Chatbots are defined as conversational agents powered by artificial intelligence (AI) designed to replicate human interaction via text, speech, and visual forms of communication, representing a novel solution to limitations regarding user experience [1,2]. Conversational agents can be designed to leverage machine learning and natural language processing to deliver personalized, interactive support to users, addressing their unique needs and preferences [1]. A recent meta-analysis highlights the effectiveness of chatbots, particularly text-based ones, in promoting lifestyle behaviors, including fruit and vegetable (F&V) consumption [3]. Moreover, these interventions have been shown to be effective across populations and age groups, reinforcing their potential for scalability and adaptability [3]. However, despite these promising findings, there remain gaps in the development, testing, and evaluation of conversational agent interventions designed specifically for high-risk, multicultural, low-income populations. These populations may face unique barriers when engaging with existing chatbots, including limited health or digital literacy, lack of linguistic accessibility, and reduced trust in automated health technologies that are not culturally or contextually tailored [1,2]. For example, some nutrition chatbots are designed for English-speaking users with high digital literacy and may not adequately address the needs or the realities of pregnant individuals navigating food access through Food is Medicine (FIM) programs [3].

Digital health innovations benefit from being grounded in human-centered design (HCD), a methodology that emphasizes iterative development through empathy, stakeholder involvement, and usability testing to ensure that tools are both acceptable and effective. In health care, HCD has been shown to improve user adoption by aligning interventions with the unique contexts and capabilities of patients and health care providers (such as a parent, nurse, or physician) [4,5]. However, even well-designed interventions often require refinement once deployed in real-world settings. To address this, the plan-do-study-act (PDSA) cycle is frequently used as a continuous quality improvement framework that supports rapid testing of targeted changes, measurement of impact, and structured adaptation [6]. Combining HCD during initial development with PDSA for ongoing evaluation creates a complementary approach that allows digital health tools such as conversational agents to evolve responsively to user feedback and use patterns over time.

One potential area for testing conversational agents is nutrition education to support interventions that promote diet quality. FIM interventions are a set of interventions that are integrated into the health care system and provide consistent access to fresh, healthy food, including produce prescriptions, medically tailored groceries, and medically tailored meals, to address the critical link between nutrition and health [7,8]. These interventions may also include nutrition education to support healthy habits.

To address the need to support FIM programs with culturally relevant and tailored nutrition information, this paper details the design of a bilingual, AI-driven nutrition conversational agent to support an FIM program for women with high-risk pregnancies using HCD and PDSA methods. This paper reports on the iterative refinement process and preliminary use patterns, specifically the development and refinement phases that shaped the conversational agent, Flora. The study design and evaluation of behavioral and clinical outcomes are reported elsewhere.

The conversational agent, Flora, was developed as a key component of an ongoing 2-arm RCT to assess the implementation and effectiveness of conversational agents on improving use and consumption of F&V among high-risk pregnant women who receive care at Harris Health obstetric clinics in Houston, Texas. The RCT randomly assigned 200 women to the intervention and control groups with 1:1 allocation.

Eligibility criteria were as follows: (1) receipt of prenatal care at Harris Health obstetric clinics, (2) age between 18 and 44 years, (3) medically confirmed viable pregnancy of 20 weeks or less, (4) designation as high risk (high-risk criteria include age of ≥35 years, overweight or obesity [BMI≥25.0 kg/m² at prepregnancy self-report], or prior history of pregnancy hypertension or gestational diabetes), (5) residence within the zip code food delivery radius, and (6) confirmed access to a smartphone capable of receiving SMS text messages.

This paper reports on the design, development, and iterative refinement of the chatbot embedded within a single, ongoing 2-arm RCT. HCD activities took place prior to trial deployment and involved community advisory group (CAG) members and implementation stakeholders who were not part of the RCT. Postdeployment PDSA cycles and informal question-and-answer (Q&A) sessions were conducted exclusively with participants randomized to the intervention arm to inform real-time refinements to the chatbot. Clinical and behavioral outcomes from the RCT will be reported separately.

Participants were required to be able to provide informed consent and communicate in either English or Spanish, the 2 languages supported by the chatbot. Informed consent was obtained verbally via phone call by credentialed community health workers prior to enrollment. Additionally, participants were required to use their own devices; smartphones and data plans were not provided by the study. Limited technology experience was not an exclusion criterion.

To support participation, all intervention participants received a personalized link to access the chatbot via SMS text message, an instructional video demonstrating how to use Flora, and follow-up phone support from a trained community health worker. Throughout the study, participants also received reminder SMS text messages and phone calls encouraging chatbot use. Participants were compensated for their time and participation with a total of US $50 in gift card incentives distributed over the course of the study.

To protect participant privacy and confidentiality, all data were deidentified for analysis and presentation. Audio files, transcripts, and analytic files were stored on University of Texas Health Science Center at Houston’s (UTHealth Houston) secure SharePoint within a restricted-access folder, with access to any identifiable information strictly limited to designated personnel in accordance with institutional privacy and security requirements.

This study was approved by the Committee for the Protection of Human Subjects at the UTHealth Houston (HSC-SPH-23-0734). The trial is registered with ClinicalTrials.gov (NCT07165990).

With feedback from our stakeholders, the next step was to co-design the conversational agent to enhance participants’ use of the FIM program (measured via self-report of F&V intake) by engaging in SMS text message conversations and delivering personalized nutrition information relevant and conducive to behavior change [3].

To achieve this, the conversational agent dialog system included both relational capacities to establish and maintain a relationship with the user and persuasive conversational capacity to modify their behavior [11]. We incorporated evidence-based behavior change techniques from social cognitive theory [12] and the health belief model [13]. We used constructs such as self-efficacy, perceived benefits, knowledge, and social support to motivate users to engage with the technology [14,15].

To engage the users and “humanize” the chatbot [16,17], we gave it the name Flora. This name was selected via a brainstorming session where our stakeholders suggested female names that could be pronounced in English and Spanish that would be relatable to the users, who were predominately Hispanic female individuals. By applying these behavioral strategies, we aimed to improve both adoption of and sustained engagement with the chatbot, ensuring meaningful evaluation of its impact on dietary outcomes.

To assess usability and refine the chatbot, we engaged both a CAG and program participants in iterative evaluation. CAG meetings were conducted as formative co-design sessions and focused on general perceptions, preferences, and design feedback rather than generating formal qualitative data. These sessions were recorded, and structured note-taking took place to capture key usability issues and suggestions. The CAG provided structured feedback on chatbot design features, including naming, tone, and cultural tailoring.

Moreover, informal usability testing and Q&A sessions with program participants were conducted to explore perceptions of accessibility, relevance, and overall experience. These sessions were designed to obtain rapid, implementation-focused feedback rather than to produce formal qualitative findings. Feedback from participants was independently reviewed and coded by 2 members of the research team to identify recurring barriers to and facilitators of chatbot use. Discrepancies were resolved through discussion until consensus was reached. Participant feedback was documented through structured notes and summarized to inform iterative refinements to chatbot content, prompts, and engagement strategies.

In addition to the structured HCD process, we anticipated the need for real-world refinements once Flora was deployed. To support this, we incorporated PDSA cycles as part of our continuous quality improvement approach. PDSA cycles, widely applied in implementation science [18], are designed to capture user experiences in practice and rapidly adapt interventions. Two PDSA cycles were planned for the postdeployment period between July 2024 and March 2025. The cycles were designed to identify usability barriers, inform iterative modifications, and document facilitators of engagement.

The iterative refinement activities involved women enrolled in the FIM program who were largely Spanish preferring and of Hispanic or Latino ethnicity, reflecting the intended target population for the chatbot. Among participants engaged in follow-up Q&A sessions (n=32) who used the chatbot once or not at all, all were female (32/32, 100%); 88% (28/32) identified as Hispanic or Latino; 90% (29/32) preferred Spanish; and the mean age was 33 (SD 6.18) years, with a mean gestational age of 26.5 (SD 4) weeks.

In February 2025, while the study was ongoing, the research team at UTHealth Houston conducted Q&A sessions with the FIM program participants via phone call. These Q&A sessions were conducted to contextualize early engagement patterns observed during deployment and inform interpretation of subsequent refinement activities.

To conduct these sessions, we developed a script that included questions regarding the following topics: confirming access to the chatbot, identifying challenges, providing technical support, encouragement to use Flora, and asking for feedback to improve their experience. Among participants contacted for Q&A sessions, themes related to low urgency and competing priorities were the most frequently reported, followed by confidence in cooking skills and low technology self-efficacy (Table 1). These findings are presented descriptively to characterize early barriers to engagement rather than to generate formal qualitative conclusions.

The RCT this study was based on was designed to evaluate the feasibility and preliminary effectiveness of 2 nutrition education strategies associated with an FIM intervention among low-income, high-risk pregnant women in a clinical setting. To our knowledge, this is the first RCT to compare standard FIM nutrition education with an enhanced strategy that integrated an AI-driven nutrition chatbot. This approach allows for a rigorous assessment of health and behavioral outcomes as well as participant engagement and satisfaction with this technology.

The feedback from participants shows the promise and challenges of implementing a conversational agent within an FIM program. The emergent themes and subthemes provide valuable insights into barriers to adoption, such as low motivation or perceived lack of necessity, as well as facilitators, including openness to digital support. These findings point to several concrete opportunities for further tailoring conversational agent–based nutrition education within FIM programs. For example, for participants who reported high confidence in cooking skills and a perceived lack of need for chatbot support, future iterations could emphasize more advanced, novel, or pregnancy-specific nutrition content, such as food safety during pregnancy or creative uses of F&V with different spices or cooking techniques. For participants with low technology self-efficacy, engagement may be improved by simplifying access through alternative modes such as SMS text messaging–based interactions sent proactively from the chatbot rather than passively expecting the user to interact with the technology. Additionally, more frequent phone-based touchpoints could be useful to troubleshoot use issues. Finally, for participants facing competing priorities and limited time, the chatbot could be adapted to deliver concise, time-efficient interactions or embedded into existing routines, such as prenatal appointment reminders or produce delivery notifications. These strategies show how conversational agents can be tailored to different user profiles to enhance relevance, accessibility, and engagement.

Within the broader literature, personalized nutrition education remains limited in several FIM programs, and studies of AI-supported nutrition tools often lack rigorous comparative designs or adequate attention to language and cultural context. For example, reviews demonstrate that, while AI chatbots are growing in health promotion, there is a lack of rigorous RCT evidence demonstrating their effectiveness in modifying and maintaining dietary behaviors for weight management or chronic disease prevention in adults [19]. Additionally, literature on personalized nutrition education increasingly shows how AI-driven tools can enhance engagement when they are culturally and linguistically adapted [20]. This study addresses these gaps by implementing a bilingual, culturally adaptable AI chatbot nested within an RCT framework aimed at high-risk pregnant women participating in an FIM program. Grounded in behavior change theory, specifically social cognitive theory and the health belief model, this intervention integrates personalization, language choice, and interactive technology into a longer-duration evaluation framework [12,13].

Finally, this work contributes to the literature on clinic-based implementation strategies. As highlighted in recent research [21], successful program adoption depends on integration into existing workflows and staff practices. Demonstrating the feasibility and health impacts of an FIM program enhanced with AI technology has important implications for health policy. If effective, this model could inform broader adoption of conversational agents to support nutrition education and preventive care among vulnerable populations.

This study has several strengths. The use of an RCT is a strength because it allows us to compare 2 different nutrition education strategies in a high-risk population of pregnant women. The evaluation data will be used to assess factors associated with implementation and adoption of these strategies. In addition, strong interinstitutional collaboration among the community, academia, and health care providers was critical to inform the design, implementation, and evaluation of implementation success.

This paper focuses on design, implementation, and iterative refinement rather than formal qualitative or engagement outcome analyses. The qualitative findings are presented descriptively to inform usability and implementation improvements and should not be interpreted as stand-alone qualitative results. A key limitation of this study is that Flora was built on a general-purpose language model (GPT-3.5), which may generate biased or incomplete responses.

To mitigate technology-related risks during the study, the research team implemented ongoing monitoring and quality assurance procedures. Chatbot interaction logs were reviewed on a weekly basis by the study team to identify potentially inaccurate, inappropriate, or unclear responses. Throughout the study, no major data safety issues were identified. However, system instructions were iteratively refined to improve accuracy, clarity, and alignment with evidence-based nutrition guidance. This process allowed for real-time oversight while preserving the scalability advantages of an AI-driven conversational agent in a clinical setting.

Emerging guidance on the use of large language models in health care emphasizes the importance of accuracy, transparency, and human oversight [22,23]. Future iterations of conversational agents such as Flora could further mitigate risks by incorporating model fine-tuning on verified, evidence-based nutrition content or using retrieval-augmented generation approaches that constrain responses to curated clinical or nutrition knowledge sources [24]. These strategies have been proposed as effective approaches to enhance reliability and reduce the likelihood of inaccurate outputs when deploying large language models in health-related contexts [25].

Another study limitation is that the web page–based implementation relied on participant-initiated engagement, which may have contributed to low overall use. Finally, reliance on self-reported measures may be subject to recall or social desirability bias; however, objective chatbot use data and process evaluation measures were included to provide additional context for interpretation and future research.

This study demonstrates the feasibility of integrating a bilingual AI-based chatbot into an FIM program among high-risk pregnant individuals in a clinical setting. The iterative design and implementation identified engagement barriers, including competing priorities, technology self-efficacy, and perceived relevance while assessing the value of the human-centered approach and continuous quality improvement throughout the program. Lessons learned can inform the optimization of this program and provide guidance for future FIM programs seeking to incorporate conversational agents as scalable, culturally responsive nutrition education tools.