The economic promise of artificial intelligence rests on a demanding assumption: better models will generate enough useful work to pay for the infrastructure built around them.
More intelligence can help. It does not automatically produce more productivity. A system becomes economically valuable when it completes useful work reliably at an acceptable total cost.
The central metric is cost per correct decision: the full cost of producing a decision that meets independently assessed requirements and leads to the intended outcome.
This distinction exposes three pressures on the AI economy: the cost of additional capability, the revenue available to finance it, and the difficulty of translating capability into dependable production. The third may determine the other two.
The triple squeeze: cost, revenue and productivity
At the frontier, additional capability can require more training resources, longer inference or more elaborate tool use. Whether that expenditure pays off depends on the task. An expensive reasoning step may be indispensable for one problem and wasteful for another.
Training costs, inference costs and customer prices must remain separate. A costly benchmark run does not establish diminishing returns to training. A cheaper API does not establish lower production costs. Nor does a modest benchmark improvement reveal its commercial value.
The revenue constraint is equally demanding. Even rapidly growing sales must eventually support infrastructure, operating expenses and a return on capital. Slower percentage growth alone would not establish a crisis; margins, utilisation and absolute revenue matter too.
For scale, McKinsey’s April 2025 scenario analysis estimated $5.2 trillion in capital expenditure for AI-capable data centres by 2030. That is a cumulative investment estimate under modelling assumptions, not annual spending or a revenue forecast. It illustrates the financial stakes without proving overinvestment.
The productivity constraint connects the two: which processes will generate sufficient customer value to sustain this spending? Benchmark intelligence alone cannot answer that question.
Productivity consists of billions of small decisions
Consider a robot wiping dust from a table. It identifies surfaces, adjusts pressure, avoids obstacles and checks whether cleaning is complete. Its value comes from reliably completing the job with little supervision.
Invoice processing, logistics and insurance administration follow a similar pattern. Many individual decisions are modest; their frequency and consequences make them economically significant.
A stronger model can improve these workflows. But repeatedly invoking a large model for a decision already captured by a reliable rule can add cost without adding value. The relevant unit is a useful, correctly completed operation.
As discussed in AI Tokenomics , token prices describe only part of that cost.
Define cost per correct decision carefully
A practical definition is:
Cost per correct decision = total workflow cost / number of independently assessed correct decisions completed.
Total workflow cost includes inference, tools, infrastructure, integration and maintenance allocated over actual volume, human review, retries, rework, and measured or expected losses from errors. Avoid double-counting realised losses and provisions for the same failures.
The numerator includes expenditure on failed and abandoned attempts. The denominator includes only decisions meeting the agreed requirements. Passing an automated check is evidence of correctness; it is not automatically correctness itself.
An illustrative comparison makes the difference concrete:
| For the same 10,000 comparable cases | Model-heavy workflow | Hybrid workflow |
|---|---|---|
| Total workflow cost | €12,000 | €8,000 |
| Independently assessed correct completions | 9,600 | 9,600 |
| Cost per correct decision | €1.25 | €0.83 |
These are hypothetical numbers, not measured results. The hybrid system wins in this example because it achieves the same correct throughput at lower total cost.
Comparisons must also hold task mix, deadlines and acceptable error severity constant. A correct rejection or escalation can count when that is the specified task outcome; repeatedly deferring difficult cases cannot masquerade as completing them. Report coverage, unresolved cases and serious errors alongside the ratio. Fix the decision unit in advance so that splitting one operation into many trivial steps cannot inflate performance.
A plausible explanation is not an audit trail
Language models can extract information, propose actions and explain answers. These abilities do not guarantee that an explanation faithfully describes the factors that produced an output.
Anthropic’s April 2025 research found that the reasoning models tested often omitted their reliance on experimentally supplied hints. This is evidence about those evaluations, rather than a universal verdict on every model or explanation.
Three properties therefore need separate assessment:
- Plausibility: the explanation sounds reasonable.
- Correctness: the decision meets the substantive requirements.
- Auditability: relevant evidence, rules, authority and state changes are recorded and inspectable.
A system can satisfy one without satisfying the others. Asking another model to review the answer may help, but adds cost and can reproduce shared errors.
Falsification can change the architecture
An alternative separates probabilistic interpretation from decision authority.
In an insurance workflow, a model extracts candidate facts from a claim and links them to source passages. An authoritative system supplies the applicable policy version. Explicit logic checks dates, coverage, exclusions, limits and missing evidence. Contradictions trigger further investigation or human review. An authorised component records any resulting state change.
The useful question becomes: what evidence would make this proposed decision fail?
That is falsification in an operational sense. A hypothesis must be challengeable, and the challenge must arrive before unacceptable consequences. The boundary for autonomous action therefore depends on the particular action.
Explicit rules have limitations. They can encode the wrong policy, operate on bad data or miss an ambiguous exception. A graph can preserve a false claim perfectly. Determinism does not establish truth .
The economic hypothesis is narrower: reusable state, explicit checks and independent evidence may reduce repeated inference and review while preserving quality. That advantage must be measured, including the cost of maintaining the checks.
Metacognition must demonstrate operational value
Synthetic metacognition would need to do more than generate reflective prose. It would need to detect missing knowledge, investigate contradictions, revise mistaken state assumptions and use calibrated uncertainty to change subsequent actions.
For the economic argument, the important test is whether those functions improve outcomes at an acceptable cost. A convincing self-critique is insufficient evidence.
Larger models, hybrid systems and specialised architectures are possible routes. None should be booked in advance as a guaranteed productivity breakthrough.
The hardware paradox
If reliable work requires less computation per decision, AI’s economic value could increase while compute intensity falls.
Persistent state, specialised models, deterministic transformations and efficient hardware could all contribute. Businesses would benefit if these approaches reduced total cost per correct decision.
Aggregate compute demand could still rise as lower costs make more applications viable. Efficiency therefore guarantees neither falling GPU demand nor proportional infrastructure growth. The uncertain variable is how much additional demand it creates.
Growing usefulness and growing demand for frontier compute are related possibilities, not interchangeable quantities.
Test the decision economy
The shift from model-centred to decision-centred AI is ultimately an empirical proposal. Compare alternative architectures on the same representative workload. Define correctness before evaluation, assess outcomes independently, include integration and failure costs, and report latency, coverage and error severity.
A hybrid architecture that costs more to maintain than it saves has failed its economic test. A stronger model that prevents costly mistakes may justify its price.
The next major productivity gain could come from making billions of ordinary decisions cheaper and more reliable. The industry is investing heavily in computational capacity; its customers need dependable outcomes.
Cost per correct decision is the bridge between those two economies.
Editorial note: The supplied draft referred to Economist charts and an approximately $3 trillion annual revenue requirement whose underlying sources were not provided. Those figures and the asserted revenue slowdown are not treated here as verified findings. The architectural benefits proposed above remain hypotheses to test.