The BIS ran the AI growth boom through a model. Its best case is 5.5%, not 18%

Every AI bull case has a growth number in it. Some recent academic models put annual output growth at 13% to 18% within two decades once AI turns transformative. A new BIS Working Paper, number 1381, published October 12, 2026, runs the same question through a different engine and gets a much smaller answer.² Let's debug it.

The bug the paper is looking for

The author, Phurichai Rungcharoenkitkul, a principal economist in the BIS Monetary and Economic Department, starts from a plain observation. Automation at AI scale means building and maintaining a lot of machines. That pushes the share of output going to investment up and the share going to consumers down.

Here is the catch. In his model, innovators earn their temporary monopoly profits on consumer products, the way they historically have with phones, cars and medicines. Research costs scale with the whole economy. Research rewards scale with consumer spending. If automation thins the consumer market enough, the next AI innovation stops paying for itself. Entry stops, automation stalls, and the growth engine slows. He calls this the demand bottleneck.

Five scenarios, one calibration

The paper simulates a pre-AI baseline with 2% trend growth and a 60% labor task share, then four AI outcomes.⁴

• Transformative: AI profits last long enough that entry keeps going. Growth reaches 5.5% at year 20 and output is 39% above the no-AI path.¹
• Demand bottleneck: trend growth rises briefly, then falls below the pre-AI 2% within a decade as innovation permanently slows. Measured output growth stays higher for longer because machines are still being built: at year 20 it is 2.3%, with output 6% above the no-AI path.¹
• Bounded boost and Automation without innovation sit in between.

The headline claim is not that the best case will happen. It is that when the model is calibrated to the available evidence, the demand bottleneck comes out as "highly plausible." That phrase appears in the abstract and the conclusion. The paper's own disclaimer says the views are the author's, not the BIS's or its member central banks'.

Why the answers differ so much

The comparison table is the most useful page for anyone reading AI growth forecasts. At year 20, Jones and Tonetti (2026) grow at 13% in their Moore's law case, and Korinek and Suh (2024) at 18% in their baseline.² The BIS model's best case is 5.5%.

The author names three reasons. First, the trigger: other studies let automation advance on an outside schedule, while this model makes each AI advance a business decision that has to be profitable. Second, calibration: the optimistic scenarios elsewhere automate a quarter of tasks in four years and everything in two decades. Third, and the one he calls most consequential: elsewhere research is a fixed share of output, so it never asks who funds it. Here it must be paid for out of consumer spending.

That is the finance angle. An AI capex boom is a bet that someone eventually buys enough of what the machines make. This paper is a formal version of that question.

Workers and rates

Real wages run below the no-AI path for decades in every scenario with more automation. At year 20, the wage is about 15% below that path under the demand bottleneck and 54% below it under the transformative case, because output rises slowly while machines are built out of savings.⁵

For central banks, the payoff is the natural rate of interest, r-star. The textbook view is that faster AI productivity lifts it. Across 2,000 parameter draws, the paper's main calibration puts the median natural rate at year 20 at 1.2%, about half a point below the pre-AI level it anchors to the Laubach-Williams estimate.³ With a wider range of assumptions, the median is 1.6%. The author's own conclusion: uncertainty is very high, and it is hard to say even which direction r-star travels.

What breaks the bottleneck

In the model, a permanent 10% cut in the cost of conventional research, for example through subsidized use of AI tools by ordinary researchers, returns trend growth to 2% after the stall.⁴ Redistributing income between households does nothing for growth, because entry depends on the share of output consumed, not on who consumes it.

The limits, in the paper's own words

This is one model with one rent mechanism. It abstracts from security risks and from supply-side bottlenecks such as compute and energy limits, which could bind before demand does. The scenario paths are projections from a calibrated theory, not forecasts with error bars. The year-20 comparisons line up studies with different start dates, and the author flags that the alignment is not exact. The paper also notes that AI was used for research assistance and editorial refinement.

The takeaway

When someone quotes an AI growth number, ask what funds the next round of AI innovation in their model. If the answer is "nothing, it just happens," you are reading the double-digit version. Not advice. What AI forecast should we put through the debugger next?