Complementarities of AI-Enabled and Human Recommendation to Jobseekers: Experimental Evidence from Kenya

Participants are offered structured career guidance, delivered through one of three channels depending on random assignment: a trained near-peer human mentor, an AI-enabled career-guidance tool (Compass), or a combination of the two. A control group receives no additional personalized guidance during the study period.

Human mentors are drawn from Swahilipot Hub and the National Council of Churches of Kenya, and are recruited from the same coastal communities as participants. They provide one-on-one counselling: eliciting skills and preferences, discussing options, recommending a path, and supporting the jobseeker to act on it. Compass (developed by Tabiya) elicits a jobseeker's skills and preferences through a structured conversation, translates informal experience into a structured skill profile, and recommends jobs and career paths matched both to the jobseeker and to local labour demand.

Within the AI-supported arms, the design also varies whether AI support stops once a recommendation has been delivered (information only) or continues with persuasion and action support (planning, encouragement, and follow-through). This separates the informational content of guidance from its persuasive, action-shifting content.

2026-06-22

2026-08-14

Primary outcomes are pre-specified in families defined by domain and by measurement wave, across two waves.
(Yet to be refined)

Short-run primary (midline, ~4 weeks):

1. Information uptake and decision quality: preference clarity and belief accuracy (pre/post preference elicitation; calibration of subjective probabilities; choice uncertainty / second-order beliefs), consideration-set breadth and calibration, dominated-option avoidance, and search targeting (a search-focus / Herfindahl measure over occupations considered and applied to).
2. Job search and advice take-up: job-search effort and application yield, active vs. passive search (on/off-path search-help decision), information-bundle choice, implementation of the recommended path and action-plan completion, and the subjective probability of pursuing the recommended occupation.

Longer-run primary (endline, ~6-9 months):
3. Economic opportunity: employment or income-generating activity (incl. a count of IGAs), earnings, hours worked, number of income sources, job applications, interviews, offers, and participation in training.
4. Match quality and persistence: skill-occupation alignment (KeSCO 2- and 3-digit), targeting of search toward a coherent occupation or sector, occupational/sectoral persistence over time, tenure and work-history length, and work satisfaction.
5. Welfare and subjective well-being: consumption, savings, perceived agency, and life satisfaction.

Long-run targets: occupation/industry persistence (2-digit ISIC) and total real earnings ~10 years out, via a pre-specified surrogate index (Kenya Life Panel Survey).

The two short-run families map to the two channels the design separates: information (uptake, decision quality, search focus) and persuasion (take-up, effort). They are primary both because the intervention targets them most directly (proximal behavioural responses) and because they identify the mechanism (RQ2). The endline families test whether these proximal effects translate into distal labour-market and welfare gains.

Each family is summarized by a standardized index (Anderson 2008), reducing each family to one primary test; across the family indices within a wave we report sharpened FDR q-values, and within each family disaggregated components use Romano-Wolf corrections.

Corrections are applied within family and within measurement wave. The short-run (midline, ~4 weeks) and longer-run (endline, ~6-9 months) primary sets are pre-specified as distinct outcome sets: measured at different times, addressing distinct questions (proximal behavioural mechanism vs. distal labour-market welfare). We therefore correct within each set and do not apply a single joint correction across the midline and endline primary outcomes. This split is registered in advance, not chosen after seeing results.

The surrogate index maps short-run proxies to the long-run targets (Athey-Chetty-Imbens-Kang 2024), under unconfoundedness, Prentice surrogacy, and cross-cohort comparability.

(Yet to be refined) Additional midline mechanism measures, each its own family: incentivized revealed preferences and WTP (counselling-format WTP; stepping-stone vignette; stated usefulness and a recommendation-regret index)
; beliefs and expectations, extended (top-option subjective probabilities; pre/post preference-shift magnitude); decision-process robustness (stated-chosen consistency; choice stability under re-elicitation); time use and substitution (displaced leisure); agency and discouragement (locus of control, self-efficacy, fatigue, aspirations, controllability slope).

Secondary treatment estimands: CACE; recommendation-vs-persuasion contrasts (T1-P vs. T1-R; T2-P vs. T2-R); "any-AI vs. control" and "any-human vs. control" pooled policy contrasts; counsellor-side mechanism outcomes.

Secondary families are corrected within themselves and not pooled with the primary families or across waves. The recommendation-vs-persuasion contrasts test whether guidance works by informing (effects in the short-run information family and in endline match quality/persistence) or by persuading (short-run take-up and effort rise while regret rises and persistence falls). The "any-AI"/"any-human" pooled-vs-control contrasts are secondary policy quantities; the clean estimands are the factorial main effects and interaction.

Individually randomized controlled trial with approximately 4,000 jobseekers. Six operational arms pool into four primary categories forming a 2×2 factorial (AI access × human counselling): Control, Human-only, AI-only, and AI+Human. Within the AI-supported arms, a secondary cross-randomization varies whether AI support stops at a recommendation (information only) or adds persuasion/action support. Primary estimands are intention-to-treat effects of the four pooled categories together with the factorial main effects (of AI and of human counselling) and their interaction (the substitutes-vs-complements test). The timing of AI introduction is separately randomized at the mentor level to diagnose counsellor learning and spillovers.

Not available

Randomization is carried out by a reproducible, seeded computer script, run after baseline data collection. The procedure combines blocking with rerandomization for covariate balance (Morgan and Rubin 2012, 2015) and is implemented batch-by-batch as recruitment proceeds, conditioning each batch on all previously randomized participants via cumulative-deficit apportionment of arm totals. The script blocks on geography (county cell), gender, and baseline labour-market state, then rerandomizes on pre-specified baseline predictors of outcomes and attrition, accepting an assignment only if it passes a calibrated, tiered Mahalanobis balance criterion. Confirmatory inference replays this exact mechanism (randomization inference).

The individual jobseeker. A separate, secondary randomization is conducted at the mentor level (timing of AI introduction: early vs. late, 1:1), used only to diagnose counsellor learning and spillovers, not to define the primary treatment effects.

No

Individual randomization (not clustered): approximately 4,000 individuals. Guidance in the human-contact arms is delivered by approximately 80 near-peer mentors; inference clusters at the counsellor × session level. The secondary mentor-level randomization (AI-introduction timing) involves approximately 80 mentor clusters.

Approximately 4,000 jobseekers.

Approximately 4,000 jobseekers, assigned across six operational arms in a 2:1:1:2:1:1 ratio:
- T0 Control: 1,000
- T1-R (AI-only, recommendation): 500
- T1-P (AI-only, recommendation + persuasion): 500
- T2-R (AI+Human, recommendation): 500
- T2-P (AI+Human, recommendation + persuasion): 500
- T3 Human-only: 1,000

Pooled into four primary categories of ~1,000 each: Control; Human-only (T3); AI-only (T1-R + T1-P); AI+Human (T2-R + T2-P).

Assuming 10% attrition, covariate adjustment, and an intra-cluster correlation of ~0.05 at the counsellor × session level (α = 0.05, power 0.80): pooled active-vs-control comparisons have a minimum detectable effect of approximately 0.13–0.16 SD; the factorial main effects of AI and of human counselling are more precise, ~0.10–0.12 SD (each compares ~2,000 vs. ~2,000); the AI×human interaction is powered for moderate departures from additivity. Effects are in standard-deviation units of each standardized (Anderson) outcome index.