The VALUE of AI-Guided Communication: Enhancing Shared Decision-Making in Metabolic Bariatric Surgery Consultations Through a Metacognitive Framework Authors: Chun Gao, Yang Fan, Fei Yao, Sheng Zhang Published: 2026 ABSTRACT BackgroundEffective communication in metabolic bariatric surgery (MBS) is essential for patient engagement and adherence, yet surgical residents often lack structured training. Artificial intelligence offers a novel approach to scaffold communication skills.ObjectiveTo evaluate the impact of an AI-Guided metacognitive framework VALUE (Validate, Align & Reframe, Link & Educate, Unite in a plan) on shared decision-making (SDM) and communication outcomes in MBS consultations compared to self-directed learning.MethodsForty surgical residents were randomized into two groups: AI-Guided (using the VALUE framework) and self-learning. The AI-Guided group used a structured prompt to interact with a large language model (DeepSeek-V3.2) to generate personalized consultation plans. Each conducted simulated consultations with standardized patients from a case library. Outcomes were measured using the Shared Decision-Making Questionnaire-9 (SDM-Q-9), Decision Conflict Scale (DCS), Four Habits Coding Scheme (4HCS), Surgeon Self-Efficacy scale (SSI-BS), Communication Outline Quality Scale (CQS), and AI Interaction Quality (AIIQ). The trial was registered on the Open Science Framework (Registration DOI: 10.17605/OSF.IO/BAQH6).ResultsThe AI-Guided group scored significantly higher on SDM-Q-9 (84.7 vs. 71.3, p < 0.01) and 4HCS (17.5 vs. 14.8, p < 0.01), and lower on DCS (19.5 vs. 32.1, p < 0.01). Communication outlines were also of higher quality (13.8 vs. 7.5, p < 0.01). Residents reported greater self-efficacy gains in information provision, values integration, decision facilitation, and emotional support. All secondary analyses remained significant after Benjamini-Hochberg correction for multiple comparisons.ConclusionAn AI-Guided metacognitive communication framework significantly improves shared decision-making, reduces decisional conflict, and enhances communication quality and self-efficacy in MBS consultations, suggesting a promising approach that requires further validation in larger, multi-site trials.Supplementary InformationThe online version contains supplementary material available at 10.1007/s11695-026-08760-9. Introduction Metabolic bariatric surgery (MBS) stands as one of the most effective interventions for severe obesity and its associated medical problems, offering durable recurrent weight gain prevention, diabetes remission, and reduced mortality [1]. Its optimal clinical response, however, extends far beyond technical mastery in the operating room; it is profoundly dependent on the quality of the preoperative consultation [2]. This conversation must navigate complex medical trade-offs, align with deeply personal life goals such as fertility or career ambitions, and build the trust necessary for lifelong postoperative partnership [3]. Yet herein lies a critical vulnerability in our healthcare system: the communication gap between surgical expertise and decision-making. Surgical residents, while skilled at learning procedures, often receive incomplete training in handling these complex, value-laden conversations [4, 5]. They meet patients who have gathered misleading information from social media, who are terrified of surgical risks, or who focus only on looking better [6]. The main challenge is not simply sharing medical facts; it is helping each patient understand how those facts apply to their own life, while also addressing their fears, priorities, and personal goals [7]. Traditional teaching methods do little to develop this advanced skill, leaving residents ill-prepared. The results are clear: poor decisions, weak preparation for long-term follow-up, and eventually, dissatisfaction or recurrent weight gain [8, 9]. Artificial intelligence (AI), particularly large language models, is catalyzing a transformation in medical education and clinical support, moving beyond knowledge retrieval towards sophisticated reasoning and communication coaching [10, 11]. In patient-facing contexts, AI-driven tools demonstrate promise in enhancing pre-consultation education and post-operative support [12]. However, a significant and under-explored opportunity exists: leveraging AI not as a patient-facing chatbot, but as a metacognitive scaffold for clinicians during the consultation planning phase–the critical window where shared understanding is architected [13, 14]. We posit that effective MBS communication is a learnable clinical skill that can be deconstructed, practiced, and reinforced. Drawing from established principles of motivational interviewing, shared decision-making, and cognitive reframing, we developed the VALUE framework–a structured, sequential approach (Validate, Align & Reframe, Link & Educate, Unite in a plan). This study investigates whether an AI-Guided interaction, structured around this metacognitive framework, can significantly enhance surgical residents’ communication performance compared to traditional self-directed learning. We measure its impact on the observable quality of the consultation, the patient’s experience of shared decision-making, and the essential reduction in their decisional conflict. Methods Study Design and Participants We conducted a prospective, randomized, controlled educational trial between August and December 2025. The study protocol was approved by the Institutional Review Board and was prospectively registered on the Open Science Framework (Registration DOI: 10.17605/OSF.IO/BAQH6). This manuscript adheres to the Consolidated Standards of Reporting Trials (CONSORT 2025) guidelines for randomized trials and the CONSORT-AI extension of the CONSORT 2010 statement [34, 35] for trials involving artificial intelligence interventions. (see supplemental material for complete checklist). We recruited surgical residents (Postgraduate Year 1–5) from a single academic medical center gastrointestinal surgery department with an accredited MBS program. Eligibility required current involvement in surgical clinics and informed consent. We excluded residents who completed a formal communication skills course in MBS care within the preceding year. This was a parallel-group, 1:1 allocation ratio, superiority randomized controlled trial. Randomization used computer-generated random numbers with block sizes of 4, stratified by PGY level (junior [PGY 1–3] vs. senior [PGY 4–5]) to ensure balanced distribution of training levels across groups. Allocation was concealed using sequentially numbered, opaque, sealed envelopes prepared by an independent statistician. The study coordinator opened the next envelope after enrolling a resident. Funders had no role in study design, data collection, analysis, interpretation, or manuscript preparation. Using a computer-generated randomization sequence with concealed allocation, we assigned 40 eligible residents to either the AI-Guided intervention group (n = 20) or the self-learning (control) group (n = 20) (Fig. 1). All participants received standard baseline didactic materials on MBS procedures (sleeve gastrectomy (SG) and Roux-en-Y gastric bypass (RYGB)).Fig. 1CONSORT flow diagram of participant enrollment, randomization, and analysis CONSORT flow diagram of participant enrollment, randomization, and analysis To address the inclusion of residents across PGY1-5, we note that senior residents (PGY4-5) may have had prior exposure to bariatric decision-making principles. However, randomization was stratified by PGY level (junior vs. senior) to balance experience across groups, and ANCOVA with PGY as a covariate showed no significant interaction. All residents completed a pre-study survey documenting prior bariatric surgery rotation, formal communication course, or structured MBS training. The duration of prior bariatric clinic exposure (Table 2) was similar between groups and was included as a covariate in the primary analysis; it was not statistically significant. Residents who had completed a formal communication skills course in MBS care within the preceding year were excluded; those with training more than one year prior were not excluded, but their prior exposure was systematically assessed and controlled for the analysis. The Value Framework and AI-Guided Intervention The intervention was built upon the VALUE framework (Table 1), a metacognitive scaffold designed to structure a values-aligned consultation. The framework’s sequential steps are: Validate the patient’s perspective (V), Align on a shared superordinate goal (A), Link medical information to that goal using comparative framing (L), and Unite in a collaborative plan (U). The “E” in the framework title refers to the metacognitive scaffold (the prompt structure) that enables the V-A-L-U steps, not a separate consultation step.Table 1The VALUE framework: a metacognitive scaffold for values-aligned communicationStepCore taskExemplar phrases/ActionsPsychological principle/GoalV-ValidateAcknowledge patient’s perspective without judgment.“It makes sense that you’re drawn to…”“Thank you for sharing that concern.”Motivational interviewing (express empathy); Reduce defensiveness.A-Align & ReframeShift focus to a superordinate, shared goal based on patient values.“If our goal is to achieve [Patient’s Value, e.g., safe pregnancy], then let’s evaluate…”Cognitive reframing; Create shared agenda.L-Link & Educate with contrastProvide medical information linked directly to the shared goal, using comparative framing.“Option A is stronger for [Value], while Option B is simpler but may compromise [Value].”Decision-aid theory; Enhance salience and understanding of trade-offs.U-Unite in a planCo-create a specific action plan, exploring barriers and next steps.“Given this, what feels like the most manageable first step for you?”“Let’s plan how to address that potential barrier.”Shared decision-making & Implementation intentions; Foster commitment and partnership.E-(Scaffold)Metacognitive process: The AI prompt requires the clinician to sequentially consider V-A-L-U before the encounter.Not a phrase, but a preparation step.Scaffolded learning; Externalizes expert thought process for skill acquisition. Table 1The VALUE framework: a metacognitive scaffold for values-aligned communicationStepCore taskExemplar phrases/ActionsPsychological principle/GoalV-ValidateAcknowledge patient’s perspective without judgment.“It makes sense that you’re drawn to…”“Thank you for sharing that concern.”Motivational interviewing (express empathy); Reduce defensiveness.A-Align & ReframeShift focus to a superordinate, shared goal based on patient values.“If our goal is to achieve [Patient’s Value, e.g., safe pregnancy], then let’s evaluate…”Cognitive reframing; Create shared agenda.L-Link & Educate with contrastProvide medical information linked directly to the shared goal, using comparative framing.“Option A is stronger for [Value], while Option B is simpler but may compromise [Value].”Decision-aid theory; Enhance salience and understanding of trade-offs.U-Unite in a planCo-create a specific action plan, exploring barriers and next steps.“Given this, what feels like the most manageable first step for you?”“Let’s plan how to address that potential barrier.”Shared decision-making & Implementation intentions; Foster commitment and partnership.E-(Scaffold)Metacognitive process: The AI prompt requires the clinician to sequentially consider V-A-L-U before the encounter.Not a phrase, but a preparation step.Scaffolded learning; Externalizes expert thought process for skill acquisition. The VALUE framework: a metacognitive scaffold for values-aligned communication “It makes sense that you’re drawn to…” “Thank you for sharing that concern.” “Given this, what feels like the most manageable first step for you?” “Let’s plan how to address that potential barrier.” AI Intervention Specification Residents in the AI-Guided group used a structured prompt interface to interact with a commercially available large language model (Deepseek v3.2, snapshot 2025-09-01; accessed via API at api.deepseek.com/v3 between Sep 1–Oct 1, 2025; temperature = 0.7, max tokens = 2000; no system prompt or fine-tuning was applied). DeepSeek-V3.2 (snapshot 2025-09-01) was selected for its strong performance in structured reasoning tasks, transparent versioning, stable API during the study period, and a 128k token context window that allowed processing of full case vignettes without truncation. Case vignette criteria for AI input: All cases met the following criteria: (1) representative of common MBS consultation challenges, (2) contained sufficient clinical and psychosocial detail to support values exploration, (3) reviewed by three bariatric surgeons for authenticity. Cases were standardized in length (300–400 words) and structure. The exact same prompt structure (Supplemental Table 2) was used for all residents and all cases, with only the case-specific details inserted. Input data handling for AI: All case vignettes were formatted identically (plain text, no markdown). Vignettes were reviewed for clarity and completeness by the research team before use. A random sample of 10% of prompts was reviewed for formatting errors; none were found. All patient identifiers were removed; cases used fictional names and demographics. For a given case from our library, the full VALUE prompt template (provided verbatim in Supplemental Table 2) required them to sequentially apply the V-A-L-U steps to generate a personalized consultation outline, including specific phrases, value-exploration questions, and contingency plans. Residents completed a single, structured prompt submission per case; iteration was not permitted to ensure standardization. Time allowed was 15 min per case, monitored by study staff. AI-generated outlines were saved and used only for the specific case; no cross-case reuse was permitted. They were instructed to use this AI-generated outline as their primary preparation tool, but were explicitly told to use it as a thinking scaffold, not as a verbatim script; they were free to modify, adapt, or discard any suggestions. This process was designed not to provide scripted dialogue but to externalize expert cognitive workflow. Three anonymized AI-generated consultation outlines (for Cases 1, 2, and 3) are provided in Supplemental Tables 1 to illustrate output format and content. During the simulated consultation (immediately following the 15-minute preparation phase), residents were not permitted to re-access the AI-generated outline or any other preparation materials. The consultation was conducted as a closed-book assessment of their ability to apply the prepared plan. No delayed (e.g., 1-week or 1-month) assessment was performed; therefore, this study evaluates short-term performance only, not long-term retention or skill consolidation. Self-Learning Group (Self-Directed Learning) Residents in the self-learning group received the same case information and baseline medical materials. They were instructed to prepare for the consultation using any resources they deemed appropriate (e.g., textbooks, guidelines, personal notes). Generative AI tools (ChatGPT, DeepSeek, etc.) were explicitly prohibited, confirmed by post-study self-report. Self-learners were given the same 15-minute preparation window as the intervention group, matched to ensure comparability. To prevent contamination, residents were instructed not to discuss their preparation methods with peers; separate preparation rooms were used; post-study debriefing confirmed no cross-group communication. Their preparation was self-directed and content focused. To characterize control group preparation, post-hoc analysis of self-reported resource use showed: textbooks (85%), clinical guidelines (70%), personal notes (60%), suggesting reasonable consistency in preparation approaches. Case Library and Standardized Patients (SPs) We developed a library of four complex, standardized patient vignettes representative of common MBS consultations (Supplemental Materials: Case Library). Cases included a young patient with fertility goals; a risk-averse patient with multiple associated medical problems; a patient with post-operative recurrent weight gain; and a patient with unrealistic aesthetic expectations. All SPs underwent standardized training (8 h) using detailed portrayal guides. Fidelity checks (random video review of 20% of encounters by the SP trainer) confirm consistent portrayal, with mean adherence score > 90% to the portrayal guides. These strategies may ensure authentic and consistent emotional and behavioral responses across all encounters. All SPs were fully blinded to resident group allocation; they received no information about the study design, the intervention, or which residents were in which group. They were instructed to focus solely on their portrayal and to complete questionnaires immediately after each encounter without discussion. Simulation Procedure and Outcomes Each resident conducted two 15-minute video-recorded consultations with two different SPs. Cases were counterbalanced across residents within each group using a Latin square design to ensure each case appeared equally often in each position (first vs. second encounter). Immediately after each encounter, the SP completed two validated patient-reported outcome measures: the 9-item Shared Decision-Making Questionnaire (SDM-Q-9) [15] and the 16-item Decision Conflict Scale (DCS) [16] (Supplemental Tables 3 & 4). For both instruments, higher scores indicate greater shared decision-making (SDM-Q-9) and greater decisional conflict (DCS). Trained, blinded raters evaluated all video recordings using the Four Habits Coding Scheme (4HCS), a reliable instrument for assessing communication as rated by trained observers [17] (Supplemental Table 5). Video recordings were de-identified (audio altered to remove any mention of preparation method; resident faces blurred). Raters were unaware of group assignment. The statistician performed analyses with groups coded as “A” and “B”; decoding occurred only after final analysis. Primary Outcome: The transformed SDM-Q-9 score (0–100), as rated by the SP. Secondary Outcomes: Included the DCS total and subscale scores, the 4HCS total and habit scores, and the change in surgeon self-efficacy using the Surgeon Self-Efficacy in MBS Communication scale (SSI-BS). (Supplemental Table 6). For the 4HCS, higher scores indicate more effective clinician‑patient communication behavior as defined by the instrument (range 4–20); for the SSI‑BS, higher scores indicate greater self‑efficacy (range 1–10). Process Evaluation For the intervention group only, we evaluated their interaction with the AI tool using the AI Interaction Quality (AIIQ) rubric (Supplemental Table 7). AIIQ was developed specifically for this study; content validity was assessed by an expert panel of three bariatric surgeons and two communication experts using a Delphi process, yielding a Content Validity Index (CVI) of 0.92. AIIQ internal consistency: Cronbach’s α = 0.84. This is a study-specific instrument developed for this trial; its preliminary status is acknowledged as a limitation. We also assessed the quality of the residents’ preparatory work using the Communication Outline Quality Scale (CQS) [18], scoring their submitted notes on structure, values integration, and medical accuracy (Supplemental Table 8). CQS was developed by three bariatric surgeons and two communication experts using a Delphi process. The development process involved three rounds: (1) item generation based on literature review and clinical experience, (2) content validity rating by an expanded panel of 10 experts, (3) final refinement and pilot testing with 10 sample outlines. A qualitative, side-by-side comparison of the consultation process and outcomes for a single representative case (case 1) was shown in Supplemental Table 9. All observer-rated instruments (4HCS, CQS, AIIQ, and SSI-BS behavioral components) were scored by two independent, trained raters: (1) 4HCS raters: Two clinical psychologists with 5+ years’ experience in communication assessment; (2) CQS and AIIQ raters: Two PhD-level medical educators with expertise in surgical education; (3) SSI-BS rater-coded components: Integrated into 4HCS ratings by the same clinical psychologists. Rater training protocol: All raters underwent intensive training including: (1) 4–8 h of didactic instruction on the instrument and scoring rubrics, (2) independent scoring of 5–10 practice samples (videos or outlines), (3) feedback sessions with the study coordinator to resolve discrepancies, (4) a certification requirement: achieving ICC > 0.80 with gold standard (trainer’s scores) on a final set of 5 practice samples before study commencement. During the study, 20% of samples were randomly selected for double-coding to monitor drift; any discrepancies > 1 point were resolved by consensus, and all final scores used in analysis are consensus scores. Inter-rater reliability: All instruments demonstrated excellent reliability (ICC > 0.75). Full details including 95% confidence intervals, F-statistics, and subscale-specific ICCs are provided in Supplemental Table 10. Summary ICC values: 4HCS total ICC = 0.87 (95% CI 0.79–0.93); CQS total ICC = 0.89 (95% CI 0.82–0.94); AIIQ ICC = 0.85 (95% CI 0.76–0.91); SSI-BS rater-coded components ICC = 0.88 (95% CI 0.81–0.93). Statistical Analysis We performed analyses on an intention-to-treat basis. All outcomes were analyzed per resident (averaged across their two consultations), recognizing that observations are clustered within residents. We used linear mixed models with resident as a random intercept to confirm robustness; results were consistent with the simpler average analysis. We presented baseline characteristics using descriptive statistics. Normality was assessed using Shapiro-Wilk tests; all p > 0.05, supporting use of parametric tests. For continuous outcomes, we compared groups using independent samples t-tests and analysis of covariance (ANCOVA) with PGY level and baseline SSI-BS as covariates to adjust for potential confounding. We reported mean differences with 95% confidence intervals (CI) and calculated effect sizes using Cohen’s *d*. To account for multiple comparisons, we applied the Benjamini-Hochberg False Discovery Rate (FDR) correction (q = 0.05) to all secondary and exploratory outcomes. The outcome hierarchy was pre-specified as: primary (SDM-Q-9), secondary (DCS, 4HCS, SSI-BS), exploratory (CQS, AIIQ, subgroup analyses). We considered a two-sided p-value < 0.05 statistically significant after FDR correction. We conducted all analyses using SPSS Statistics (Version 28, IBM Corp) and R software version 4.3.0. Sample size justification: Based on a pilot study (n = 5 per group) showing SDM-Q-9 mean difference of 12 points (SD = 10), alpha = 0.05, power = 0.80, a priori sample size calculation required n = 18 per group. We enrolled 20 per group to account for potential attrition. The minimal clinically important difference (MCID) for SDM-Q-9 is ≥ 10 points [15]. Post-hoc power analysis with observed effect size d = 1.42 achieved power > 0.99. Results Participant Flow and Baseline Characteristics We assessed 44 surgical residents for eligibility. Four were excluded (two declined, two had prior training), and 40 were randomized (Fig. 1). All residents completed the study and were included in the intention-to-treat analysis. No missing data occurred. As shown in Table 2, the groups were well-balanced at baseline. There were no statistically significant differences in age, sex, training level, prior MBS clinic exposure, or clinical context (gastrointestinal specialty) (all p > 0.05). PGY distribution: AI-Guided Group: Junior (PGY 1–3) = 8 (40.0%), Senior (PGY 4–5) = 12 (60.0%); Self-Learning Group: Junior = 7 (35.0%), Senior = 13 (65.0%) (p = 0.78).Table 2Baseline characteristics of surgical residents by study group (N = 40)CharacteristicAI-Guided group (n = 20)Self-learning group (n = 20)p valueDemographicsAge, mean (SD), years30.8 (2.1)31.1 (2.4)0.56Sex, n (%)MaleFemale17 (85.0)3 (15.0)18 (90.0)2 (10.0)0.91Training level, n (%)0.78Junior resident (PGY 1–3)8 (40.0)7 (35.0)Senior Resident (PGY 4–5)12 (60.0)13 (65.0)Prior bariatric clinic exposure, mean (SD), months6.5 (5.5)7.1 (6.2)0.54Clinical contextGastrointestinal surgery specialty, n (%)17 (85.0)18 (90.0)0.81PGY Postgraduate Year, SD Standard Deviation. Notes: Categorical variables are presented as n (%). Continuous variable (Age) is presented as mean (SD). P-values were derived from independent samples t-test for Age, prior clinic exposure, and Chi-square tests (or Fisher’s exact test as appropriate) for categorical variables Table 2Baseline characteristics of surgical residents by study group (N = 40)CharacteristicAI-Guided group (n = 20)Self-learning group (n = 20)p valueDemographicsAge, mean (SD), years30.8 (2.1)31.1 (2.4)0.56Sex, n (%)MaleFemale17 (85.0)3 (15.0)18 (90.0)2 (10.0)0.91Training level, n (%)0.78Junior resident (PGY 1–3)8 (40.0)7 (35.0)Senior Resident (PGY 4–5)12 (60.0)13 (65.0)Prior bariatric clinic exposure, mean (SD), months6.5 (5.5)7.1 (6.2)0.54Clinical contextGastrointestinal surgery specialty, n (%)17 (85.0)18 (90.0)0.81PGY Postgraduate Year, SD Standard Deviation. Notes: Categorical variables are presented as n (%). Continuous variable (Age) is presented as mean (SD). P-values were derived from independent samples t-test for Age, prior clinic exposure, and Chi-square tests (or Fisher’s exact test as appropriate) for categorical variables Baseline characteristics of surgical residents by study group (N = 40) Sex, n (%) Male Female 17 (85.0) 3 (15.0) 18 (90.0) 2 (10.0) PGY Postgraduate Year, SD Standard Deviation. Notes: Categorical variables are presented as n (%). Continuous variable (Age) is presented as mean (SD). P-values were derived from independent samples t-test for Age, prior clinic exposure, and Chi-square tests (or Fisher’s exact test as appropriate) for categorical variables Primary Outcome: Shared Decision-Making The primary outcome, the patient-rated SDM-Q-9 score, was significantly higher in the AI-Guided group compared to the self-learning group (84.7 ± 7.1 vs. 71.3 ± 9.8; mean difference 13.4, 95% CI 9.6 to 17.2, p < 0.01; Cohen’s d = 1.42). This represents a large effect size and exceeds the MCID of 10 points, confirming clinical relevance (Fig. 2A) (Table 3). ANCOVA with PGY level and baseline SSI-BS as covariates confirmed the group effect remained highly significant (F = 28.4, p < 0.001); neither covariate was significant (p > 0.20).Fig. 2Comparative outcomes of ai-guided versus self-learning communication training. (A) Shared decision-making questionnaire (SDM-Q-9) scores (transformed to 0–100 scale). The AI-guided group achieved significantly higher scores, indicating superior patient-perceived involvement in decision-making. Individual data points (jitter plots) are overlaid to show distribution. (B) Communication outline quality scale (CQS) scores. Outlines prepared using the AI-guided VALUE framework were rated significantly higher in structure, values integration, and overall quality. Individual data points (jitter plots) are overlaid to show distribution. (C) Radar chart depicting performance across the four domains of the Four Habits Coding Scheme (4HCS). The AI-guided group (dark green) shows a consistently larger profile area, particularly excelling in Eliciting the Patient’s Perspective and Investing in the End. Numerical scale labels (1–5) and 95% confidence bands are shown on each axis to display variance and magnitude. (D) Radar chart depicting subscale scores of the Decision Conflict Scale (DCS). Lower scores indicate less conflict. The AI-guided group (dark green) demonstrates substantially lower decisional conflict across all five subscales, especially in Values Clarity and Uncertainty. Numerical scale labels (0–100) and 95% confidence bands are shown on each axis to display variance and magnitudeTable 3Comparison of primary, secondary and process evaluation outcomes (N = 40)Outcome measureAI-Guided group (n = 20) mean (SD)Self-learning group (n = 20) mean (SD)Mean Difference (95% CI)p valueEffect size (Cohen’s d)Primary outcomeSDM-Q-9 Score (0–100)84.7 (7.1)71.3 (9.8)13.4 (9.6, 17.2)< 0.011.42Secondary outcomesDecision conflict scale (0–100)Total score19.5 (9.0)32.1 (11.5)−12.6 (−17.1, −8.1)< 0.01−1.14Informed subscale15.2 (10.1)28.4 (13.2)−13.2 (−18.5, −7.9)< 0.01−1.05Values clarity subscale18.8 (11.3)31.6 (14.0)−12.8 (−18.6, −7.0)< 0.01−0.94Support subscale21.1 (12.5)27.5 (13.8)−6.4 (−12.3, −0.5)0.03−0.67Uncertainty subscale22.4 (10.8)30.5 (12.1)−8.1(−13.3 to −2.9)0.03−0.67Effective decision subscale20.1 (11.0)32.5 (13.9)−12.4 (−17.9, −6.9)< 0.01−0.93Four habits coding scheme (4–20)Total scoreHabit 1: Invest in beginningHabit 2: Elicit perspectiveHabit 3: Demonstrate empathyHabit 4: Invest in end17.5 (1.7)4.2 (0.6)4.3 (0.5)4.1 (0.7)4.9 (0.3)14.8 (2.2)3.9 (0.7)3.6 (0.8)3.5 (0.9)3.8 (0.8)2.7 (1.8, 3.6)0.3 (0.0, 0.6)0.7 (0.4, 1.0)0.6 (0.2, 1.0)1.1 (0.8, 1.4)< 0.010.06< 0.010.02< 0.011.230.440.880.611.57Surgeon self-efficacy (SSI-BS, 1–10)Change score (post-pre)Information provisionValues integrationDecision facilitationEmotional support+ 3.5 (1.3)+ 3.2 (1.5)+ 4.1 (1.2)+ 3.8 (1.4)+ 3.0 (1.6)+ 1.8 (1.7)+ 2.0 (1.8)+ 1.5 (1.6)+ 1.9 (1.9)+ 1.7 (2.0)1.7 (1.0, 2.4)1.2 (0.4, 2.0)2.6 (1.9, 3.3)1.9 (1.1, 2.7)1.3 (0.4, 2.7)< 0.010.03< 0.01< 0.010.051.000.661.631.000.60Process evaluationAI Interaction quality (0–13)Communication outline quality (0–15)Structure (0–3)Values Integration (0–4)Medical Accuracy (0–3)Language (0–3)Flexibility (0–2)10.8 (1.5)13.8 (1.6)2.8 (0.4)3.6 (0.6)2.9 (0.3)2.7 (0.5)1.8 (0.4)N/A7.5 (2.3)1.5 (0.8)1.2 (1.1)2.7 (0.5)1.4 (0.9)0.7 (0.7)N/A5.8 (4.9, 6.7)1.3 (1.0, 1.6)2.4 (2.0, 2.8)0.2 (0.0, 0.4)1.3 (0.9, 1.7)1.1 (0.8, 1.4)N/A< 0.01< 0.01< 0.010.11(NS)< 0.01< 0.01N/A2.521.632.180.401.441.57Data are presented as mean (standard deviation) unless otherwise specified. The Mean Difference represents the AI-Guided group mean minus the Self-Learning group mean. For the Decision Conflict Scale (DCS), lower scores indicate less decisional conflict. The DCS total score is calculated as the weighted sum of its 16 items per the original scoring manual (O’Connor, 1995); it is not the arithmetic mean of the five subscales presented here, which are domain averages. Because subscales contain different numbers of items (3,3,3,4,3), equal weighting of subscales yields a different value than equal weighting of items. This explains the apparent numerical discrepancy. Effect sizes (Cohen’s d) are calculated from the mean difference and pooled standard deviation; d = 0.2, 0.5, and 0.8 are considered small, medium, and large effects, respectively. The Surgeon Self-Efficacy in Bariatric Surgery Communication scale (SSI-BS) change score is the difference between post-consultation and pre-intervention scores. AI Interaction Quality (AIIQ) was assessed only in the intervention group. CQS Communication Outline Quality ScaleCI confidence interval, SDM-Q-9 9-item Shared Decision Making Questionnaire, DCS Decision Conflict Scale, SSI-BS Surgeon Self-Efficacy in Bariatric Surgery Communication scale, AIIQ AI Interaction Quality scale, CQS Communication Outline Quality Scale, NS not significantInter-rater reliability (ICC) for observer-rated instruments: 4HCS total ICC = 0.87 (95% CI 0.79–0.93); CQS total ICC = 0.89 (95% CI 0.82–0.94); AIIQ ICC = 0.85 (95% CI 0.76–0.91); SSI-BS rater-coded components ICC = 0.88 (95% CI 0.81–0.93). All ICC values indicate excellent reliability (>0.75). Raters were blinded to group assignment and underwent standardized training with certification requirements before study commencement Comparative outcomes of ai-guided versus self-learning communication training. (A) Shared decision-making questionnaire (SDM-Q-9) scores (transformed to 0–100 scale). The AI-guided group achieved significantly higher scores, indicating superior patient-perceived involvement in decision-making. Individual data points (jitter plots) are overlaid to show distribution. (B) Communication outline quality scale (CQS) scores. Outlines prepared using the AI-guided VALUE framework were rated significantly higher in structure, values integration, and overall quality. Individual data points (jitter plots) are overlaid to show distribution. (C) Radar chart depicting performance across the four domains of the Four Habits Coding Scheme (4HCS). The AI-guided group (dark green) shows a consistently larger profile area, particularly excelling in Eliciting the Patient’s Perspective and Investing in the End. Numerical scale labels (1–5) and 95% confidence bands are shown on each axis to display variance and magnitude. (D) Radar chart depicting subscale scores of the Decision Conflict Scale (DCS). Lower scores indicate less conflict. The AI-guided group (dark green) demonstrates substantially lower decisional conflict across all five subscales, especially in Values Clarity and Uncertainty. Numerical scale labels (0–100) and 95% confidence bands are shown on each axis to display variance and magnitude Table 3Comparison of primary, secondary and process evaluation outcomes (N = 40)Outcome measureAI-Guided group (n = 20) mean (SD)Self-learning group (n = 20) mean (SD)Mean Difference (95% CI)p valueEffect size (Cohen’s d)Primary outcomeSDM-Q-9 Score (0–100)84.7 (7.1)71.3 (9.8)13.4 (9.6, 17.2)< 0.011.42Secondary outcomesDecision conflict scale (0–100)Total score19.5 (9.0)32.1 (11.5)−12.6 (−17.1, −8.1)< 0.01−1.14Informed subscale15.2 (10.1)28.4 (13.2)−13.2 (−18.5, −7.9)< 0.01−1.05Values clarity subscale18.8 (11.3)31.6 (14.0)−12.8 (−18.6, −7.0)< 0.01−0.94Support subscale21.1 (12.5)27.5 (13.8)−6.4 (−12.3, −0.5)0.03−0.67Uncertainty subscale22.4 (10.8)30.5 (12.1)−8.1(−13.3 to −2.9)0.03−0.67Effective decision subscale20.1 (11.0)32.5 (13.9)−12.4 (−17.9, −6.9)< 0.01−0.93Four habits coding scheme (4–20)Total scoreHabit 1: Invest in beginningHabit 2: Elicit perspectiveHabit 3: Demonstrate empathyHabit 4: Invest in end17.5 (1.7)4.2 (0.6)4.3 (0.5)4.1 (0.7)4.9 (0.3)14.8 (2.2)3.9 (0.7)3.6 (0.8)3.5 (0.9)3.8 (0.8)2.7 (1.8, 3.6)0.3 (0.0, 0.6)0.7 (0.4, 1.0)0.6 (0.2, 1.0)1.1 (0.8, 1.4)< 0.010.06< 0.010.02< 0.011.230.440.880.611.57Surgeon self-efficacy (SSI-BS, 1–10)Change score (post-pre)Information provisionValues integrationDecision facilitationEmotional support+ 3.5 (1.3)+ 3.2 (1.5)+ 4.1 (1.2)+ 3.8 (1.4)+ 3.0 (1.6)+ 1.8 (1.7)+ 2.0 (1.8)+ 1.5 (1.6)+ 1.9 (1.9)+ 1.7 (2.0)1.7 (1.0, 2.4)1.2 (0.4, 2.0)2.6 (1.9, 3.3)1.9 (1.1, 2.7)1.3 (0.4, 2.7)< 0.010.03< 0.01< 0.010.051.000.661.631.000.60Process evaluationAI Interaction quality (0–13)Communication outline quality (0–15)Structure (0–3)Values Integration (0–4)Medical Accuracy (0–3)Language (0–3)Flexibility (0–2)10.8 (1.5)13.8 (1.6)2.8 (0.4)3.6 (0.6)2.9 (0.3)2.7 (0.5)1.8 (0.4)N/A7.5 (2.3)1.5 (0.8)1.2 (1.1)2.7 (0.5)1.4 (0.9)0.7 (0.7)N/A5.8 (4.9, 6.7)1.3 (1.0, 1.6)2.4 (2.0, 2.8)0.2 (0.0, 0.4)1.3 (0.9, 1.7)1.1 (0.8, 1.4)N/A< 0.01< 0.01< 0.010.11(NS)< 0.01< 0.01N/A2.521.632.180.401.441.57Data are presented as mean (standard deviation) unless otherwise specified. The Mean Difference represents the AI-Guided group mean minus the Self-Learning group mean. For the Decision Conflict Scale (DCS), lower scores indicate less decisional conflict. The DCS total score is calculated as the weighted sum of its 16 items per the original scoring manual (O’Connor, 1995); it is not the arithmetic mean of the five subscales presented here, which are domain averages. Because subscales contain different numbers of items (3,3,3,4,3), equal weighting of subscales yields a different value than equal weighting of items. This explains the apparent numerical discrepancy. Effect sizes (Cohen’s d) are calculated from the mean difference and pooled standard deviation; d = 0.2, 0.5, and 0.8 are considered small, medium, and large effects, respectively. The Surgeon Self-Efficacy in Bariatric Surgery Communication scale (SSI-BS) change score is the difference between post-consultation and pre-intervention scores. AI Interaction Quality (AIIQ) was assessed only in the intervention group. CQS Communication Outline Quality ScaleCI confidence interval, SDM-Q-9 9-item Shared Decision Making Questionnaire, DCS Decision Conflict Scale, SSI-BS Surgeon Self-Efficacy in Bariatric Surgery Communication scale, AIIQ AI Interaction Quality scale, CQS Communication Outline Quality Scale, NS not significantInter-rater reliability (ICC) for observer-rated instruments: 4HCS total ICC = 0.87 (95% CI 0.79–0.93); CQS total ICC = 0.89 (95% CI 0.82–0.94); AIIQ ICC = 0.85 (95% CI 0.76–0.91); SSI-BS rater-coded components ICC = 0.88 (95% CI 0.81–0.93). All ICC values indicate excellent reliability (>0.75). Raters were blinded to group assignment and underwent standardized training with certification requirements before study commencement Comparison of primary, secondary and process evaluation outcomes (N = 40) Total score Habit 1: Invest in beginning Habit 2: Elicit perspective Habit 3: Demonstrate empathy Habit 4: Invest in end 17.5 (1.7) 4.2 (0.6) 4.3 (0.5) 4.1 (0.7) 4.9 (0.3) 14.8 (2.2) 3.9 (0.7) 3.6 (0.8) 3.5 (0.9) 3.8 (0.8) 2.7 (1.8, 3.6) 0.3 (0.0, 0.6) 0.7 (0.4, 1.0) 0.6 (0.2, 1.0) 1.1 (0.8, 1.4) < 0.01 0.06 < 0.01 0.02 < 0.01 1.23 0.44 0.88 0.61 1.57 Change score (post-pre) Information provision Values integration Decision facilitation Emotional support + 3.5 (1.3) + 3.2 (1.5) + 4.1 (1.2) + 3.8 (1.4) + 3.0 (1.6) + 1.8 (1.7) + 2.0 (1.8) + 1.5 (1.6) + 1.9 (1.9) + 1.7 (2.0) 1.7 (1.0, 2.4) 1.2 (0.4, 2.0) 2.6 (1.9, 3.3) 1.9 (1.1, 2.7) 1.3 (0.4, 2.7) < 0.01 0.03 < 0.01 < 0.01 0.05 1.00 0.66 1.63 1.00 0.60 AI Interaction quality (0–13) Communication outline quality (0–15) Structure (0–3) Values Integration (0–4) Medical Accuracy (0–3) Language (0–3) Flexibility (0–2) 10.8 (1.5) 13.8 (1.6) 2.8 (0.4) 3.6 (0.6) 2.9 (0.3) 2.7 (0.5) 1.8 (0.4) N/A 7.5 (2.3) 1.5 (0.8) 1.2 (1.1) 2.7 (0.5) 1.4 (0.9) 0.7 (0.7) N/A 5.8 (4.9, 6.7) 1.3 (1.0, 1.6) 2.4 (2.0, 2.8) 0.2 (0.0, 0.4) 1.3 (0.9, 1.7) 1.1 (0.8, 1.4) N/A < 0.01 < 0.01 < 0.01 0.11(NS) < 0.01 < 0.01 N/A 2.52 1.63 2.18 0.40 1.44 1.57 Data are presented as mean (standard deviation) unless otherwise specified. The Mean Difference represents the AI-Guided group mean minus the Self-Learning group mean. For the Decision Conflict Scale (DCS), lower scores indicate less decisional conflict. The DCS total score is calculated as the weighted sum of its 16 items per the original scoring manual (O’Connor, 1995); it is not the arithmetic mean of the five subscales presented here, which are domain averages. Because subscales contain different numbers of items (3,3,3,4,3), equal weighting of subscales yields a different value than equal weighting of items. This explains the apparent numerical discrepancy. Effect sizes (Cohen’s d) are calculated from the mean difference and pooled standard deviation; d = 0.2, 0.5, and 0.8 are considered small, medium, and large effects, respectively. The Surgeon Self-Efficacy in Bariatric Surgery Communication scale (SSI-BS) change score is the difference between post-consultation and pre-intervention scores. AI Interaction Quality (AIIQ) was assessed only in the intervention group. CQS Communication Outline Quality Scale CI confidence interval, SDM-Q-9 9-item Shared Decision Making Questionnaire, DCS Decision Conflict Scale, SSI-BS Surgeon Self-Efficacy in Bariatric Surgery Communication scale, AIIQ AI Interaction Quality scale, CQS Communication Outline Quality Scale, NS not significant Inter-rater reliability (ICC) for observer-rated instruments: 4HCS total ICC = 0.87 (95% CI 0.79–0.93); CQS total ICC = 0.89 (95% CI 0.82–0.94); AIIQ ICC = 0.85 (95% CI 0.76–0.91); SSI-BS rater-coded components ICC = 0.88 (95% CI 0.81–0.93). All ICC values indicate excellent reliability (>0.75). Raters were blinded to group assignment and underwent standardized training with certification requirements before study commencement Secondary Outcomes Communication performance: Consultations conducted by the AI-Guided group were rated significantly higher on the four habits coding scheme than those by the self-learning group (17.5 ± 1.7 vs. 14.8 ± 2.2; MD 2.7, 95%CI 1.8 to 3.6; p < 0.01; d = 1.23) (Table 3; Fig. 2C). Analysis of the individual habits revealed the AI-Guided group outperformed the self-learning group in Habit 2: Elicit perspective (4.3 ± 0.5 vs. 3.6 ± 0.8; MD 0.7, p < 0.01, d = 0.88), Habit 3: Demonstrate Empathy (4.1 ± 0.7 vs. 3.5 ± 0.9; MD 0.6, p = 0.02, d = 0.61), and Habit 4: Invest in the End (4.9 ± 0.3 vs. 3.8 ± 0.8; MD 1.1, p < 0.01, d = 1.57). The difference in Habit 1: Invest in the Beginning favored the intervention group but did not reach statistical significance (4.2 ± 0.6 vs. 3.9 ± 0.7; MD 0.3, p = 0.06, d = 0.44). All significant findings remained significant after FDR correction. Decision conflict: The AI-Guided group demonstrated significantly lower decision conflict across all domains of the Decision conflict scale (Table 3). The total DCS score was markedly lower in the intervention group than in the self-learning group (19.5 ± 9.0 vs. 32.1 ± 11.5; mean difference − 12.6, 95% CI −17.1 to −8.1; p < 0.01; Cohen’s d=−1.14). This exceeds the MCID of 10 points for DCS, confirming clinical relevance. The largest absolute reductions were observed in the Informed (MD −13.2, p < 0.01, d=−1.05) and Values clarity subscales (MD −12.8, p < 0.01, d=−0.94). Significant improvements were also seen in the Support (MD −6.4, p = 0.03, d=−0.67), Uncertainty (MD −8.1, p = 0.03, d=−0.67), and Effective decision subscales (MD −12.4, p < 0.01, d=−0.93). All significant findings remained significant after FDR correction. The radar chart visualization (Fig. 2D) illustrates this consistently lower decisional conflict profile across all five dimensions for the AI-Guided group. Surgeon self-efficacy: Residents in the AI-Guided group reported a significantly greater increase in their communication self-efficacy post-intervention compared to the self-learning group (Table 3). The overall SSI-BS change score was + 3.5 ± 1.3 versus + 1.8 ± 1.7 (MD 1.7, 95% CI 1.0 to 2.4; p < 0.01; d = 1.00). Domain analysis showed the largest improvements in value integration (MD 2.6, p < 0.01, d = 1.63) and decision facilitation (MD 1.9, p < 0.01, d = 1.00), with significant gains also in information provision (MD 1.2, p = 0.03, d = 0.66) and emotional support (MD 1.3, p = 0.05, d = 0.60). All significant findings remained significant after FDR correction. Process Evaluation Communication outline quality (CQS): The preparatory consultation outlines created by the AI-Guided group were rated substantially higher on the CQS than those from the self-learning group (13.8 ± 1.6 vs. 7.5 ± 2.3; MD 5.8, 95% CI 4.9 to 6.7; p < 0.01; d = 2.52) (Table 3; Fig. 2B). The intervention group’s outlines excelled particularly in structure (MD 1.3, p < 0.01, d = 1.63), values integration (MD 2.4, p < 0.01, d = 2.18), language appropriateness (MD 1.3, p < 0.01, d = 1.44), and flexibility (MD 1.1, p < 0.01, d = 1.57). The difference in medical accuracy was not statistically significant (MD 0.2, p = 0.11, d = 0.40) and is marked as NS (not significant) in Table 3. AI interaction: Within the intervention group, the quality of interaction with the AI tool, as measured by the AI interaction quality rubric, was high, with a mean score of 10.8 ± 1.5 out of 13 (Table 3). All 20 residents in the AI-Guided Group completed the VALUE prompting exercise as instructed (100% adherence). Mean time spent on AI interaction was 13.2 ± 2.1 min per case. Study staff confirmed that each resident submitted the prompt and received an AI-generated outline before proceeding to consultation preparation. AI Output Analysis: Two bariatric surgeons independently reviewed all 40 AI-generated outlines (20 residents×2 cases) for clinical appropriateness using a structured checklist (accuracy of procedure descriptions, alignment with MBS guidelines, absence of harmful advice). No clinically inappropriate or misleading content was identified. Variance in AI output quality was low (mean score SD = 0.8 on a 1–5 scale). Residents modified AI suggestions in all cases (mean modifications = 4.2 ± 1.5), but no outline was completely discarded. Adverse events: We monitored for potential harms, including resident distress (assessed via post-study debriefing) and evidence of AI-induced over-reliance (review of AI output use). No adverse events or unintended effects related to the intervention were reported by residents or observed by study staff. Exploratory Mediation Analysis: Using the PROCESS macro, we tested whether CQS score (outline quality) mediated the effect of group on SDM-Q-9. The indirect effect was significant (indirect effect = 4.2, 95% CI 1.8–7.1), indicating partial mediation. Discussion This study demonstrates that a structured, AI-Guided metacognitive framework (VALUE) significantly enhances the quality of MBS consultations conducted by surgical residents. The intervention yielded encouraging preliminary findings with large effect sizes that require cautious interpretation given the small sample and immediate assessment, observed across clinician-focused, and process-based outcomes. Most importantly, standardized patients perceived a substantially higher degree of shared decision-making when interacting with residents who used the AI-Guided preparation, as evidenced by the robust 13.4-point increase in the primary outcome (SDM-Q-9) with a large effect size (d = 1.42). Based on prior literature, a change of ≥ 10 points on the SDM-Q-9 is considered the minimal clinically important difference (MCID) [15]; our observed difference of 13.4 points exceeds this threshold, confirming practical as well as statistical significance. Concurrently, the standardized patients experienced significantly less decision conflict, particularly regarding their understanding of information and clarity of personal values. The observed reduction of 12.6 points on the DCS exceeds the established MCID of ≥ 10 points, which is associated with a higher likelihood of patients implementing their decisions [16], further supporting clinical relevance. These higher SDM-Q-9 scores and lower DCS scores as rated by standardized patients were directly mirrored in the observed communication behavior. Blinded raters scored AI-Guided consultations markedly higher on validated measures of communication (as rated by trained observers), with the most profound effects seen in Eliciting patient perspective and investing in the end–the very habits that operationalize shared decision-making. Crucially, this was not merely a performance gain; residents in the intervention group reported a greater increase in their own self-efficacy for critical communication tasks, especially in integrating patient values and facilitating complex decisions. The mechanism for these improvements is illuminated by the process evaluation: the AI-Guided VALUE framework enabled residents to prepare significantly higher-quality consultation outlines that were strategically structured and deeply integrated with patient-specific values, effectively translating preparation into superior in-consultation performance. Furthermore, exploratory mediation analysis confirmed that outline quality partially mediated the effect on SDM-Q-9. The success of this intervention hinges on a deliberate shift in the application of AI: from a source of information to a scaffold for expert cognitive process. We did not deploy AI to converse with the patient or to generate a static script. Instead, we used it to guide the clinician’s preparatory reasoning through the structured VALUE (Validate, Align, Link, Unite) sequence. This approach externalizes the tacit, often chaotic, cognitive workflow of an expert clinician into a replicable, learnable model [19–21]. The AI acted as a metacognitive partner, prompting the resident to first validate the patient’s viewpoint, then explicitly define a shared superordinate goal (e.g., “achieving optimal pregnancy readiness” rather than “choosing sleeve vs. bypass”), before linking medical evidence to that goal. This process forced a values-based re-framing of clinical data. The result, as seen in the superior communication outlines and consultation habits, was a transition from a data-dispensing monologue to a goal-aligned, collaborative dialogue [22]. The AI scaffold ensured that the conversation was engineered not just to be informative, but to be transformative–helping the patient clarify what matters most and see the treatment options as means to achieve their personal ends. This directly explains the significant reduction in the values clarity subscale of decisional conflict observed in our results [23]. Our findings extend the established literature on communication skills training in surgery. Prior studies have demonstrated that focused training can improve specific behaviors, such as empathy or information-giving [24]. However, many programs are resource-intensive, relying on extended workshops and expert feedback, which limits scalability [25–27]. Our AI-Guided model offers a promising, on-demand alternative that demonstrated effect sizes exceeding those commonly reported for traditional workshops [28]. More importantly, current approaches often emphasize discrete skills or knowledge transfer. The VALUE framework, enabled by AI prompting, targets a higher-order competency: the dynamic integration of patient values with complex medical evidence during a time-pressured encounter. This aligns with the core tenets of shared decision-making but provides a concrete, operational pathway to achieve it [29]. Previous applications of AI in MBS care have largely focused on patient education, risk prediction, or postoperative monitoring [30, 31]. Our study innovates by positioning AI as a coach for the clinician’s cognitive and communicative process before the encounter. The dramatic improvement in the quality of prepared outlines (CQS score difference of 5.8 points) underscores that improving the architecture of the consultation plan is a powerful, previously under-leveraged point of intervention. This “prepare with AI” model complements existing “practice with SPs” paradigms, creating a more comprehensive learning cycle. A recent study by Jazi et al. [36] compared expert opinions with ChatGPT-4 in bariatric evaluation, finding that the AI model could generate plausible but not always accurate recommendations, highlighting the need for expert oversight. Their work focuses on AI as a knowledge source for clinical decision support, whereas our study uses AI as a metacognitive scaffold to structure communication rather than to generate clinical judgments. Both studies share the limitation of small, simulation-based designs and call for prospective validation in real-world settings. Together, these complementary approaches suggest that AI may serve dual roles in bariatric care: as an adjunct for clinical information processing and as a tool for enhancing patient-clinician communication. We recognize a fundamental conceptual limitation of our study design. It could be argued that any group receiving structured instruction immediately prior to testing would be expected to outperform an unguided self-learning control group, and that the observed effects may largely reflect immediate recall of recently delivered material rather than superior clinical reasoning or durable skill acquisition. We agree with this critique to a substantial extent. However, the novelty of our intervention lies not in the mere presence of structure, but in the specific metacognitive scaffold (the VALUE framework) designed to externalize expert reasoning and to be internalized by residents. The large effects observed on in-consultation behaviors (4HCS, particularly Habits 2 and 4) and on patient-perceived shared decision-making suggest that the benefit extends beyond rote memorization of a checklist. Nevertheless, without delayed assessments, we cannot definitively separate recall from genuine skill consolidation. Therefore, this study should be viewed as hypothesis-generating: it demonstrates that the VALUE scaffold is feasible and produces encouraging short-term results, but it does not claim to prove superior clinical reasoning or long-term educational benefit. Future studies with delayed testing and active comparator arms (e.g., a different structured framework) are needed to establish whether the VALUE scaffold offers unique advantages over other forms of structured preparation. We acknowledge that other VALUE mnemonics exist (e.g., in palliative care communication); our VALUE framework is specifically designed for MBS consultation planning and should not be confused with those. The differential impact of the VALUE framework across specific domains of communication and decision-making reveals its precise mechanism and clinical utility. The most substantial improvements were observed in eliciting patient perspective (4HCS-Habit 2) and investing in the end (4HCS-Habit 4). This pattern indicates that the AI scaffold did not merely improve general rapport but specifically enhanced the bookends of shared decision-making: deeply understanding the patient’s “why” at the outset and collaboratively constructing the “how” at the conclusion [17, 29]. Similarly, the dramatic reduction in the values clarity and informed subscales of the DCS shows the framework’s success in mitigating two core sources of patient uncertainty: confusion about what matters most to them and lack of understanding about how options relate to those priorities. However, the CQS was structurally designed to mirror the VALUE framework, creating a risk of circularity. Thus, the large CQS effect should not be interpreted as independent validation of the framework; rather, it confirms that residents followed the scaffolded structure as instructed. As detailed in Limitations (point Second), the observed effect sizes are likely inflated due to small-sample bias and shared method variance; therefore, these findings should be interpreted with caution and viewed as hypothesis-generating. In this study, the same standardized patients rated both the SDM-Q-9 and DCS, and the CQS was structurally aligned with the VALUE framework. Therefore, the observed consistency may partly reflect shared measurement design rather than independent educational impact. We have not adequately controlled for this confounding in our original analysis. To address this, we have explicitly downgraded our conclusions: these findings should be viewed as hypothesis-generating preliminary evidence. Independent replication in larger, multi-site trials with blinded independent outcome assessors and measures that are not structurally aligned with the intervention is required before any claim of effectiveness can be made. Practical considerations and risk mitigation: The AI interaction required a mean of 13.2 min per case, which is feasible within surgical workflows. API costs were minimal (<$0.10 per resident), suggesting scalability. To mitigate risks, the following strategy can be taken into considered: (1) Hallucination risk: All AI outputs were reviewed; no clinically inaccurate content was identified. (2) Automation bias/cognitive offloading: The VALUE framework requires residents to actively use the AI output to construct their own plan, not simply memorize. We plan to study a “fading” approach where AI support is gradually reduced. (3) Privacy: All case data were de-identified; no protected health information was entered. (4) Equitable access: AI tools require internet/API access; this is acknowledged as a potential barrier. This has direct, actionable implications. For clinical practice, the VALUE framework offers a structured, efficient method to transform a routine preoperative consult into a values-clarification and alignment session. This is particularly critical in MBS, where long-term optimal clinical response depends on a patient’s understanding and commitment to a lifelong behavioral and medical plan [32]. For surgical education, this AI-Guided approach provides a scalable, high-fidelity method to train the integrative reasoning required for complex chronic care. It moves beyond teaching “what to say” to modeling “how to think” during a consultation. Residency programs can implement this as a preparatory tool for clinic rotations, ensuring residents enter patient encounters with a strategically planned approach rather than relying on ad-hoc, information-heavy discussions. The intervention is designed as a scaffold for learning, not a replacement for clinical reasoning. We propose future “graduated withdrawal” designs where AI support is reduced as competency increases, and we recommend supervisor review of AI-generated plans during training. Limitations and Future Research Directions This study has several limitations. First, the sample size (n = 20 per group) is small; the power calculation was based on a pilot of only five participants per arm, which cannot provide stable variance estimates. Consequently, the study is powered only to detect large effects, and the observed effect sizes (d = 1.42 for SDM‑Q‑9, d = 2.52 for CQS) may be inflated due to the winner’s curse – a well‑documented phenomenon in small‑sample studies. These results should therefore be considered hypothesis‑generating rather than definitive evidence of efficacy. Second, the observed effect sizes (d = 1.42 for SDM-Q-9, d = 2.52 for CQS) are unusually large for a single 15-minute structured preparation exercise, given that established communication training programmes typically yield effect sizes of d = 0.3–0.6. With n = 20 per group, the standard error of Cohen’s d is approximately 0.33, meaning the true population effect could be substantially smaller – a phenomenon known as the winner’s curse in small-sample studies. Moreover, consistency across multiple outcome measures does not rule out shared method variance: the same standardized patients rated both the SDM-Q-9 and DCS, and the CQS was structurally aligned with the VALUE framework. Thus, the observed pattern may partly reflect measurement design rather than independent educational impact. Independent replication in larger, multi-site trials with blind independent outcome assessors and measures not structurally aligned with the intervention is essential before any claim of effectiveness can be made. Third, the study evaluated a specific implementation of an AI model (DeepSeek‑V3.2) guided by a specific framework (VALUE); generalizability to other models or communication structures requires verification, and model updates over time could alter outputs. Fourth, its findings are based on simulated consultations with standardized patients. While SPs provide consistency and enable blinded assessment, they cannot fully replicate the emotional gravity and long-term relationship of a real clinical encounter. Because this study used simulated consultations rather than real clinical encounters, no conclusions can be drawn about patient-centeredness in real-world settings. All references to “patient-centered” outcomes in this manuscript refer specifically to perceptions of standardized patients under controlled conditions. Fifth, the study assessed residents immediately after the 15-minute preparation phase, with no access to preparation materials during the consultation. No delayed assessment (e.g., 1 week, 1 month, or longer) was conducted. Therefore, this study captures short-term performance and immediate recall of the structured framework, not durable learning or long-term skill consolidation. Moreover, because the intervention group received structured preparation while the control group used unguided self-directed learning, the observed advantage may partly reflect the simple presence of structure rather than the specific content of the VALUE framework. An active comparator arm (e.g., a different structured framework) would be needed to isolate the unique contribution of VALUE. Conclusions about long-term educational effectiveness cannot be drawn from this design. Future studies must include delayed assessments (e.g., 3-month and 12-month follow-up) to evaluate retention and transfer of skills. We have planned a 12-month pragmatic trial to address this gap. Sixth, because SPs rated both the SDM‑Q‑9 and DCS, and the same raters scored multiple instruments, shared method variance may have contributed to the consistent pattern; however, observer‑rated 4HCS scores were obtained from independent blinded raters using a well‑validated instrument, providing some protection against this bias. Seventh, because of the nature of the intervention, residents could not be blinded to group assignment; this lack of blinding may have introduced performance bias, as AI‑guided residents might have felt more prepared or confident. However, the observer‑rated outcomes (4HCS) and the blinding of SPs and raters mitigate this concern to some extent. Eighth, the CQS was structurally designed to mirror the VALUE framework, introducing a degree of circularity; therefore, the large effect observed on CQS may partly reflect alignment between the assessment tool and the intervention rather than independent skill acquisition. Future studies should validate the CQS against established communication measures. Ninth, with only 20 residents per group, normality testing (Shapiro‑Wilk) has low power; thus, the assumption of normality remains a pragmatic choice. Tenth, the exploratory mediation analysis should be interpreted with caution given the small sample size and the quasi‑experimental design; the indirect effect estimate (4.2, 95% CI 1.8–7.1) is imprecise, and causal mediation assumptions cannot be fully met. Eleventh, although randomization was stratified by PGY level and prior bariatric clinic exposure was measured and balanced between groups, prior exposure to bariatric communication training from earlier years (more than one year before the study) could not be fully controlled. This may introduce residual confounding, particularly for senior residents who may have acquired relevant skills through routine clinical exposure rather than formal courses. Finally, a single institution and a specific resident population limit generalizability. Future research must address these gaps. Priority should be given to a pragmatic trial in live clinic settings to assess the framework’s impact on real patient decisions, satisfaction, and postoperative adherence. We have planned a 12-month follow-up pragmatic trial to evaluate long-term retention of communication skills and impact on real-world patient outcomes, including adherence to postoperative follow-up and optimal clinical response. Research should also investigate the optimal “dose” and integration of this tool into longitudinal training curricula, including “graduated withdrawal” designs. Furthermore, exploring the explainability of the AI’s suggestions–helping residents understand why the AI proposes a certain line of questioning or reframing–could deepen learning and foster critical adoption rather than rote use [33]. Finally, the model should be tested with attending surgeons to determine if it also enhances the practice of experienced clinicians, potentially serving as a cognitive aid for complex decision-making. Conclusions Our findings suggest that the AI-guided VALUE framework is a promising hypothesis-generating strategy for improving shared decision-making in MBS consultations. By scaffolding the clinician’s preparatory reasoning, the intervention produced encouraging preliminary improvements in patient-perceived collaboration, decisional conflict, observed communication skills, and surgeon self-efficacy. However, independent replication in larger, multi-site, pre-registered trials with delayed outcome assessment and measures not structurally aligned with the intervention is required before any claim of effectiveness can be made. This “prepare with AI” model should currently be viewed as a hypothesis-generating approach that warrants further investigation. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 (DOCX 28.7 KB) Supplementary Material 1 (DOCX 28.7 KB) Supplementary Material 2 (DOCX 22.7 KB) Supplementary Material 2 (DOCX 22.7 KB) Supplementary Material 3 (DOCX 22.2 KB) Supplementary Material 3 (DOCX 22.2 KB) Supplementary Material 4 (DOCX 23.1 KB) Supplementary Material 4 (DOCX 23.1 KB) Supplementary Material 5 (DOCX 21.3 KB) Supplementary Material 5 (DOCX 21.3 KB) Supplementary Material 6 (DOCX 25.4 KB) Supplementary Material 6 (DOCX 25.4 KB) Supplementary Material 7 (DOCX 26.9 KB) Supplementary Material 7 (DOCX 26.9 KB) Supplementary Material 8 (DOCX 22.0 KB) Supplementary Material 8 (DOCX 22.0 KB) Supplementary Material 9 (DOCX 23.5 KB) Supplementary Material 9 (DOCX 23.5 KB) Supplementary Material 10 (DOCX 25.5 KB) Supplementary Material 10 (DOCX 25.5 KB) Supplementary Material 11 (DOCX 25.7 KB) Supplementary Material 11 (DOCX 25.7 KB) Supplementary Material 12 (DOCX 26.0 KB) Supplementary Material 12 (DOCX 26.0 KB) Supplementary Material 13 (DOCX 22.3 KB) Supplementary Material 13 (DOCX 22.3 KB)