Background:

Task-shifting—the delivery of evidence-based mental health interventions by trained lay providers—has shown promise in closing the treatment gap in low- and middle-income countries. But the effectiveness of task-shifted interventions depends critically on ongoing supervision and monitoring, yet traditional supervision models are difficult to scale. Artificial intelligence (AI) tools capable of automatically processing session recordings and generating structured fidelity feedback for supervisors could offer a scalable alternative, but no such system has been developed or validated for lay-delivered interventions in multilingual, low-resource settings.

Objective:

We developed and pilot-validated shamiriAI, an automated fidelity monitoring tool for lay-delivered mental health interventions, embedded within the Shamiri school-based mental health program in Kenya.

Methods:

We conducted a pilot validation study across six secondary schools in Ngong Hub, Kajiado County, Kenya (May–September 2025). shamiriAI follows a five-stage pipeline: audio ingestion and preprocessing, multilingual automatic speech recognition (ASR) with prosodic feature extraction, personal identifiable information (PII) scrubbing, large language model (LLM)-based fidelity inference, and structured feedback report delivery to supervisors. We pursued two pilot aims: (1) ASR performance on a held-out test set of manually transcribed sessions; and (2) interrater reliability between shamiriAI-generated fidelity ratings and independent human supervisor ratings across 52 recorded sessions, spanning six domains (Required Contents, Specifics, Thoroughness, Clarity, Skill, Purity) rated on a 1–7 scale. Reliability was assessed using intraclass correlation coefficients (ICC), Bland-Altman analysis, adjacent agreement rates, paired-sample t-tests with Holm–Bonferroni correction, and Gwet's AC2.

Results:

The ASR model achieved a Character Error Rate of 0.19, Word Error Rate of 0.34, and cosine semantic similarity of 0.77, indicating strong meaning preservation despite surface-level transcription errors in code-switched speech. On fidelity ratings, AI scores were systematically lower than the human composite overall (M = 5.14, SD = 0.77 vs. M = 5.93, SD = 0.57; Δ = −0.79, 95% CI [−1.04, −0.53], d = −1.16, p < .001). Reliability varied markedly by dimension: ICCs ranged from −0.06 to 0.20 across all six domains. Three distinct patterns emerged: large systematic underrating on holistic interpretive dimensions (Required Contents d = −3.48; Clarity d = −1.56); a bidirectional medium-effect pattern on facilitation dimensions (Thoroughness d = −0.99; Skill d = +0.87); and no significant bias on structured detection dimensions (Specifics 78.8% adjacent agreement; Purity 73.1% adjacent agreement), where performance approached the human–human AC2 benchmark of 0.42–0.60.

Conclusions:

shamiriAI demonstrates technically feasible multilingual ASR and a coherent, interpretable reliability profile. Underperformance was concentrated on holistic inferential dimensions — particularly Required Contents and Clarity — while structured detection tasks already approach operational utility. The underperformance pattern reflects diagnosable misalignments in rubric interpretation and prompt design, with clear engineering solutions identified. These findings provide the foundational validation evidence and dimension-specific diagnostics needed to guide the development of AI-augmented supervision for lay-delivered adolescent mental health programs in sub-Saharan Africa and other multilingual settings.