AI productivity gains vs the AI bubble: what the data shows

What do the data say about AI productivity gains and the AI bubble?

The data say two things at once. Projected gains are large, and delivered gains are uneven. The EQOS Asia AI study lists productivity gains of +40% beside a 95% failure rate for deployed projects. A bubble reading focuses on the second number, and a boom reading focuses on the first. Productivity data matters because it is the only evidence that comes from the buyer’s side.

A useful way to read it is by stage: projection, delivery, quality, people, and recovery. Each stage below takes one part of the study and asks how much of the gain survives. By the end you can tell which share of the +40% is a forecast and which share a business can bank. Keep the stages in this order, because a gain lost at an early stage cannot reappear at a later one.

How large are the projected AI productivity gains in the study?

The headline figure is +40%, a projected productivity gain for firms that integrate AI well. Beside it the study places 300 million jobs that change, which shows the reach of the technology across sectors. Both numbers describe potential rather than results, and they carry the date April 2026. Ask which sector and which task each figure covers before you reuse it in a plan.

McKinsey, cited by the study, estimates that 30% of hours worked could be automated by 2030. Automation of hours frees time, and freed time becomes productivity when it moves to work that earns more. Treat these figures as a ceiling for the boom reading, and use the next sections to find the floor.

Where do AI productivity gains get lost between pilot and production?

The study traces failures to causes inside the buyer’s organisation. Insufficient or biased data appears in 42% of failed projects, a missing link to strategy in 31%, team resistance in 24%, unproven return in 22%, and ethics or bias issues in 18%. Each cause holds a pilot back before it adds to output.

These causes explain why a large projected gain and a high failure rate coexist. The technology can work in a demo and still lack the data, the sponsor, or the team to work in daily operations. Gains arrive where those three exist, so a buyer can read its own readiness before a purchase. Score each of the five causes from 0 to 1 inside your own firm, and start with the lowest.

How do AI errors reduce the productivity gains you can count?

Errors cost time. Large language models produce false answers between 0.7% and 30% of the time depending on the model and the domain, and the rate reaches 67% in a medical context (JMIR 2025, cited by the study). A worker who checks every answer keeps part of the time saved and spends the rest on review. Measure review time per answer in a small test, and subtract it from the hours you claim.

Confidence adds a second cost. The study reports that models are 34% more confident when wrong, and that 47% of enterprise users took at least one major decision based on hallucinated content. A decision built on a false answer can erase many hours of saved effort, so quality control belongs inside every productivity estimate.

How does workforce retraining change AI productivity gains?

People turn tools into output. The study notes that fewer than 15% of companies have a retraining plan for affected employees, and it scores the Care dimension of the market at 0.338. A workforce that expects to be replaced adopts tools slowly, which delays the gain.

The study recommends a skills plan for each automated role, funded at 2% to 5% of total AI investment. At the market’s $200B a year, that range runs from $4B to $10B, small beside the spending it protects. The recommendation targets Care from 0.338 to 0.50 and positions people as the multiplier of the gain.

How much AI productivity value could better deployment recover each year?

The study models four corrections: measure data quality before deploying, require explainability, align the AI strategy with the field, and fund retraining. Together they lift the market score φ(S) from 0.337 to 0.401 and lower the predicted failure rate from 94% to 82%. The study lists the corrections in a fixed order and starts with a data quality audit before any deployment.

The study converts the change into money: recovering 40% of the $200B spent each year gives $80B a year of recoverable value. This figure bridges the two readings. A boom reading sees the $80B as proof that gains exist, and a bubble reading sees the other $120B as spending that still waits for a return.

How should you weigh productivity data against the AI bubble?

Weigh the data by stage. Projections support the boom reading, delivery data supports whichever side matches your own sector, and the gap between them sizes the bubble risk. Update the weights each quarter as buyers report results, and write the source and date beside each number. A reader who repeats the exercise four times in a year holds a trend, and a trend carries more weight than any single estimate.

Your own organisation may sit above or below the market score. The Λ simulator gives a first estimate in 10 minutes, free and anonymous, and the EQOS Asia 30-minute pre-diagnostic applies the same scoring to a decision in Singapore, ASEAN, or Hong Kong (EQOS Asia AI study, April 2026).

Frequently asked questions

About the author

Stéphane Jambu

President, EQOS Forecast Asia

Stéphane Jambu leads EQOS Forecast Asia and works as a consultant in e-commerce, search visibility, and AI. LinkedIn profile