AI is usually discussed as a productivity story. If firms can automate cognitive tasks at lower cost, output should rise, margins should improve, and resources should move toward more productive uses.

But productivity is only one side of the equation.

In his October 2026 working paper Burning the Bridge: Imported Intelligence, Reversibility Loss, and External Liquidity, Eduardo Levy Yeyati asks whether foreign-provided AI can make an economy more productive while simultaneously making it more externally dependent. His answer is yes — under a specific combination of capability erosion, concentrated foreign supply, and scarce external liquidity. 2026_13 Burning the Bridge

Levy Yeyati describes foreign agentification as a form of “black-box offshoring.” A firm is not simply importing software. It is delegating a productive task to a technological system supplied, maintained, and updated abroad — potentially combining models, compute, orchestration, proprietary tools, workflow design, and support. 2026_13 Burning the Bridge

The immediate economics can be compelling. But as firms stop performing those tasks locally, the domestic capability required to bring them back can begin to erode.

In Levy Yeyati’s framework, if aggregate agentification is a, domestic capability evolves as:

More agentification can mean less domestic capability.

That capability is broader than employment. It includes workers, training pipelines, organizational routines, suppliers, and the tacit knowledge required to perform the task locally.

And that matters because domestic capability is also an outside option.

If firms can credibly bring an activity back in-house, foreign providers have less room to extract rents. If reconstructing that capability becomes expensive, the bargaining position changes.

Levy Yeyati represents the foreign-agent price as:

where m is the provider’s base cost and captures the share of the weakened domestic outside option that foreign suppliers can monetize.

The implication is subtle: today’s adoption can make tomorrow’s imported intelligence more expensive.

Each individual company may rationally adopt AI because the productivity gain exceeds its expected cost. But no individual firm fully internalizes how its decision contributes to the erosion of the broader sectoral pool of workers, routines, and suppliers.

That is the first externality.

The paper expresses the additional social cost from this loss of reversibility as:

The intuition is straightforward. One additional unit of agentification may not only affect the marginal task. If it weakens the domestic fallback, it can raise the price paid on the entire installed base of imported agents.

But the second channel is even more important from a macro perspective.

Agentification can change the currency composition of production costs.

A firm that previously paid domestic salaries may begin paying recurring fees to foreign AI, software, and cloud providers instead.

Productivity rises.

But part of the domestic wage bill has become an imported-services bill.

This creates a distinction that is particularly relevant for emerging markets: productivity generation is not the same thing as foreign-currency generation.

A domestic retailer can automate accounting, customer service, or administrative work and become significantly more efficient without earning one additional dollar abroad.

An exporter using the same technology may increase output and generate enough additional foreign revenue to finance the imported-intelligence bill.

So the relevant macro question is not simply:

Does AI raise productivity?

It is:

Does AI generate enough tradable income to finance the foreign-currency obligation it creates?

Levy Yeyati’s external-liquidity condition captures exactly that tension:

The left-hand side is the marginal foreign-currency income generated by agentification. The right-hand side is its marginal foreign-currency cost, including the possibility that greater dependence raises the price paid on existing use.

When the inequality holds, more agentification tightens the external constraint even while productivity rises. 2026_13 Burning the Bridge

The chart illustrates the core mechanism. As foreign agentification expands, the recurring FX bill associated with imported intelligence can rise faster than the FX earnings generated by those productivity gains.

The widening gap is the external restriction.

This is where the argument becomes particularly relevant for emerging markets.

Development has historically been constrained not only by the ability to produce, but by the ability to finance imports of capital goods, technology, energy, and intermediate inputs without eventually running short of foreign currency.

AI could introduce a new version of the same problem.

Instead of importing only machines, economies may increasingly import productive intelligence itself — not as a one-off investment, but as a recurring foreign-currency obligation.

The paper is careful not to turn this into an argument against AI adoption. If agentification raises exports or substitutes for other imports, the external-liquidity effect can shrink or even reverse. Likewise, if foreign AI provision becomes highly competitive and systems become genuinely portable across vendors, the rent-extraction channel weakens sharply. 2026_13 Burning the Bridge

That makes the policy question less about preserving old jobs and more about preserving valuable options: human capability, training pipelines, interoperability, data portability, supplier competition, contractual protection, and access to alternative compute. 2026_13 Burning the Bridge

For investors, the same logic applies at the company level.

The question is no longer simply whether AI lowers costs.

It is whether those savings create a genuinely more productive business — or exchange one operating cost for a new dependency that becomes progressively harder and more expensive to unwind.

Levy Yeyati’s paper puts a useful macro framework around that risk.

Cheap intelligence can create enormous value. But when the capability behind it sits somewhere else, cheap intelligence can also become expensive dependence.

Levy Yeyati, E. (2026). Burning the bridge: Imported intelligence, reversibility loss, and external liquidity (Documento de Trabajo 2026/13). Escuela de Gobierno, Universidad Torcuato Di Tella. 2026_13. https://www.utdt.edu/ver_contenido.php?id_contenido=27390&id_item_menu=31549