A reproducible index for 'agentic' AI in banks promises to separate autonomous execution from advisory AI in disclosure data and enable credible causal analysis; the paper supplies a validated AAII measure and a battery of complementary identification strategies but stops short of applying them to produce outcome estimates.

Main Finding

This Method Article develops and validates the Agentic AI Intensity Index (AAII), a disclosure‑based, bank‑level measure that distinguishes genuine agentic (delegated‑execution) AI from advisory or generic AI adoption. The paper specifies a fully reproducible five‑step construction and validation protocol (including human + large‑language‑model classification with adjudication), an authority‑weighted scoring rule and normalization to isolate agentic intensity net of overall AI talk. It also provides a matched identification strategy (multiple causal designs and threat mitigations) required to estimate AAII’s effects on bank efficiency, risk detection, operational risk and market reaction. The contribution is methodological: the index and the inferential machinery are presented as a directly executable protocol rather than as applied empirical estimates.

Key Points

• Measurement gap: existing text/term‑frequency AI/digitalisation indices do not separate autonomous execution from decision support and lack psychometric validation.
• Conceptual definition: AAII = extent to which a bank has delegated consequential financial decision and execution authority to autonomous systems (authority, irreversibility and delegated mandate are central).
• Theoretical anchors:
  • Dynamic Capabilities Theory — AAII signals a firm’s capacity to sense, seize and transform (governance/maturity moderate outcomes).
  • Technology–Organisation–Environment (TOE) — technological, organisational and environmental context (e.g., regulation, fintech pressure, CBDC exposure) condition effects.
• AAII is intended to enable joint evaluation of productivity gains and priced risk (efficiency vs systemic/operational risk trade‑offs), and to support tests of nonlinearity and conditional effects.

Principal methodological design elements - Five‑step AAII construction protocol: 1. Theory‑driven seed keyword pool (appendix lists seeds). 2. Two‑round expert Delphi refinement to finalise items. 3. Inter‑rater reliability assessment with prespecified acceptance thresholds and a coding manual (excerpted in appendix). 4. Exploratory and confirmatory factor analysis to establish dimensional structure. 5. LLM‑assisted document classification with mandatory human adjudication; authority‑weighted scoring rule separates delegated execution language from advisory language; normalization isolates agentic intensity net of general AI‑discussion volume. - Construct validity: convergent, discriminant, criterion and known‑groups tests against independent deployment indicators (external benchmarks). - Threats and mitigations: - Reverse causality (richer banks adopt more AI): addressed via multiple, overlapping identification designs. - AI‑washing and strategic disclosure: explicit detection and robustness checks; human validation of model labels emphasised. - Vendor concentration, explainability and adversarial manipulation of text signals are acknowledged and logged as limitations/controls. - Reproducibility: full protocol, coding manual excerpts, seed keywords and supplementary tables provided; acceptance thresholds specified at each validation step (protocol oriented).

Data & Methods

• Source documents: annual reports, Management Discussion & Analysis (MD&A), and earnings‑call transcripts for banks (panel design described; sample design and variable definitions in appendices).
• Index construction:
  • Keyword generation grounded in theory (Dynamic Capabilities + TOE) and refined via Delphi with domain experts.
  • Human coders perform reliability checks (inter‑rater metrics and thresholds reported in the protocol).
  • Factor analysis (EFA/CFA) to confirm dimensionality and to justify authority weights.
  • Classification pipeline: LLMs used to scale text classification; every LLM classification is subject to human adjudication (protocol includes rules for disagreement resolution).
  • Scoring: authority‑weighted rule assigns higher weight to language indicating delegated authority/execution; normalization controls for total AI discussion volume.
• Validation strategy:
  • Construct validity: convergent (correlation with independent deployment signals), discriminant (distinguish from generic digitalisation indices), criterion (predictive relationships for outcomes), and known‑groups (e.g., early adopters versus non‑adopters).
• Identification strategies for causal inference (detailed, overlapping designs to triangulate effects):
  • Two‑way fixed effects and event‑study setups.
  • Staggered‑adoption difference‑in‑differences (with careful handling of heterogeneous timing).
  • Instrumental variables (instruments and exclusion logic specified).
  • System GMM for dynamic panels.
  • Propensity‑score matching and synthetic control for case studies / single‑treatment events.
  • Robustness: placebo tests, falsification, handling of anticipatory disclosure and measurement error.

Implications for AI Economics

• Provides a validated, replicable bank‑level measure that distinguishes agentic (delegated execution) AI from advisory AI — enabling more precise empirical work on automation, productivity and risk.
• Enables rigorous causal testing of conflicting hypotheses in the literature (e.g., productivity gains vs increased operational/systemic risk; nonlinearity and heterogeneous effects by governance/regulatory context).
• Facilitates joint assessment of market pricing, lending outcomes (including fairness/disparate impact), operational‑risk disclosures and systemic vendor concentration, by providing a continuous, theory‑grounded exposure variable.
• Encourages best practice for text‑based measurement: transparent item generation, expert refinement, psychometric validation, and mandatory human oversight of model classifications to reduce AI‑washing and adversarial manipulation.
• Policymaking and supervision: gives regulators an operational instrument (and a prespecified identification plan) to evaluate agentic AI deployments, test containment measures, and monitor correlated model/vendor risks across the banking sector.
• Research agenda opened: heterogeneous effects across functions/workflows, interactions with governance quality and regulatory regimes, macro‑financial amplification channels from correlated agentic adoption, and evaluation of mitigation policies (e.g., vendor diversification, approval thresholds for delegated authority).

Notes - The article is a methods/protocol paper: it specifies the index and causal designs but does not report AAII application estimates. Supplementary files include Tables 1–4, seed keywords and coding manual excerpts.