Automated insulin delivery increases, not reduces, routine specialist engagement: adopters become substantially more likely to see a diabetologist within semesters after initiation (≈16 ppt at 6 months, ≈39 ppt at two years), suggesting patient-facing automation can create new clinical monitoring and interpretation tasks and raise demand for specialist care.

Main Finding

Automated insulin delivery (AID) adoption is associated with an increase in routine outpatient specialist engagement. After matching and accounting for differential pre-adoption trends, the probability of a diabetologist visit rises by ~16.3 percentage points one semester after adoption and by ~39.3 percentage points four semesters after adoption (reference pre-adoption visit probability = 55.2%). HbA1c testing and a process-of-care index move in the same direction but with weaker evidence. The visit estimates, however, are sensitive to the identifying extrapolation assumption and should be interpreted conditional on that restriction.

Key Points

• Research question: Does patient-facing medical automation (AID) substitute for or complement professional clinical care?
• Theoretical framing: Competing forces
  • Substitution (Grossman-style productivity gains; risk compensation): automation reduces need for professional input.
  • Reinstatement/task-creation (Acemoglu & Restrepo-style): automation displaces some tasks but creates new interpretive, supervisory, and adjustment tasks that raise demand for clinician time.
• Empirical finding: Evidence is consistent with complementarity—AID appears to increase routine specialist involvement rather than decrease it.
• Caveat: Estimates rely on (and are sensitive to) an explicit extrapolation of pre-adoption trends into the post-adoption period; results are presented as conditional on that identifying assumption (following Rambachan & Roth, 2023).

Data & Methods

• Data
  • Sample: 1,608 adults with type 1 diabetes treated at four specialist clinics in Emilia-Romagna (Italian NHS).
  • Adopters: 283 patients who initiated AID; after matching, 181 treated adopters retained with comparable untreated controls.
  • Outcomes: Probability of diabetologist visit (semester-level), HbA1c testing, and a process-of-care index.
  • Context: Publicly funded care with chronic-disease exemptions (little/no point-of-care price); prescribing diabetologists retain post-adoption monitoring responsibility.
• Identification challenges
  • Adoption is clinically targeted and staggered over time; untreated controls may have different pre-adoption trends.
  • Staggered adoption creates potential negative-weighting biases with conventional two-way fixed effects/event-study estimators.
• Estimation strategy (three steps)
• Risk-set coarsened exact matching: matches adopters to not-yet-adopted/never-adopted patients with comparable pre-adoption clinical histories within risk sets (retained 181/283 adopters).
• Interaction-weighted event-study (Sun & Abraham, 2021): constructs dynamic treatment effects using comparisons of newly treated vs. not-yet-treated/never-treated groups, avoiding negative-weighting of cohorts.
• Adjustment for differential pre-adoption trends: extrapolate untreated trajectory in event time into the post-adoption window (following Dobkin et al. and Freyaldenhoven et al. approaches) and assess sensitivity; interpret results conditional on the low-dimensional event-time path restriction.
• Robustness & sensitivity
  • The diabetologist-visit effects persist after the pre-trend adjustment but are sensitive to departures from the extrapolation assumption; authors therefore emphasize conditional interpretation.
  • Results for HbA1c testing and the process-of-care index are in the same direction but less precise.

Implications for AI Economics

• For cost and capacity assessments, device cost and clinical effectiveness are necessary but not sufficient. Patient-facing automation can reconfigure task composition and increase demand for specialist labor (interpretation, adjustment, supervision), shifting resources toward sustained specialist involvement.
• Workforce and capacity planning: Increased routine specialist engagement implies higher demand for specialist time even when automation improves clinical metrics—important for budgeting, staffing, and wait-list management in public systems.
• Reimbursement and procurement: Payment models that focus only on device unit costs risk underestimating full system costs if adoption systematically increases clinician follow-up and monitoring needs.
• Regulatory and implementation design: Policies that clarify roles and responsibilities (who monitors, who is liable for device decisions) and invest in training can shape whether automation complements or substitutes clinician input.
• Evaluation methodology for AI/automation interventions:
  • Empirical strategies must account for clinical targeting and staggered adoption; standard TWFE/event-study estimators can be biased under treatment timing heterogeneity.
  • Combining risk-set matching with interaction-weighted event-study estimators and explicit sensitivity checks on pre-trend extrapolation is a promising approach but requires transparent reporting of identifying assumptions.
• Open questions and limits for generalization:
  • Mechanism ambiguity: The study cannot definitively separate task reinstatement (new clinical tasks) from a scale effect (patients/clinicians choosing tighter control because control is cheaper) or from administrative/training processes associated with adoption.
  • External validity: Data come from four specialist clinics in a single Italian region and a publicly funded system—responses may differ in other health systems, patient populations, or with different AID implementations (e.g., DIY systems excluded here).
  • Policy takeaway: Adoption of patient-facing AI/automation can increase—rather than reduce—demand for professional input; economic evaluations of such technologies should explicitly model downstream changes in clinical workload and care composition, not only device costs and acute-care offsets.