AI and the Boundary of the Firm: Evidence from Radiology (CESifo WP 13086)
A · Strong evidence University of Essex economists combine a formal model with US Medicare claims for 2013-2024 and a pathology placebo test; the method is laid out in full and the national figures can be checked against public data.
Summary based on the full text·Original: about 90 minutes to read
In short
CESifo Working Paper 13086 (October 2026) by Marco Francesconi and Xiaoyu Xia asks why radiology, the specialty most exposed to AI, has more radiologists than before and no rise in reads per radiologist. Their answer: AI lowers the cost of verifying remotely supplied interpretations, so hospitals outsource reads to multi-market groups. Work leaves small local markets and concentrates in large platforms, without the task itself being automated.
Key numbers
- 16 percentrise in radiologists billing Medicare, 2013 to 2024, while interpretations per billing radiologist fell 16 percent
- 9 to 47 percentshare of radiologists in groups operating in five or more states, 2017 to 2025
- about a fifthextra loss of locally billing radiologists by 2024 in markets one standard deviation more exposed
- approximately a quartershare of interpretation production the calibrated model says was reorganised across markets
- about 2.5 percentaggregate reader capacity saved by that reorganisation, per the model
As reported by the source.
English·Web page·opens d.repec.org
What it reports
- Radiologists billing Medicare rose from 30,708 (2013) to 35,705 (2024), 16 percent; billed interpretations fell from 81.9 million to 79.8 million as the fee-for-service population shrank. Interpretations per billing radiologist fell 16 percent.
- The share of radiologists in groups spanning five or more states rose from 9 to 47 percent (2017-2025); the number of such groups rose from 79 to 448. Radiologists listing practices in several local markets rose from 20 to 36 percent.
- By 2024, a market one standard deviation more exposed had lost about a fifth more of its locally billing radiologists. The loss is concentrated in markets that started with fewer than ten radiologists and absent in larger ones.
- Imaging demand by patient residence did not fall the same way: small markets bill fewer reads, large markets bill more than their own residents generate.
- Of radiologists billing from small markets in 2017, about 30 percent had stopped billing Medicare by 2024, against 21 percent in the largest markets.
- Placebo: pathology serves the same hospitals and payer but had almost no reader-facing AI; its local billing shows no similar pattern.
- The calibrated model puts about a quarter of interpretation production as reorganised across markets, equal to about 2.5 percent of aggregate reader capacity.
How they got it
A model of AI-assisted verification of credence-good work, tested on Medicare claims, physician practice directories and the diffusion of FDA-cleared reader-facing AI. The design interacts national AI diffusion with each market's prior radiology intensity, with market and year fixed effects, across 769 metropolitan markets. A calibration reproduces the decline in local billing in small markets.
Limits
- The design identifies how exposure relates to local billing as AI spreads, not the causal effect of any one practice adopting AI (the authors say so).
- Medicare fee-for-service only; Medicare Advantage and private payers are outside the data.
- The 2.5 percent capacity figure is a model output, not a measured saving.
- Entry by documented AI-adopting platforms is not significantly different from entry by other multi-state groups.
Why it matters for companies
AI can change who does expert work and where before it changes how much one expert produces. Here the gain shows up as consolidation and outsourcing, not as more reads per head. Firms that buy professional services should expect providers with verification tooling to pool work across clients, and small local providers to lose volume. Headcount forecasts built on task automation alone miss this organisational channel.
Summary based on the full text. Read the original for the full argument and data.Read the original on d.repec.org