AI, Human, and Hybrid Career Guidance in Kenya: Does Advice Inform or Activate?

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 behavioral activation and action support (planning, encouragement, and follow-through). This separates the informational content of guidance from its behavioral activation content. The activation session is delivered by the AI directly to the jobseeker, in a conversation separate from any human counseling and never routed through the counselor. Counselors receive identical material for both subarms and are blind to which of their mentees receive the activation session.

2026-06-22

2026-09-18

Primary Outcomes (end points) All primary outcomes are measured at the endline (~6-9 months) to align with what the design can support. They consist of three pre-specified objects:
1. Labor-market index: Equal-weighted average of paid employment (income-generating work in the last seven days) and monthly earnings.
2. Quality-adjusted placement: The utility of the respondent's main occupation evaluated at their own baseline preferences.
3. Recommendation-directed action index: Averages applications in the target career family, fields searched, and training discipline (the primary outcome for the activation contrast).

Long-run targets: Four-year job-quality index, log labor earnings, and longest continuous employment spell, projected via a pre-specified surrogate index using the Kenya Life Panel Survey (KLPS).

Detailed construction and contrasts can be found in the uploaded PAP.

Each family is summarized by an equal-weighted standardized index (Kling, Liebman, Katz 2007). Inverse-covariance weighting (Anderson 2008) is used as a robustness check. The registered index carries the within-family correction using Romano-Wolf step-down, with remaining disaggregated members reported beneath it unadjusted. Sharpened FDR q-values are reported alongside the Romano-Wolf p-values.

Corrections are applied within family and not across families. Midline outcomes (~6 weeks) are secondary and descriptive. The surrogate index maps endline proxies to the long-run four-year targets (Athey-Chetty-Imbens-Kang 2024).

Secondary Outcomes (end points) Midline mechanism measures (~4-6 weeks): Career-direction index, advice take-up index, beliefs, agency, intentions, and delivery fidelity.

Endline secondary outcomes: Allocation quality (misallocation gap, skill-occupation alignment), labor market search and training (job-search index), psychological well-being index, consumption, savings, and life satisfaction. Unsigned measures (hours, work multiplicity, reservation wage, activity type) are descriptive and enter no correction family. I also test a budget-equivalence criterion for cost-effectiveness (AI vs. Human).

Secondary treatment estimands: Within-channel activation contrasts (T1-A vs. T1-R, T2-A vs. T2-R). The AI-by-human interaction (factorial superadditivity) is exploratory.

Detailed construction and contrasts can be found in the uploaded PAP.

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 2x2 factorial in AI access (Compass on or off) and human counseling (on or off): Control, Human-only, AI-only, and AI-and-Human. Within the two AI-receiving categories, a nested randomization varies whether AI support stops at a recommendation (recommendation only) or adds a behavioral-activation session. Primary estimands are intention-to-treat effects of the pooled active guidance offer vs. control, the factorial main effects of AI and human counseling, and the pooled activation contrast (recommendation-plus-activation vs. recommendation-only). The AI-by-human interaction (the substitutes-vs-complements test) is exploratory. The timing of AI introduction is separately randomized at the mentor level to diagnose counsellor learning and spillovers.

Not available

Randomization was carried out by a reproducible, seeded computer script, run after baseline data collection, and is now complete across five recruitment batches. The procedure combines blocking with rerandomization for covariate balance (Morgan and Rubin 2012, 2015) and was implemented batch by batch as recruitment proceeded, conditioning each batch on all previously randomized participants through cumulative-deficit apportionment of arm totals.
The randomization is hierarchical. The first level assigns jobseekers in equal proportion (1:1:1:1) across the four primary categories: Control, AI-only, AI-and-Human, and Human-only. The second level applies a nested 1:1 split between recommendation and activation inside the two AI-receiving categories only. Collapsing the two levels yields six atomic cells in the marginal ratio 2:1:1:2:1:1.
The script blocks on geography, 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, varying the timing of AI introduction (early versus late, 1:1). It is used only to diagnose counselor learning and implementation spillovers and does not define any primary treatment effect.

No

Individual randomization, not clustered: approximately 4,000 individuals. Guidance in the human-contact arms is delivered by roughly 80 to 100 near-peer mentors. Because randomization is individual, no design-based clustering enters the power calculations. Model-based standard errors are clustered on the assigned counselor, with wild cluster bootstrap p-values, as a secondary diagnostic alongside the primary randomization-inference statements. The secondary mentor-level randomization of AI-introduction timing involves those same 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-A (AI-only, recommendation + behavioral activation): 500
- T2-R (AI and Human, recommendation): 500
- T2-A (AI and Human, recommendation + behavioral activation): 500
- T3 Human-only: 1,000

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

Assuming 10% attrition and covariate adjustment, with individual-level randomization (no design-based clustering) ($\alpha = 0.05$, power 0.80): the pooled guidance offer vs. control has a minimum detectable effect (MDE) of 0.096 SD (one-sided); the factorial main effects of AI and human counseling are more precise at 0.084 SD (one-sided); the pooled activation contrast is detectable at 0.118 SD (two-sided). The AI×human interaction is exploratory (MDE 0.167 SD). Effects are in standard-deviation units of each equal-weighted standardized outcome index (Kling, Liebman, Katz 2007).