Guiding the AI disruption to the Good Place

[MUSIC]

[MUSIC FADES INTO SWEEPING SOUND]

DAVID ROTHSCHILD: Hi, my name is David Rothschild, and I’m an economist at Microsoft Research in New York City.

I’m here to talk about a series of papers I’ve been working on with teams in New York and New England in economics and computation. We’re excited about this work because, rather than focusing only on short-term AI impacts, we’re looking more broadly and long term—what happens when AI and agents become ubiquitous? How does that affect markets, society, and all of us?

Let’s start with a puzzle. AI capabilities feel extraordinary, yet large-scale economic disruption still seems muted. Our explanation isn’t about model quality or compute—it’s about how AI is integrated into workflows.

The first key point is that we distinguish three phases of disruption.

In Stage 1, augmentation, AI improves and accelerates individual tasks like writing, summarizing, and coding within workflows designed for humans.

In Stage 2, automation, routine tasks move “under the hood.” Humans supervise at a high level, but the underlying workflows—forms, approvals, queues—remain largely unchanged. These human-centered structures become bottlenecks that limit further gains.

The real disruption comes in Stage 3: reconstruction. Here, workflows and markets are redesigned from the ground up around AI’s native strengths—parallelism, memory, continuous monitoring, and machine-to-machine interaction. Human-centered interfaces are no longer the constraint.

Most gains today are still in Stage 1 and early Stage 2. But the largest long-term benefits will come from Stage 3—and getting there is institutionally difficult.

This leads to the second key point: the main barriers to Stage 3 adoption are not technical. Reconstruction requires delegating real authority to agents, building trust and accountability systems, ensuring machine-readable data and constraints, and aligning incentives to reward redesign rather than incremental improvement. It’s an innovator’s dilemma—replacing existing systems with fundamentally new ones.

Like past general-purpose technologies—electricity, computers, the internet—AI has a long adoption curve. The bottleneck lies in organizations and governments, not raw capability.

The third key point is that as we move into Stage 3, coordination becomes central. Communication frictions and system design choices matter more than raw compute.

Lower communication barriers and less lock-in enable broader coordination of tasks and payments, increasing returns to innovation and distributing value more equitably. Open ecosystems tend to benefit society, while closed systems concentrate value among incumbents.

This applies to workers as well—their share of value depends on how delegation, monitoring, and access are structured.

The fourth point is more cautionary: the default trajectory favors entrenchment. Incumbents with established user bases are incentivized to maintain control, while building open, safe ecosystems requires costly and complex coordination across organizations and governments.

The result may be acceleration without transformation—AI makes existing platforms faster, but doesn’t fundamentally change them, keeping us stuck in Stage 2.

So how do we reach a better outcome?

We propose building synthetic agent-based markets to study and influence how transitions to Stage 3 unfold. These environments allow agents representing consumers and businesses to interact under controlled conditions.

They help us observe dynamics that are difficult to isolate in real-world settings—such as shifts from human-to-agent to agent-to-agent interaction, or what happens when agents are given real authority rather than advisory roles.

We can also study governance mechanisms—constraints, monitoring, auditability—and test resilience by introducing adversarial or malicious actors.

These experiments make abstract ideas concrete. They show that simply layering AI onto existing workflows yields limited benefits compared to redesigning systems to fully leverage AI capabilities.

Alongside experiments, we develop theoretical models to study these markets at a higher level. These models focus on comparative insights rather than precise predictions—helping us understand how different design choices, such as open vs. closed ecosystems or centralized vs. decentralized coordination, shape outcomes.

The key insight is that architectural choices strongly influence market behavior and value distribution—even when underlying AI capability is held constant.

So what are the takeaways?

AI disruption is real, but slower than it may appear. Its largest impacts will come from full system reconstruction, not just task-level improvements. Reaching that future requires combining systems thinking, experiments, and theory.

Early decisions will matter. The way we design these systems now will shape long-term outcomes for markets, society, and overall welfare.

Waiting is itself a choice—and one with consequences.

Thank you.