence would support broader use. If benefits disappear once those costs are counted, the remedy may be a smaller model, a redesigned task or a human-led process—not a louder promise.
Picture the colleague who opens the next morning's queue. The assistant has already produced its answer and moved on. That colleague sees what remains: the exceptions, the disputed transactions, the awkward explanations. A genuinely better AI system should make that person's day better too. Until we look at their desk as carefully as we admire the interface, we have not finished counting.
Read the reporting
Opinion is ours. The factual record is linked below.
Capgemini — AI economics survey release and methodology United — Restrictions on unauthorized third-party AI agents MIT — Training climate emulators with optimized scenarios PNAS — Optimal scenario design for climate emulation