This study is a secondary analysis of data collected during a real-world implementation of a group PMT program enhanced by an AI CA called “Pat.” The dataset has previously been examined in a separate secondary analysis focusing on qualitative parent experiences [19]. The current manuscript analyzes the quantitative components of the same implementation dataset, examining indicators of feasibility and acceptability, including parent engagement, satisfaction, and perceived therapeutic alliance with Pat.

As previously described in Rivera-Cepeda et al [19], the participants consisted of 88 primary caregivers, the majority of whom were female (n=75, 86%). Caregivers ranged in age from 34 to 65 years (mean 44.31, SD 5.73) and had a child aged between 3 and 14 years (mean 7.98, SD 2.45). Most children were male (n=70, 80%), while 20% (n=18) were female. Of the 88 caregivers, 74 (84%) identified as White or of European descent, 9 (10%) identified as another race or ethnicity, and 3 (3%) identified as Indigenous or Native people. An additional 3 (3%) caregivers did not report their race or ethnicity, as this information was optional. In terms of relationship status, over half of caregivers were married (n=49, 56%), 17% (n=15) were living with a partner, 9.2% (n=8) were single, 9.2% (n=8) were divorced, and 8.0% (n=7) were separated. Educational attainment was generally high: 42% (n=37) had completed a postgraduate degree (eg, specialization, master’s, or doctorate), 38% (n=33) held a university degree, 15% (n=13) completed technical or nonuniversity training, 4.5% (n=4) completed 12 years of schooling, and 1.1% (n=1) reported fewer than 7 years of formal education [19].

Caregivers in Argentina and Paraguay were recruited through online outreach conducted by licensed mental health professionals in each country. Recruitment occurred on a rolling basis from January 2025 to November 2025 and targeted caregivers of children aged 3 to 12 years with disruptive behavior concerns. Interested caregivers completed online registration and were enrolled in the ParenteAI program.

As described in Rivera-Cepeda et al [19], the program included access to the ParenteAI digital platform and participation in a live, therapist-led parenting group delivered via videoconference. Study measures were administered at midintervention (wk 4) and postintervention (wk 8 or wk 12).

The intervention followed the hybrid PMT model integrating therapist-led group sessions with the ParenteAI chatbot (“Pat”), as detailed in Rivera-Cepeda et al [19]. Caregivers attended live sessions and engaged with Pat between meetings to support skill acquisition and implementation. Live sessions focused on core PMT principles (eg, reinforcement strategies, effective instruction, praise, and consequences), while Pat provided structured modules and personalized in-the-moment support. Caregivers interacted with Pat through conversational exchanges that included psychoeducation, skill practice, role-play activities, personalized reminders, responses to parenting questions, and support for applying PMT strategies to real-life situations between sessions. Therapists reviewed caregiver engagement with Pat, provided feedback, and assigned weekly modules.

Cohorts 1 and 2 were delivered over a 12-week period and included 8 required modules. Subsequent cohorts were condensed to an 8-week format. While the number of required modules remained consistent across all cohorts, later cohorts were offered 9 additional optional PMT modules reflecting iterative refinements to Pat. Module completion was scheduled according to consistent group timelines across cohorts. Although the core intervention content remained consistent, changes in program duration and the availability of optional modules led to a more condensed delivery and greater opportunities for additional engagement.

As described in Rivera-Cepeda et al [19], the University of Chile Institutional Review Board approved this study (approval number: 23136643/2023). This study was a secondary data analysis of deidentified data. Caregivers consented to the use of anonymized data at enrollment through the ParenteAI Terms of Service. Further consent was obtained when participants completed the surveys as part of their participation in the hybrid group PMT. Participants could leave the study at any time.

Engagement, NPS, WAI, and TAI descriptives were reported using descriptive analysis. Descriptive statistics were calculated to characterize user engagement across weeks 4, 8, and 12, including total messages exchanged, Q&A interactions, modules completed, and days of platform use. Additionally, ANOVAs were run for NPS, WAI, and TAI. For the NPS, there were a total of 117 observations analyzed, corresponding to 88 unique participants. Of these, 59 participants contributed 1 measurement and 29 contributed 2 measurements; no participant contributed more than 2 measurements. NPS categories (promoters, passives, and detractors) were calculated at the observation level, using all scores provided rather than aggregating by participant. A linear mixed-effects model was used to examine differences in NPS across weeks 4, 8, and 12, accounting for partially repeated measurements. Measurement week was included as a fixed effect, and participant identifier was included as a random intercept. The linear mixed-effects model was selected because some participants contributed data at multiple time points, resulting in partially repeated measurements. This approach retained all available NPS scores while accounting for within-participant clustering. To facilitate interpretation and comparison across data points, descriptive statistics and exploratory ANOVAs were also conducted. Although some participants contributed data at multiple time points, week was treated as a between-subjects factor for these exploratory analyses due to the limited overlap of participants across data points.

Rivera-Cepeda et al [19] reported an excellent NPS score at the end of the treatment (76.92) but did not report the NPS across weeks. For week 4, based on 54 respondents, the NPS was 77.7. For week 8, based on 43 respondents, the NPS was 76.8, and for week 12, based on 20 respondents, the NPS was 75. Results from the linear mixed-effects model indicated no significant effect of week on NPS (F_{2, 38.67}=1.66; P=.20). Estimated marginal means of NPS remained high across all assessment points (wk 4: mean 9.13, 95% CI 8.81-9.46; week 8: mean 9.34, 95% CI 9.00-9.68; and wk 12: mean 9.43, 95% CI 8.96-9.91), showing a descriptive upward trend, although differences between time points were not statistically significant. Descriptive statistics indicated consistently high NPS ratings across all time points. Mean NPS scores were 9.33 (SD 1.13; n=54) at week 4, 9.35 (SD 1.04; n=43) at week 8, and 9.20 (SD 2.12; n=20) at week 12. The overall mean NPS across the sample was 9.32 (SD 1.31; n=117), with scores ranging from 2 to 10 (Table 2). The Levene test for equality of variances was nonsignificant (F_{2, 114}=2.59; P=.09) indicating that the assumption of homogeneity of variance was met. The overall ANOVA was not significant (F_{2, 114}=0.10; P=.90; η²=0.002) suggesting that NPS did not differ across time points. The week accounted for less than 1% of the variance in NPS scores. CIs around the effect size estimate were small and included zero, further supporting the absence of meaningful differences over time. These findings indicate that participants reported consistently high levels of likelihood to recommend the program across weeks 4, 8, and 12, with no evidence of change over time.

Mean scores for TAI across weeks were week 4 (mean 4.22, SD 0.47), week 8 (mean 4.49, SD 0.37), and week 12 (mean 4.49, SD 0.48). Regarding the ANOVA comparing TAI scores across weeks, although some participants contributed data at multiple time points, week was treated as a between-subjects factor for these exploratory analyses. The Levene test was nonsignificant (F_{2, 114}=0.80; P=.45) indicating that the assumption of homogeneity of variance was met. The overall ANOVA was significant (F_{2, 114}=5.69; P=.004; η²_p=0.09) suggesting that TAI scores differed across time points, with week accounting for approximately 9% of the variance in therapy attitudes. Post hoc comparisons using Tukey’s Honestly Significant Difference indicated that TAI scores at week 4 were significantly lower than at week 8 (P=.008) and week 12 (P=.047), whereas weeks 8 and 12 did not differ (P>.99). Overall, these findings indicate that therapy attitudes increased between weeks 4 and 8 and were maintained through week 12.

At week 8, the items with the highest frequencies of positive responses were #10 (“My general feeling about the program I participated in is”), #9 (“I feel the type of program that was used to help me improve the behaviors of my child was”), and #7 (“Regarding the progress my child has made in his or her general behavior, I am”; Figure 1).

For the ANOVA comparing WAI scores across the weeks, Levene’s tests indicated that the assumption of homogeneity of variance was met for all analyses. Mean total scores were similar at week 4 (mean 4.10, SD 0.63), week 8 (mean 4.10, SD 0.55), and week 12 (mean 4.04, SD 0.63), and the ANOVA was not significant (F_{2, 114}=0.182; P=.83). Likewise, no significant differences were found for the Goal subscale (wk 4: mean 4.09, SD 0.92; wk 8: mean 4.19, SD 0.66; wk 12: mean 4.13, SD 0.70; F_{2, 114}=0.182; P=.84), the Bond subscale (wk 4: mean 4.03, SD 1.07; wk 8: mean 3.92, SD 1.06; wk 12: mean 4.06, SD 1.01; F_{2, 114}=0.162; P=.85), and the Task subscale (wk 4: mean 3.93, SD 0.74; wk 8: mean 4.06, SD 0.52; wk 12: mean 3.94, SD 0.70; F_{2, 114}=0.570; P=.57; Figure 2).

Orange is week 8 overall, yellow is week 8 agreement with goals, blue is positive bond, and green is agreement with tasks.

This study examined caregiver engagement, satisfaction, and therapeutic alliance in a hybrid PMT group program supported by Pat. Consistent with the study aims, caregivers demonstrated high levels of engagement with Pat, reported strong satisfaction across multiple time points, reported progress in their major challenges, and formed a strong therapeutic alliance with Pat. The findings support the feasibility and acceptability of integrating an AI CA into a therapist-led PMT program. These findings should be interpreted in the context of a descriptive study design and are intended to inform the feasibility and acceptability of this hybrid intervention model. Caregivers showed substantial engagement with Pat across the intervention. By week 4, participants had already exchanged an average of 383.09 (SD 185.64) messages with the platform, and message volume was even higher among those assessed later in treatment, with averages of 605.00 (SD 268.61) messages in the 12-week cohorts (who had fewer modules) and 726.47 (SD 405.41) messages in the 8-week cohorts (who had more modules available). This pattern suggests that caregivers interacted with Pat extensively throughout the program. Engagement was also reflected in module completion. On average, caregivers completed 5.41 (SD 1.93) modules by week 4 and 8.67 (SD 3.39) by week 8, indicating that most participants progressed through a substantial portion of the intervention content. Additionally, the optional Q&A use showed a similar pattern, rising from 5.17 (SD 4.75) interactions at week 4 to 10.49 (SD 9.18) at week 8, suggesting that parents completed at least 1 module and 1 in-the-moment support chat per week. Additionally, engagement with Pat was frequent, with an average of 10.00 (SD 5.15) days at week 4, 17.42 (SD 9.26) days at week 8, and 18.90 (SD 9.53) days at week 12. Taken together, these findings suggest that caregivers engaged with structured learning modules and returned to Pat for in-the-moment support to apply parenting skills in real-life situations.

The combination of high message volume, substantial module completion, and repeated Q&A use is consistent with, and notably higher compared with, the existing literature on CAs for parenting support. Engagement in digital interventions is often reflected in sustained participation over time, repeated interactions, and continued use beyond the required intervention activities. In this study, caregivers completed the assigned modules as well as initiated optional in-the-moment support chats, reflecting an engagement level beyond the minimum program requirement. More specifically, to our knowledge, the number of messages exchanged at week 8 (mean 726.47, SD 405.41 messages) is the highest reported in any AI study on mental health. The most comparable study using generative AI for parenting was by Rivera-Cepeda et al [18], which found that four sessions combining therapist support and Pat resulted in an average of 376 messages exchanged. Other parenting studies have generally used earlier conversational technologies and/or shorter interventions and have reported substantially lower levels of interaction [14,16]. A study by Heinz et al [23] using generative AI for depression, anxiety, and body shape concerns with college students yielded an average of 260 messages (sent by users only) at week 8. Thus, the engagement level seen in this study is notably high.

The high level of engagement can be explained by the hybrid nature of the intervention. Pat was integrated into a therapist-supported model in which caregivers received professional guidance from the therapist, personalized reminders from the AI, structured modules, and opportunities to practice skills through role plays with Pat. During the live meetings, the professional assigned caregivers a specific module to be completed with Pat, and at the following meetings, those modules were discussed. Unlike traditional digital parenting programs that primarily deliver educational content, Pat used conversational AI to teach the foundational aspects of parenting skills, promote skills practice through role plays and reminders, and provide immediate support tailored to caregivers’ ongoing parenting challenges. Interestingly, caregivers completed 1 module a week on average and did 1 “in-the-moment” support chat every week. These chats were spontaneously initiated by the parents; therefore, engagement extended beyond the assigned activities. Notably, instead of having just 1 group session a week, parents had 3 support sessions total: the live session, the module with Pat, and the in-the-moment support with Pat, meaning Pat could provide considerably more support to parents.

Rivera-Cepeda et al [19] showed that caregivers experienced Pat as useful and supportive, and they attributed more progress to their chats with Pat than to their participation in the group. It is plausible that when caregivers experience the intervention as helpful in managing real-life parenting challenges, this reinforces continued engagement with the platform. Overall, these findings suggest that generative AI can support considerable caregiver engagement and that such engagement may be amplified when the technology is embedded within a human-supported model of care.

Caregivers reported high levels of satisfaction with the intervention, as reflected in an overall NPS of 76.92, which is considered excellent and indicative of a strong likelihood of recommending the program to others. NPS scores remained consistently high across time points, with values of 77.8 at week 4, 76.7 at week 8, and 75 at week 12. This pattern suggests that positive perceptions of the intervention were established early and maintained throughout the program and did not decline with continued use of the CA. One possible explanation is that early interactions with Pat facilitated the formation of positive initial impressions, as caregivers were able to immediately access guidance and support during real-life parenting challenges, allowing them to experience the relevance and usefulness of the intervention early on. This was reinforced as caregivers continued to engage with the platform and apply parenting strategies in real-world contexts. Rivera-Cepeda et al [19] reported that a qualitative analysis of caregivers’ experiences showed that they valued constant availability and in-the-moment support as some of the aspects most valued by the parents.

These findings are consistent with prior research on parenting CAs, which has demonstrated high levels of satisfaction and acceptability [14]. Similarly, systematic reviews have found high acceptability and retention rates for CA interventions in parenting and mental health contexts for children and adults [15]. The strong NPS observed in this study aligns with this growing body of literature and further supports the acceptability of AI-augmented parenting programs. Importantly, much of the existing literature assesses satisfaction at a single postintervention time point [15]. Taken together, these findings suggest that an approach that combines human therapists who encourage the use of the AI can sustain parents’ satisfaction across extended engagement, which is critical for the delivery of the parenting skills that Pat taught the parents. Although the acquisition of parenting skills was not measured, this represents an important direction for future research.

Beyond their high willingness to recommend Pat, caregivers reported high levels of satisfaction with the intervention, as measured by the TAI, with mean scores increasing from week 4 to weeks 8 and 12. Overall, the responses from the TAI showed that caregivers reported positive impressions of the program, reported improvements in their children’s behavior, gained confidence, learned practical parenting techniques, and experienced improvements in their relationships with their children. These patterns are consistent with emerging evidence suggesting that perceived therapeutic alliance is associated with engagement and clinical outcomes in digital interventions [24]. Together, these findings indicate that alliance ratings remained stable across the weeks of administration, with consistently strong agreement on therapeutic goals, tasks, and relational bonds.

As caregivers gained experience applying the strategies introduced in the intervention, their confidence and perceived effectiveness may have increased, contributing to more favorable evaluations in later weeks. These findings are consistent with prior research on parenting interventions, which have demonstrated high levels of satisfaction and perceived improvement in both caregiver confidence and child behavior outcomes [22,25]. Previous studies have reported similarly high TAI scores, with caregivers indicating meaningful improvements in parenting skills and child behavior alongside strong satisfaction with the intervention process [25]. However, most prior research has assessed satisfaction at a single time point, limiting the understanding of how caregiver perceptions evolve across treatment. The present study examined satisfaction across treatment, demonstrating that positive perceptions increased across multiple time points. These findings suggest that caregivers responded positively to the intervention and perceived meaningful changes in both their parenting practices and their children’s behavior over time.

Therapeutic alliance ratings with Pat remained consistently high across the intervention, with mean scores of 4.10 (SD 0.63) at week 4, 4.10 (SD 0.55) at week 8, and 4.04 (SD 0.63) at week 12, and no significant differences were observed across time. This pattern suggests that caregivers were able to establish a strong working alliance with the CA early in the intervention and maintain it throughout the program. The stability of these scores indicates that continued interaction with the AI did not diminish the perceived relational bond, agreement on goals, or collaboration on tasks. Overall, the therapeutic alliance scores reported in this study are relatively high compared with both CA and human-delivered interventions [21,26]. Prior research examining alliance with digital CAs has reported lower scores, including means of 3.75 for Wysa [27] and 3.36 for Woebot [26]. Other studies have also demonstrated the formation of therapeutic alliances using alternative measures, with reported scores that range approximately between 4.22 and 4.49 on the Working Alliance Questionnaire [28]. Alliance scores in this study also compare favorably with those reported in some human-delivered psychotherapy contexts, such as the approximately 3.8 reported by Munder et al [21]. The therapeutic alliance scores observed with Pat suggest that caregivers may have experienced a strong sense of connection, collaboration, and support when interacting with Pat. One possible explanation for these findings is that Pat provided caregivers with immediate access to support, opportunities to practice parenting skills, personalized guidance, and assistance applying PMT strategies between sessions. These features may have contributed to perceptions of collaboration, support, and shared goals, which are central components of a therapeutic alliance. It is possible that the high scores reported about Pat are influenced by the hybrid approach that included a human therapist. As such, the alliance formed with Pat should not be interpreted as equivalent to an alliance in individual psychotherapy. Rather, these findings suggest that when humans and AI are combined, this can result in a stronger therapeutic alliance than when each component is isolated. Within the context of a hybrid PMT program, Pat appears to function as more than a supplementary tool; it serves as a relational and collaborative component that supports caregivers in applying parenting strategies in real-world contexts.

The findings of this preliminary study have several implications for clinicians, program developers, and health systems. For clinicians, hybrid models that combine live PMT through video conferencing with AI support can enhance scalability and continuity of care, allowing families to receive personalized guidance from their homes and receive support between sessions without increasing therapist workload. AI tools like Pat can function as engagement boosters (ie, reminding, reinforcing, and tracking caregivers’ use of skills). While this study did not analyze engagement, other studies on group interventions with adults with depression have shown that AI could enhance engagement with the treatment and improve outcomes [29]. For developers, including accessible resources (such as apps) and voice features to avoid typing will be essential to ensure smoother engagement with the CAs. Finally, for health systems, integrating AI-supported interventions could lead to greater accessibility and cost efficiency, reducing logistical barriers such as time, transportation, and provider shortages that often limit families’ participation in evidence-based parenting programs.

Several limitations should be considered when interpreting these findings. This study was conducted as a real-world implementation without a control group, which limits the ability to draw causal conclusions about the effects of the intervention. Because the design did not include a comparison condition, it is not possible to isolate the specific contribution of the AI CA relative to the therapist-led group or the combined hybrid model. Conclusions regarding the specific effects of Pat should be interpreted cautiously, as engagement, satisfaction, and alliance ratings may reflect the combined influence of the therapist and the AI-supported component. The findings of these analyses should be interpreted as evidence of feasibility, engagement, and acceptability rather than efficacy.

The analysis relied primarily on self-report measures, which may be influenced by social desirability, recall bias, or participant expectations. However, these self-report assessments are commonly used in psychotherapy and digital intervention research, particularly when evaluating subjective experiences such as satisfaction and alliance. Nonetheless, the absence of objective child-level or family-level outcome measures limits the ability to determine whether perceived improvements translated into measurable changes in child behavior or family functioning. Additionally, the analyses are based on parents who chose to complete the surveys, which biases the sample.

The sample consisted largely of highly educated, digitally literate caregivers with consistent internet access, which may limit the generalizability of the findings to more diverse or underserved populations. Additionally, participants who remained engaged in the study and completed assessments may differ systematically from those who did not, introducing potential attrition bias and limiting insight into the experiences of less engaged users.

The intervention evolved over the course of the study, with differences in program duration and the availability of optional modules across cohorts. These variations may have influenced engagement patterns and user experiences, making it challenging to determine whether observed differences were due to the intervention structure or participant behavior. Due to the hybrid format of the intervention, which included therapist-led group sessions alongside interactions with the AI agent, perceptions of satisfaction and therapeutic alliance with Pat may have been shaped by the broader treatment context. As such, it is not possible to determine the extent to which the observed alliance and satisfaction are attributable specifically to the AI agent versus the combined human-AI model.

Future research should address these limitations by incorporating randomized controlled designs that allow for direct comparisons between intervention conditions. In particular, studies comparing group PMT with and without the inclusion of Pat would help clarify the unique contribution of Pat to the outcomes. Additionally, future studies should expand beyond experimental comparisons and examine the relationships and mechanisms underlying these outcomes. While the present analysis provides evidence of feasibility, engagement, and acceptability, it does not examine how these variables interact or contribute to one another. Understanding these relationships will be crucial for identifying the mechanisms through which AI-supported interventions influence caregiver experiences and outcomes.

Although the inclusion of caregivers from Latin American countries represents a strength in terms of cultural context, future research should aim to include more socioeconomically and educationally diverse samples, particularly those with varying levels of digital literacy and access to resources, in order to better assess the generalizability and scalability of Pat. Finally, because of the hybrid format, ratings of engagement, satisfaction, and therapeutic alliance may reflect the combined influence of the therapist-led and AI-supported components rather than the AI component alone and therefore should be interpreted cautiously.

Caregivers reported high engagement, satisfaction, and therapeutic alliance with a hybrid PMT program combining a human therapist and an AI CA. Caregivers interacted extensively with the platform, completing assigned modules and using optional in-the-moment support chats, suggesting that the intervention was used consistently and integrated into daily parenting practices. Satisfaction remained high across time, and caregivers reported positive therapy attitudes, indicating that the program was perceived as helpful, relevant, and worth recommending to other caregivers. Therapeutic alliance was also strong and stable across the intervention, which suggests that caregivers were able to establish a meaningful working relationship with Pat. A notable finding was the value of the in-the-moment support, which allowed caregivers to apply parenting strategies directly to real-life situations. These findings suggest that integrating AI with a human therapist is feasible and acceptable within a group PMT program and may provide additional support between scheduled sessions. Parents were open to this hybrid model and utilized the support provided by the AI in their everyday lives.