A clearly defined goal is the first step in the design process and the foundation that will guide the development and evaluation of the CA. This step consists of 3 interlinked areas of evaluation—completing a thorough needs assessment, defining the aim, and characterizing the end user and objectives—which, in turn, determine the parameters to be tested and reported. The CA goal was described in 64% (35/55) of the papers in our reviews [3,5,25-28,65,68-76,81-99].

The design process should commence with an in-depth needs assessment to understand existing gaps that may be filled by the CA. These may be informed by a literature review [83,90,91,96,100] to assess potential research areas and the needs and challenges of the target population, including not only patients but also caregivers, health care providers (HCPs), and other experts [25,26,87,89,95,98]. Researchers should also involve end users in this initial phase by using surveys and a variety of qualitative methods [70,87] such as in-depth interviews and focus group discussions to gather their views.

Aligned with the needs assessment, the design team should formulate clear, attainable, and relevant objectives to drive the CA design and development process. It is important to consider the CA temporal profile, which characterizes 4 types of CAs according to the type and frequency of CA-user dialogs [101]. The CA temporal profile will also determine the type of objectives included, broadly classified as short term or long term [101]. A short-term objective refers to an outcome to be completed as soon as the interaction with the CA ends, such as medication reminders [30]. A long-term goal would involve several CA-user interactions being completed over a period, as in mental health interventions to promote mental well-being in the general population [3] or young people with cancer [70]. Complex CA interventions may include short- and long-term goals, such as CA Vik [30] providing medication reminders (short-term goal) and health education (long-term goal) to patients with breast cancer. Furthermore, Kowatsch et al [73] used prompts and reminder SMS text messages to enhance children’s discipline and routine, which are essential for the self-management of asthma.

The next important design consideration is to determine the target population. An initial assessment should establish whether the CA will be offered to healthy users or individuals with a specific medical condition, caregivers, or HCPs. It is important to generate a detailed and accurate portrayal of the target user, including gender, age group, cultural beliefs and socioeconomic concerns, digital and health literacy, access to digital devices, and smartphone penetration rate. If the intervention is educational, a knowledge test should be implemented [73]. The acceptability of CAs by the target population and the perceived risk of using a CA for health care matters should be evaluated, particularly for severe or highly stigmatizing conditions [102] such as mental health disorders [103,104].

User interaction with CAs appears to be enhanced when the CA displays a well-defined, positive, and empathic personality [105,106]. In general, giving a name and profile picture to the CA may enhance its social presence and user acceptance [107], although its effect appears to be small [106]. In health care settings, using a human-like avatar rendering realistic features, including medical attire, may increase user satisfaction [105], although avatars displaying highly realistic features may upset users and decrease engagement, an experience referred to as the “uncanny valley” [108].

Studies have consistently shown that CAs displaying empathy, relational behavior, and self-disclosure enhance the user experience [78,105] and increase the working alliance [109]. Conversely, users would notice if the CA did not convey empathy [69].

Acceptability may be further enhanced if the CA design acknowledges the specific cultural or demographic traits of the target population [73] or offers options to personalize the interface (eg, offering a male and female persona) [31,73]. Alternatively, CAs may explicitly disclose their identity [70] to reduce user expectations about their capabilities. Finally, CA personality should align with its intended function. For example, health care CAs often display one of two personality types: a more approachable, empathic coach-like personality, particularly if delivering behavior change interventions [77] supporting self-management of chronic disorders [73,78] and mental health conditions [3], or a health care professional persona to emphasize the legitimacy of the CA and its content [1].

The language recommended for text-based interventions should be encouraging, positive, friendly, polite, and light-hearted and may include light humor while at the same time being formal [110]. To maintain the flow of the conversation, it may be advisable to use visual cues such as successive moving dots signaling that the CA is “typing” the next message.

The text should be written in clear, short sentences using simple language and avoid scientific jargon. The National Institutes of Health recommends that patient education materials be written at or below the sixth-grade reading level (ages of 11 and 12 years) to reach a diverse range of individuals with varying levels of literacy [111]. The readability of the text can be assessed using a scale such as the Flesch-Kincaid grade level to determine its suitability [112]. Furthermore, the CA should use the target population’s native language in its communications [75] and, if needed, the conversational content may be translated to one or more languages, particularly if the CA will be deployed in multiethnic, multilanguage societies.

With regard to the tone of the conversation, despite the text-based nature of the CA, it may be advisable to simulate more casual, verbal speech while avoiding the use of “textese” [113], a form of abbreviated written or typed language characterized by unconventional spelling and grammar (eg, “tonite” instead of “tonight”) and abbreviations and contractions (eg, “pls” instead of “please” or “wanna” for “want to”) [114]. Furthermore, words written in full capital letters should be avoided as they equate to shouting [110].

Emojis may be used to articulate emotions or other expressions more efficiently than text [70]. However, emojis are vulnerable to varied interpretations across cultures and contexts and should be used mindfully. Fadhil et al [115] noted a context-specific nature of emojis whereby they increased efficacy in a mental health intervention but did not help in promoting physical well-being.

CAs designed to address sensitive topics such as HIV and AIDS, sexually transmitted infections, or mental health disorders may emphasize the confidential nature of the messages or include code words to protect users’ privacy. This is particularly relevant in low- and middle-income settings, where family members may share a single smartphone [116].

Conditional to the CA’s aim, the design and development team may consider a “hybrid” intervention where the interaction with the CA would be complemented by regular interactions with HCPs offering timely feedback on a self-management technique or regular support and motivation [33,72,73]. Alternatively, as presented in the study by Stasinaki et al [79], the CA may be fitted with multiple channels, where the user can converse with the CA in one channel and directly with an HCP in another.

Peer support is recognized to play an important role in promoting adherence to self-management interventions [117] and a further point of human involvement to be considered. The CA intervention may include an additional communication channel for users to interact, share experiences, and receive peer support. For example, Wang et al [75] developed a WeChat intervention to support smoking cessation where the CA not only responded to individual users but also acted as a group moderator.

CAs may be delivered through a variety of channels, such as stand-alone apps [3,73] and existing messaging platforms [68,71,75] such as Facebook Messenger, Telegram, WeChat, and WhatsApp, or embedded in a website [69]. Each channel possesses its own set of complexities, and the decision regarding the delivery channel should be based on the target population needs and the expertise of the CA development team [118]. If the research team does not include app developers or computer scientists, the CA may be embedded in a messaging platform or may be developed using a CA development platform that offers templates or other design solutions for individuals with no previous programming knowledge [5,118], such as Chatfuel, ManyChat, and others. CAs are generally web-based, and some of these platforms are free of charge. Alternatively, if the team expertise or project budget allows, the CA may be delivered through a stand-alone app. This approach offers design flexibility, such as a variety of data collection sources including smartphone sensors, health programming interfaces, connected medical devices, and patient self-reported data [119]. The combination of subjective patient reports with objective, real-time data may reduce users’ responsibility to update their progress and at the same time receive relevant, dynamic coaching based on the current data [120], which in turn may increase adherence to the intervention.

In addition, factors associated with the target population may also affect the selection of the most suitable delivery channel and operating system (eg, Android or Apple’s iOS). For example, Kamita et al [71] implemented their CA on the messaging platform “LINE” as it was the most popular social network service in Japan, and Wang et al [75] selected WeChat, the most common messaging app in Hong Kong.

Aligned with the framework focus, text would be the CA’s main input and output modality. Messages should be brief, fit the mobile screen without scrolling [69], and be of an adequate font size to allow for comfortable reading. Moreover, if the CA targets populations for whom reading might be challenging, such as older adults or visually impaired individuals, text-to-speech assistive technology may be incorporated into the app.

Visual aids such as images or videos are useful to adapt content to audiences with lower educational attainment [121], deliver personal narratives relevant to the end users (eg, young people with cancer), or decrease the amount of textual information [76]. When using multimedia content, it is important to use high-resolution files to avoid pixelated or blurred images. Furthermore, if pictures are obtained from the web, developers should abide by copyright regulations and either source the pictures from free stock photo repositories, acquire the image rights, or produce the images in-house.

All health-related information included in the CA should be derived from reputable sources and adequately referenced. Sources of evidence-based information include comprehensive literature reviews; clinical practice guidelines; Cochrane systematic reviews; and reputable organization websites such as the World Health Organization, MEDLINE Plus, and the Centers for Disease Control and Prevention in the United States or the National Health Service Health A to Z in the United Kingdom [65]. For example, Kowatsch et al [73] used evidence from multiple sources such as published literature on the improvement of asthma management in children [122], technology acceptance research [123], and user-CA working alliances [124] to inform their intervention for asthma management.

Another important aspect of content development is to ensure an adequate understanding of user requests, particularly for potentially serious or life-threatening health conditions. Safeguards to be implemented within the dialog include the request for clarification if the CA receives an unfamiliar input or directing the user to contact an HCP or a human administrator [125,126]. These strategies were included in TensioBot, an intervention to facilitate self-measurement of blood pressure where, after obtaining confirmation of a blood pressure measurement value outside the normal range, the CA alerted the attending physician [68]. Important strategies to manage unintended errors include using validated data entry fields; limiting the data input to predetermined number ranges, words, or characters; or including predefined options for the user to select.

In general, health care CAs should include a disclaimer clearly stating that the intervention “does not replace healthcare provider’s advice.” Furthermore, in the case of health conditions associated with rapid deterioration of patient status leading to medical emergencies, such as cardiovascular conditions, diabetes, chronic pulmonary disorders, or mental health conditions that increase the risk of suicide, information should be included to assist users in managing an emergency situation, such as the provision of emergency services or crisis helpline telephone numbers [127], links to contact their primary physician, or clear advice on first aid treatments such as offering a sugary drink to manage a hypoglycemic event in a person with diabetes [128].

The content and style of the messages should be aligned with the health condition and CA aim. Broadly, the messages may be educational [30,78] or motivational [34,77,79] or deliver reminders to perform a self-management task [68], input data [77], comply with preset tasks [73], take a medication, or attend an HCP appointment [68]. For CAs tasked with engaging with the user during clinic visits, it may be useful to include a status report or summary of the consultation [126].

CAs assuming a coach-like persona might emphasize sympathy, empathy, and participants’ achievements [78]. Interventions attempting to modify users’ behavior may deliver messages with higher emotional content, as reported in the study by Carfora et al [80], where only emotional messages led users to reduce red meat consumption. In addition, the Wang et al [75] CA used 4 types of messages to deliver a smoking cessation intervention: group announcements, health-related information, reminders to share positive results and progress, and fixed answers to frequently asked questions or requests.

CAs are increasingly used to promote behavior change [1,129]. Behavior change interventions are complex [130] and often comprise one or more behavior change techniques (BCTs) to induce change. In our assessment, 4% (2/55) of the studies used a behavior change theory to guide the intervention design, including the Health Action Process Approach [78] and the technology acceptance model [71]. In addition, 13% (7/55) of the studies [31,72-75,77,80] reported the use of specific BCTs such as goal setting, self-monitoring, tracking and feedback, social support, use of rewards, and anticipated regret.

For example, a study described a multicomponent behavior change intervention incorporating several BCTs, such as goal setting, self-monitoring, stimulus control, and behavioral contract, to support a healthy lifestyle for adolescents with obesity [34,79]. Furthermore, including group chats where peers or HCPs offer relevant information and emotional support may also assist in promoting positive behavior change, such as using a CA-led WeChat peer group to promote smoking cessation [75].

Depending on the purpose of the CA, it may be appropriate to integrate data from external devices such as glucometers [131] or activity trackers [119]. Alternatively, access to smartphone sensor data [132] may facilitate passive monitoring of the user’s activity [79] or determine novel digital biomarkers to assess the user’s mood [133] or disease status [134]. The use of smartphone sensors for passive monitoring may further allow for real-time information sharing with HCPs, caregivers, or peers, a feature that may be particularly useful to monitor older people living alone, who may be at higher risk of falling, or individuals with severe chronic illnesses and multiple hospital admissions.

For a good conversation flow, the predefined answer options should be sufficient and appropriate to align with the user intent, defined as the user goals or intentions in each conversation turn. Constructing a mind map outlining the different facets associated with a topic (eg, medication adherence) and the likely influencing factors (lifestyle components or emotional state) would help predict the most relevant answer options to provide to the user [135].

A mind map is a diagram representing concepts, ideas, or tasks generated from a key concept, which is generally represented in the center of the graph [136]. Mind maps are an effective method of brainstorming [137] that can be applied to building the conversation flows. Several web-based programs and platforms are available to organize the conversation flow, including tools specifically designed to build the CA conversation, such as SAP Conversational AI [138] or MobileCoach [35]. Conversation flows may also be built using nonspecific mind mapping software such as Xmind [139]. Mind mapping is useful to assist in recording the flow of conversations between different topics or different user interactions. A well-constructed conversation flow leads the conversation, guides the user, and can address all relevant questions about its purpose. Furthermore, interactivity, personalization, and consistent messaging have been noted as valued qualities [140].

Interventions should be tailored to individual participant needs [110]. When compared with generic CAs, context, situational, or individually aware agents promote a more positive user experience [132]. Personalized interventions include addressing the user by their name or nickname [141]; delivering notifications and reminders tailored to individual needs [110], such as medication or appointment reminders; and notifications for missed activities or unread messages [30,78]. For example, an intervention promoting self-management of chronic pain offered personalized content based on the user’s type and duration of pain and personal interests [78].

An important caveat involves the design of interventions offering personalized advice based on user measurements, such as suggesting a treatment based on individually reported data (eg, blood glucose levels or blood pressure readings), as these interventions may require regulatory oversight and be considered a “mobile medical application” [142].

The timing and frequency of messages are important components when planning the intervention and may be determined by the intervention scope as well as user preference. Earlier studies on SMS text messaging interventions have suggested a preference for weekly messaging [113]. However, different intervention types may require a more adaptive message delivery system, such as smoking cessation programs that often require an increased volume of messages close to the desired quit date [143] or high-risk behavior prevention programs targeting binge drinking or inappropriate sexual behaviors timing their messages to when the risky behavior is expected to occur, for example, on a Friday night [116,141]. Therefore, strategies for message delivery and frequency could be adapted to suit the CA intervention.

Just-in-time adaptive interventions (JITAIs) leverage smartphone sensor data to “provide the right type (or amount) of support at the right time” [22]. Smartphone sensor data would determine and even predict “states of vulnerability” (susceptibility to negative health outcomes) [144] and “states of receptivity” (the capacity to receive, process, and use the intervention) [120] in the user when the intervention may be required and more useful. This novel approach may be particularly useful for behavior change interventions supporting a healthy lifestyle, such as increasing physical activity or adhering to a healthier diet, or supporting substance use remission [22,120]. Nevertheless, researchers considering this approach should take into account human and economic resources as JITAI design may require a larger development team that includes computer scientists and app developers.

Strategies to keep the users engaged for the intended duration of the intervention are particularly important in health care settings. These aspects were discussed in 11% (6/55) of the studies in our reviews [3,30,31,73,78,79]. Reported strategies included notifications, weekly summaries, reminders, motivational statements, persuasive techniques, a high frequency of messages to promote habit formation, and daily encouragement. In addition, CA-specific engagement strategies included building rapport and attachment with the user [72,73] or adding gamified components to incentivize CA use for rewards and points [73,79].

Digital health interventions often report high rates of participant attrition, which may limit the validity of research findings and, more importantly, the effectiveness of the intervention. Therefore, the assessment of the CA-led intervention should be complemented by regular evaluations of end-user adherence to as well as engagement with and acceptability of the intervention. Several assessment methods are commonly used, including quantitative, data-driven analyses and qualitative assessments of users’ opinions.

The definition of adherence to digital health interventions refers to the extent to which a user has interacted with the intervention [159]. This term may be used to define the degree to which a user interacts with the CA (greater adherence equals more time engaging with the intervention) or the degree to which the user-CA interaction complies with the prescribed recommendation (intended use of the intervention) [159]. In health care interventions, the concept of “intended use” is preferred, and it should be clearly defined during the CA design and development stage for the subsequent adherence measurements to be meaningful. Increased adherence to an intervention may be related to its increased effectiveness [75,160], although the data are not conclusive [3,161,162].

User engagement with the CA may be evaluated using data metrics such as the times the user opened the app, time spent interacting with the CA, the extent of the dialog, or the number of screens opened if the CA also includes other functions [70]. Chaix et al [30] measured use duration, interest in various educational contents, and level of interactivity as indicators of engagement. Nevertheless, researchers should consider the challenges of defining engagement with digital interventions, which may include other user-related variables such as the severity or stage of the disease as well as the long-term engagement with the CA [163].

Other aspects of CA use, such as underused or missing topics or CA functionalities not working as intended, may also be assessed. CA use analytics are often embedded in host platforms. Commercial platforms such as ManyChat [164] may offer a variety of built-in analytics tools such as the number of times the CA is accessed. Some of these platforms offer free-of-charge services. For health care CAs, the open-source MobileCoach platform [35] offers flexible, customizable use analytics.

Acceptability refers to the “affective attitudes towards a new digital health intervention” [165]. It is a dynamic concept comprising the intention to engage with the novel CA, the actual interaction with the CA, and the postengagement satisfaction [165].

Acceptability is a subjective term that is generally assessed using questionnaires or other qualitative methods such as focus groups or interviews. For example, Kowatsch et al [73] evaluated the acceptance of a CA to support asthma self-management using a 7-point Likert scale (strongly agree-strongly disagree) for perceived usefulness, ease of use, enjoyment, and use intention, and Echeazarra et al [68] used a survey with questions on ease of use, preference for the CA over existing methods, CA usefulness for its intended purpose, and whether the user had stopped using it as measures of acceptability and satisfaction. Furthermore, Gabrielli et al [69] facilitated a participatory design workshop where suggestions for improvement were provided via open-ended questions, and Ly et al [3] conducted semistructured interviews on the benefits, opportunities, and challenges associated with the CA for mental health. Yan et al [166] described a very involved process of evaluation of an mHealth intervention to promote physical activity. A focus group discussion was organized whereby each SMS text message was displayed and participants were required to respond either with “Yes, I like it” or “No, let’s change it to make it better.” This voting was then followed by a discussion in which suboptimal messages were improved and the strengths of effective messages were noted. Finally, participants may also be questioned about their willingness to recommend the conversation to others, which is a good indicator of satisfaction and acceptability [70].

Several aspects of user engagement and acceptability may be measured using one of several app quality rating tools, of which the most commonly used one is the Mobile App Rating Scale [167]. The use of standardized, validated rating scales may improve the reproducibility of this research area and facilitate the reporting of trial results, although they are not specific for CAs.

Safeguarding the privacy and security of CA users’ data is essential and should be a part of the entire design and development cycle. Health information is considered personal, sensitive information that should be protected at all times. The level of data protection should align with the data collected by the CA, if any. Therefore, the functionalities of the CA will determine the type of sensitive data to be collected and guide the inclusion of data protection software such as firewalls and encryption.

In general, developers should minimize the amount of personal and sensitive information collected from users by asking specific questions to avoid oversharing or simply providing predetermined responses instead of using free text. Furthermore, all CAs should include a privacy policy that is brief and written in clear language outlining the data collected and the uses of these data. All data must be encrypted during transit (when the message is being sent) and at rest (when the message has been delivered) [192]. The platform on which the CA will be deployed may also vary according to the CA functionalities. For example, a CA collecting users’ personal data should not be deployed on proprietary or messaging platforms as the platform data management policies may not be clearly reported [65] or data sharing with third parties may occur without informing the user [193]. This might create an ever-increasing digital footprint, potentially allowing for user identification from data aggregation rather than actually identifiable information [194].

A 2020 framework for governing the responsible use of CAs in health care highlighted the importance of safeguarding data privacy, including user health data, history of interactions, and disclosure of user data even if unintended [195]. In addition, the framework highlighted the user’s right to access their personally identifiable information, the requirement of user consent before recording or saving health-related data, and the preclusion of using the stored data as a means of surveillance or to discriminate users against health care privileges or coverage [195].

Health care CAs that collect users’ sensitive data must comply with country-relevant data privacy laws, such as the Health Insurance Portability and Accountability Act in the United States [192] or the General Data Protection Regulation (GDPR) in the EU [196]. These laws’ jurisdiction is generally limited to the issuing country; however, the GDPR applies to any EU citizen within or outside the EU. The GDPR, which went into effect in 2018, is an overarching law that aims to enhance the rights of individuals over their personal data, defined as any data that may allow for the identification of a person on their own or combined with other data, including pseudonymized data [196]. Alternatively, the Health Insurance Portability and Accountability Act is industry-specific and applies only to health-related data [197]. Other countries have adopted their own data protection laws and regulations. In Singapore, the Personal Data Protection Act is a baseline regulatory framework informing the collection, distribution, and use of personal data [198].

In addition to the aforementioned GDPR, children’s data are generally more stringently safeguarded. For example, in the United States, the Children’s Online Privacy Protection Act [199] requires that verifiable parental consent be obtained by all digital operators (not restricted to health care) collecting data from children (aged <13 years). Similar considerations are included within the GDPR and the Singapore Personal Data Protection Act, with the caveat that, in some European countries, parental consent is required for children and adolescents aged <16 years.