Helping people identify and pursue personally meaningful career goals at scale remains a key challenge in applied psychology. Career coaching can improve goal quality and attainment, but its cost and limited availability restrict access. Large language model (LLM)-based chatbots offer a scalable alternative, yet the psychological mechanisms by which they might support goal pursuit remain untested. Here we report a preregistered three-arm randomised controlled trial (N = 517) comparing an AI career coach ("Leon," powered by Claude Sonnet), a matched structured written questionnaire covering closely matched reflective topics, and a no-support control on goal progress at a two-week follow-up. The AI chatbot produced significantly higher goal progress than the control (d = 0.33, p = .016). Compared with the written-reflection condition, the AI did not significantly improve overall goal progress, but it increased perceived social accountability. In the preregistered mediation model, perceived accountability mediated the AI-over-questionnaire effect on goal progress (indirect effect = 0.15, 95% CI [0.04, 0.31]), whereas self-concordance did not. These findings suggest that AI-assisted goal setting can improve short-term goal progress, and that its clearest added value over structured self-reflection lies in increasing felt accountability.
Figures & tables
Figure 1: Screenshots of the study application across the three conditions. G1 (left): the final goal-entry screen, where participants typed their goals directly. G2 (centre): an example structured reflection prompt (“Priorities & Impact”), one of five written questions completed before goal entry. G3 (right): a mid-conversation exchange with “Leon,” the AI career coach, probing the participant’s work situation before collaborative goal formulation.
Variable
M
SD
1
2
3
4
5
1. Goal Progress (T2)
3.27
1.33
—
2. Accountability (T1)
3.51
1.27
.19∗
—
3. Self-Concordance (T1)
1.91
2.45
−.07
−.01
—
4. NPS (T1)
5.68
2.65
.30∗
.50∗
.17∗
—
5. Goal Specificity
1.26
0.96
.14∗
.09
−.04
.18∗
—
Table 1: Descriptive statistics and intercorrelations for primary study variables ( N=323 ; goal specificity n=292 ).
Figure 2: Primary outcome and main mediator by condition. Panel A shows goal progress at Time 2; Panel B shows perceived accountability at Time 1. Bars show condition means; error bars are 95% confidence intervals. Key preregistered contrasts are annotated. n per condition is shown within each panel.
Figure 3: Preregistered two-mediator parallel mediation model for the AI Chatbot versus Questionnaire contrast. Green paths are statistically significant (95% CI excludes zero, 5,000 bootstraps); grey dashed paths are non-significant. Path coefficients shown are unstandardised. The indirect effect through Accountability ( ab=0.15 , 95% CI [0.04,0.31] ) indicates mediation of the AI-versus-Questionnaire contrast; the indirect effect through Self-Concordance was negligible ( ab=−0.002 , 95% CI [−0.06,0.04] , ns).
Figure A1: Participant flow. N=517 randomised; 89 excluded at Time 1 (84 did not complete the session, 5 failed the attention check), leaving n=428 who completed Time 1. The high T1 dropout in the Questionnaire condition ( n=53 , all non-completions) reflects the platform-enforced 20-minute minimum engagement period. At Time 2, 93 did not return and 12 failed the attention check, yielding an analytic sample of n=323 (62.5% of randomised).
Figure A2: Exploratory parallel mediation model for the AI Chatbot versus Questionnaire contrast. Green paths: accountability; orange paths: goal specificity; grey dashed: self-concordance. Indirect effects: accountability ab=0.14 , 95% CI [0.03,0.29] ; goal specificity ab=0.27 , 95% CI [0.02,0.52] . All based on 5,000 bootstraps, n=211 .
Variable
n
M
SD
Goal Progress (T2)
323
3.27
1.33
Accountability (T1)
323
3.51
1.27
Accountability (T2)
323
3.13
1.51
Self-Concordance (T1)
323
1.91
2.45
Self-Concordance (T2)
323
2.05
3.56
Goal Specificity
292
1.26
0.96
Table A1: Descriptive statistics for all study measures. N=323 except goal specificity ( n=292 ). Self-concordance is the standard composite of intrinsic + identified − introjected − external motivation (range −8 to +8 ). NPS = Net Promoter Score (0–10). Goal specificity = exploratory LLM-coded time-frame + measurability (0–4).
Variable
1
2
3
4
5
6
1. Goal Progress (T2)
2. Accountability (T1)
.19∗
3. Accountability (T2)
.29∗
.50∗
4. Self-Concordance (T1)
−.07
−.01
−.10
5. Self-Concordance (T2)
.03
−.02
−.15∗
.74∗
6. Goal Specificity
.14∗
.09
.13∗
−.04
−.06
Table A2: Intercorrelations among focal outcome and mechanism variables. N=323 except goal specificity ( n=292 ). ∗p<.05 .
b
SE
t
p
Panel A: Paths from condition (X) to mediators
a1
X → Self-Concordance
0.122
0.339
0.36
.720
a2
X → Accountability
0.509∗
0.161
3.17
.002
Panel B: Paths from mediators to outcome (controlling for X)
b1
Self-Concordance → Goal Progress
−0.016
0.036
−0.44
.659
b2
Accountability → Goal Progress
0.290∗
0.076
3.81
<.001
Table A3: Path coefficients for the preregistered two-mediator parallel mediation model (H4c: AI Chatbot vs. Questionnaire; n=218 ). X = AI Chatbot condition dummy (1 = AI, 0 = Questionnaire). Mediators: M 1 = Self-Concordance (T1); M 2 = Accountability (T1). Outcome: Goal Progress (T2). All paths estimated via ordinary least squares. Indirect effects from 5,000 bootstrapped samples (seed = 42). ∗p<.05 .
Figure A3: Descriptive goal specificity by condition (exploratory LLM coding). Left: total specificity score (0–4); centre: time-frame subscale (0–2); right: measurability subscale (0–2). Bars show participant-level means averaged across each participant’s coded goals; error bars are 95% confidence intervals. Pairwise Cohen’s d values for AI versus Control and AI versus Questionnaire are annotated descriptively.
Domain
G1 Control
G2 Questionnaire
G3 AI
Career
211 (86.8%)
279 (89.4%)
191 (59.7%)
Non-career (total)
32 (13.2%)
33 (10.6%)
129 (40.3%)
Fitness & health
3 (1.2%)
3 (1.0%)
48 (15.0%)
Personal finance
5 (2.1%)
4 (1.3%)
28 (8.8%)
Personal development
8 (3.3%)
15 (4.8%)
15 (4.7%)
Leisure
4 (1.6%)
3 (1.0%)
17 (5.3%)
Table A4: Goal domain classification by condition (LLM-coded, claude-sonnet-4-6 , blind to condition). Each participant contributed up to three goals. Percentages are of total classified goals per condition. The non-career subtotal row corresponds to the aggregate used in the main text ( χ2(2)=96.75 , p<.001 ; G3 vs. G1: h=0.63 ; G3 vs. G2: h=0.71 ).
Code
G1 Control
G2 Questionnaire
G3 AI
Positive helpfulness
49.0%
52.3%
61.0%
Neutral helpfulness
26.0%
28.8%
24.8%
Negative helpfulness
25.0%
18.9%
14.3%
Clarity / focus
45.2%
55.9%
59.0%
Reflection
25.0%
24.3%
21.9%
Motivation
7.7%
18.0%
18.1%
Table A5: Qualitative feedback coding summary by condition (exploratory; claude-sonnet-4-6 , blind to condition). Values are percentages of scorable responses within each condition. Benefit, challenge, and request rows are non-exclusive.
Group
n
T1 SC M (SD)
T2 SC M (SD)
G1 Control (overall)
105
1.76 (2.35)
1.70 (3.60)
G1 — non-career ( ≥ 1)
22
1.09 (2.65)
0.71 (4.00)
G1 — career-only
83
1.94 (2.24)
1.96 (3.47)
G2 Questionnaire (overall)
111
1.92 (2.60)
2.25 (3.61)
G2 — non-career ( ≥ 1)
24
2.79 (2.56)
3.17 (3.75)
G2 — career-only
87
1.68 (2.57)
1.99 (3.55)
Table A6: Self-concordance (SC) by condition and goal domain (exploratory). Subgroups are defined by whether participants surfaced at least one non-career goal. SC range: −8 to +8 ; higher scores indicate stronger autonomous motivation. The within-condition non-career versus career-only difference was clearest in G3 (T1: d=0.44 , p=.033 ; T2: d=0.53 , p=.011 ), though smaller differences were also present in G1 and G2. T1 and T2 used the same SC items but different response scales (5-point at T1; 7-point at T2), so change scores should be interpreted cautiously.