Innovation Economics · Citation chart

How much do the canonical productivity studies actually claim?

Five studies, one chart. The bar is the headline percentage uplift. The row beneath is the methodology meta — sample size, study year, whether the experiment was blinded — because the difference between a 14% measured field experiment and a 45% modelled scenario is the entire argument.

BCG × Harvard Management consultants, complex tasks

Blinded N = 758 2023 Navigating the Jagged Frontier ↗

GPT-4 inside the frontier. Lower uplift, even quality decline, on tasks beyond it.
MIT NBER Customer support agents

Not blinded N = 5,179 2023 Generative AI at Work ↗

Largest gains accrued to the lowest-tenure agents. Compresses the experience curve.
GitHub × Microsoft Software engineers, narrow task

Not blinded N = 95 2022 Quantifying Copilot impact ↗

Time-to-complete on a narrow JS authoring task. Self-reported satisfaction also up.
McKinsey Global Institute Knowledge work, modelled scenario

Not blinded N = modelled 2023 The economic potential of generative AI ↗

Top-of-range modelled estimate. Treat as upper bound, not measured outcome.
Stanford × MIT Call centre operators

Not blinded N = 5,179 2024 Generative AI at Work (call centre) ↗

Same dataset as MIT NBER, follow-up analysis. Skill ceiling effect for top performers.

Read the chart this way

Field experiments cluster at 12–14%. Narrow vendor-conducted demos and modelled scenarios reach 45–55%. The defensible expectation for an enterprise rollout is the lower band, conditional on the work falling inside the model's frontier and the workflow being rebuilt around it. The upper band is what shows up in board decks; the lower band is what shows up on the income statement.

← All models