The reviewed studies demonstrated a wide- array of AIBC platforms, including conversational agents, just in time adaptive interventions, digital twins and hybrid models. Despite differences in form and delivery, these platforms share core functionalities, delivering structured behavioral interventions, analyzing user generated data (e.g., physical activity, diet, sleep), and providing tailored feedback. Most were integrated into mobile apps or wearable devices, aiming to enhance motivation, engagement and sustained behavior change. Common behavioral techniques used across platforms included goal-setting, self-monitoring, feedback loops and gamification. The subsequent results are organized by health outcomes and subcategorized into study type.

Engagement and adherence were key metrics evaluated across nearly all included studies and serve as critical indicators of the feasibility, scalability, and acceptability of AI-based behavioral coaching interventions. Retention rates and frequency of interaction with digital tools varied by platform type, intensity, and user population.

In the PROTEIN (gamified nutrition) app RCT, Petra et al. (2024) found that 57% of users were still active at one year, highlighting long-term adherence in a general wellness context [6].

Among observational studies, Stein et al. [2], who implemented a fully automated conversational AI coach, reported that 62% of participants remained actively engaged after 6 months, with an average of 3.1 to 4.5 logins per week in an observational study [2]. Meanwhile, Graham et al. [9] observed a clear association between higher frequency of self-monitoring (e.g., weigh-ins) and greater weight loss over a 1 year retrospective longitudinal study. Supporting results of previously established relationships between self-monitoring and weight loss [24]. Participants who engaged daily with the AI coach were significantly more likely to achieve clinically meaningful outcomes, demonstrating that sustained interaction amplifies effectiveness [9, 24].

Further, Khokhar et al. [1, 20] reported exceptionally high engagement, supported by weekly weight logs and app activity. In the 2024 cohort, the platform demonstrated near-universal usage, with > 90% of participants meeting minimum adherence benchmarks through 26 weeks [1, 20]. Lockwood et al. [21] also observed high adherence, with over 70% of participants achieving sustained interaction thresholds and consistent logging behaviors [16].

Collectively, these findings support the feasibility of deploying AI-based health interventions at scale, with engagement outcomes that compare favorably to traditional in-person programs. Importantly, adherence appeared to be enhanced when programs integrated multimodal feedback, gamification, or limited human support, allowing for flexible and sustained user participation.

Of the included studies, 6 were randomized control trials, and 8 were observational in design. All but one study reported statistically significant weight reduction from baseline, with several demonstrating clinically meaningful effects.

Among the RCTs, a 12-month trial conducted by Joshi et al. [17] compared a digital twin coaching system to standard care, showing robust result, with 73.8% of the intervention group participants achieved > 5% body weight loss (BWL), and 41.6% achieved > 10% BWL, Statistically significant differences were observed in mean weight change (−7.4 kg vs. −0.4 kg; p < 0.001), BMI reduction (−2.7 vs-0.1; p < 0.001) and waist circumference reduction (−9.5 cm vs. 1.2 cm; p < 0.001) [17]. Shamanna et al. [5] also reported significant improvements in the intervention group at 1 year compared to controls: with mean weight change of − 5.2 kg versus − 1.1 kg (p < 0.0001) and BMI reduction of − 2.3 kg/m² versus − 0.4 kg/m² (p < 0.0001) [5]

Nakata et al. (2022) evaluated a 3-month AI-assisted lifestyle intervention, observing a mean weight loss of 2.4 ± 4.0 kg in the intervention group compared to 0.7 ± 3.3 kg in the control group. Adjusted analyses confirmed a significant between-group difference of − 1.60 kg (95% CI − 2.83 to − 0.38; p = 0.011). Ruize-Leon et al. [4] also conducted a 12-week RCT, there’s resulted in a mean weight loss of 0.8 kg (p = 0.03), BMI reduction of 0.3 kg/m² (p = 0.03), and a 1.0 cm decrease in waist circumference (p = 0.046) [10].

In contrast, Forman et al. (2019) evaluated a 10-week digital cognitive-behavioral therapy intervention (OnTrack) for weight loss with the weight watchers (WW) platform Beyond The Scale (BTS) vs. the weight watchers platform by itself as the control group. However, half way through the RCT the dynamics of the trial changed from using BTS to a Freestyle (FS) diet plan - a newly developed WW digital platform that was released during the study period. With these changes with design during the study, the results were unique in that only BTS-WW participants weight loss was greater for the intervention group (Mean change of = 4.7%, SE = 0.55) than for the control group (Mean change = 2.6%, SE = 0.80). Interestingly, the results were reversed for the FS-WW intervention participants (mean of 2.9%, SE = 0.38) vs. control group (mean change of 4.5%, SE = 0.52) [25]. Similarly, Persell et al. [23] conducted a 6-month RCT focused on changes in blood pressure, but measured secondary outcomes, one of which being change in BMI. They found a mean BMI change of–0.15 (95% CI − 0.51 to 0.20; p = 0.39) [23].

Among observational studies two single-arm fields trials by Khokhar et al. [1, 20] reviewing cardiometabolic effects of the same participant population base experienced a mean weight loss of 17.27 lbs., equating to a total body weight loss (BWL) of 13.9% with a mean BMI reduction of 8.6 points. Weekly weight loss averaged 0.71 kg [20]. Furthermore, 98.7% of participants achieved ≥ 5% weight loss, 75% achieving ≥ 10%, 43% achieved ≥ 15%, and 9% achieved ≥ 20% BWL [1, 20].

In a large cohort (n > 3000), Graham et al. [9] found a mean weight nadir of 4.2% (4.4 kg) achieved at day 150, with 35% of the population achieving > 5% BWL. Among CDC-qualified program participants, mean weight loss increased to 7.0% (7.3 kg). Daily AI coach interaction was associated with a 0.8% increased likelihood of achieving 5% weight loss. Also, higher self-weighing frequency was independently associated with greater weight loss, supporting previously established literature on impact of self-weighing frequency on weight loss [9, 24]. Paz et al. [14] conducted a retrospective analysis on engagement cardiometabolic changes after engagement with a digital lifestyle-medicine app tied with self-monitoring, with a weight cohort of 16,402 individuals. They classified reduction of weight in pounds by baseline BMI category and found that weight reduction was inversely related to BMI, with an average loss of 12.0 lbs (SEM 0.3) equating to a 5.1% weight loss among users with BMI ≥ 30 [14]. Colwell et al. [13] observed a combined BMI and weight loss improvement of 6.5% over 12 weeks in a large real-world cohort (n = 320) based on lifestyle course lengths from 4 to 32 weeks, with statistically significant results (p < 0.01) [13].

Additionally, 3 observational studies supported a modest weight loss: Lockwood et al. [21] completed a single-arm pilot study over 3 months (n = 509) found an average weight loss of 3.8% (SD 2.9%; 95% CI 3.5%−4.1%) of baseline body weight. A total of 71.2% of participants achieved ≥ 2% weight loss, and 26.5% achieved ≥ 5% [16]. Stein et al. [2] evaluated a fully automated AI coach and found a mean weight loss of 2.4 kg (2.4% of body weight) over 15 weeks, with 75.7% of participants achieving weight loss during the intervention.² Maher et al. (2020) documented a mean weight loss of 1.1 kg by week 6 and a cumulative 1.3 kg by week 12 (95% CI − 2.5 to − 0.7; p = 0.01). As well as a waist circumference decline of 2.1 cm over the 12-week period (p = 0.003) [11].

Only 1 observational study reviewed pediatric populations: Zarkogianni et al. [3] evaluated the ENDORSE platform in a pediatric population and found a significant reduction in BMI z-score (mean − 0.21 ± 0.26; p < 0.001). A negative correlation was observed between use of an activity tracker and BMI z-score (r = − 0.355; p = 0.017) [3].

Taking both RCTs and observational study data together, these findings illustrate that AI-driven behavioral coaching intervention can achieve weight loss across diverse populations, settings, and delivery models. While the magnitude of change varied, ranging from modest (–0.8 kg) to highly clinically significant (–13.9% BWL). Many studies met or exceeded the 5% threshold considered beneficial for clinically significant weight loss. The most effective interventions often involved personalized feedback loops, predictive modeling, or integration with human coaching, suggesting that tailored engagement and hybrid designs may enhance outcomes.

Several studies reported outcomes related to blood pressure, highlighting the potential for AI-assisted behavioral interventions to positively influence cardiovascular risk factors. Shamanna et al. [5], completed an RCT that found the intervention group achieved significant reductions in systolic blood pressure (SBP) (−7.6 vs. −3.2 mm Hg; p < 0.007) and diastolic blood pressure (DBP) (−4.3 vs. −2.2 mm Hg; p = 0.046) after 1 year compared with the control group. They also found that among participants with HTN, the intervention group achieved higher rates of normotension (40.9% vs. 6.7%; p = 0.0009) and HTN remission (50% vs. 0%; p < 0.0001) than the control group [5].

Ruiz-Leon et al. [4], conducted a 12-week RCT evaluating a digital lifestyle intervention in older adults with overweight and obesity. The intervention group experienced − 4.5mmHg (CI: −9.0 to 0.0) change in systolic BP compared to a −5.0mmHg (CI: −9.4 to −0.7) change in the control group with a difference of 1.3 (CI: −10.2−12.8), as well as a Diastolic BP, −2.4mmHg (CI:−4.4 to −0.3) in the intervention group vs. −0.4mmHg (CI: −2.5 to 1.6) in the control, a difference of −3.5 (CI: −8.9 to 1.8) over 12 weeks [4]. Thus all the data is not statistically significant.

Another RCT, Persell et al. [23] found that over the 6-month trial, the corresponding mean systolic blood pressures were 132.3mmHg (from a 140.6 mm Hg baseline) and 135.0 mm Hg (from a 141.8 mm Hg baseline), with a between-group adjusted difference of − 2.0 mm Hg (95% CI, − 4.9 mm Hg to 0.8 mm Hg; p = 0.16). Interestingly they also found that at 6 months, self-confidence in controlling blood pressure was greater in the intervention group (0.36 point on a 5-point scale; 95% CI, 0.18 to 0.54 point; p < 0.001) [23].

Branch et al. [16], completed a large observational study evaluation the effectiveness of an AI powered hypertension care app. Among 717 participants with baseline hypertension, the program resulted in a mean reduction of −5.4 mm Hg (95% CI −6.5 to −4.3; p < 0.01) in SBP and − 1.2 mm Hg (95% CI −2.1 to −0.5; p < 0.002) in DBP over a 3-month engagement period. Notably, participants who experienced > 5% weight loss had significantly greater SBP and DBP reductions compared to those who did not lose weight, suggesting a synergistic effect on combined weight loss and digital engagement. The study also highlighted the level of engagement with the digital platform (e.g., app use frequency, logging behavior) was positively correlated with the magnitude of blood pressure improvement. To note, they also found that SBP and DBP changes were greatest in those with stage 2 hypertension [16].

Among the other observational study findings, Paz et al. [14] documented a substantial reduction in SBP of 18.6 mmHg after 24 weeks of AI-guided lifestyle modifications [14]. Similarly, Leitner et al. [15] led a trial utilizing an autonomous lifestyle-guidance system. They found an average SBP reduction of 8.1 mmHg over the 24-week course of the intervention [14]. In a large digital diabetes prevention program (DPP) analyzed by Graham et al. [9], mean SBP decreased by 5.4 mmHg at three months, and this reduction was sustained through six months.⁹ These findings align with reductions commonly observed in traditional intensive lifestyle interventions and underscore the potential of AI tools to serve as scalable adjuncts to hypertension management.

Diastolic blood pressure (DBP) outcomes were specifically reported in three studies: Paz et al. [14], Graham et al. [9], and the autonomous lifestyle-guidance system trial. In the Paz et al. study, DBP declined by 9.4 mmHg; in the autonomous guidance system, DBP decreased by approximately 6.2 mmHg; and in the digital DPP analyzed by Graham et al., DBP was reduced by 3.2 mmHg [9, 14]. These consistent reductions in DBP reinforce the capacity of AI-driven tools to influence both components of blood pressure meaningfully. The degree of reduction was influenced by baseline hypertension severity, level of program adherence, and integration with other behavioral strategies like weight tracking and dietary logging.

Collectively, these findings suggest that AI-enabled coaching platforms may have beneficial effects on blood pressure regulation, particularly when integrated with weight loss strategies and real-time health monitoring. The magnitude of SBP reduction observed in some studies approaches the efficacy of first-line antihypertensive medications, supporting the role of digital coaching as a non-pharmacologic complement in managing hypertension in individuals at risk for obesity-related cardiometabolic disease.

Several studies reported on the effects of AI-assisted behavioral coaching on diabetes-related outcomes, particularly glycemic control. In the Greenhabit randomized controlled trial [4], a more modest yet clinically relevant HbA1c reduction of 0.4% points was observed, over a 1 week period [4]. Additionally, Joshi et al. [17] RCT, noted a significant decline in fasting glucose and HbA1c among individuals (−2.9 [1.8] vs. −0.3 [1.2]; p < 0.001) at 1 year with 72.7% remission of T2D [4].

Colwell et al. [13], conducted an observational study on the Redicare digital twin–enabled platform showed one of the most significant improvements in glycemic metrics, with a reduction in hemoglobin A1c (HbA1c) of 1.2% points (10.9%, n = 80, p < 0.01) over the course of the intervention [13].

Improvements in lipid parameters were also documented in one randomized control trials reviewed. Nakata et al. (2020) RCT over a 12-week timeframe calculated difference in HDL, LDL and triglycerides. They found that the HDL levels adjusted mean between groups at 12 weeks was 1.42 mg/dL (95% CI: −1.05, 3.89; p = 0.22), LDL levels adjusted mean between groups at 12 weeks was − 2.45 mg/dL (95% CI: −7.64, 2.74; p = 0.34) and a triglyceride adjusted mean difference of −23.73 mg/dL (95% CI: −62.06, 14.6o; p = 0.22) [10]. Although all p values indicate that the data was not significant it still provides a level of evidence for us to make assumptions and base future research on results.

A few observational studies also found some interesting data on lipid changes. Paz et al. [14] demonstrated a significant reduction in low-density lipoprotein cholesterol (LDL-C) of −66.6 mg/dL, accompanied by a 28% decline in triglyceride levels among users of a digitally guided behavior-change platform [14]. In a separate analysis, Colwell et al. [13] reported a 30% decrease in triglyceride levels after 12 weeks of participation in a hybrid AI-human coaching program [13].

Limited data were available regarding hepatic outcomes; Joshi et al. [17] completed an RCT and tracked participants in numerous hepatic markers of MASLD. Following intervention, they found a drastic increase in users towards normal MASLD-Liver Fat Scores (from 11.8 to 67.4% in the intervention group, whereas it reduced from 16 to 9.9% at 1 year in the standard care group (p < 0.00001). Additionally, of the 10 patients in the intervention group with abnormal Fib-4 scores, 9 (90%) fell into the normal range at 1 year, compared to only 2 of the 8 (25%) in the standards of care group [17]. While preliminary, these findings warrant further investigation into the hepatometabolic impacts of AI-driven lifestyle modification.