WorkFit: Designing Proactive Voice Assistance for the Health and Well-Being of Knowledge Workers

3.1.1 Health and Well-Being Apps. Previous studies have shown that knowledge workers express a desire for VA to assist them in managing health and well-being issues [2]. We hypothesized that knowledge workers would be more likely to assess an application designed to support their health and well-being specifically in the context of their office work environment. Drawing insights from existing literature, we identified three prevalent health concerns among knowledge workers at their workplaces: sedentary behavior [33, 40], hydration [2, 27, 39], and stress [9, 47].

Coping exercise with a defined goal. Existing research highlights the success of goal-based interventions in motivating users to engage in activities [6]. Our objective was twofold: not only to detect and deliver health and well-being interventions but also to suggest coping exercises to address these issues. We chose to suggest coping activity as a health recommendation for each intervention. These coping activities need to be actionable and feasible within the work environment, enabling easy execution for knowledge workers. Furthermore, each suggested activity should include a time-based or performance-oriented goal.

Stress. Controlled slow breathing exercises, like the ‘4-7-8 technique’, offer benefits for relaxation, heart rate regulation, and blood pressure management [20, 34]. This technique, also known as

pranayama, is an ancient yogic practice that triggers the parasympathetic state, promoting rest and relaxation. The exercise involves three steps: (1) Inhale for 4 seconds, (2) hold breath for 7 seconds, and (3) exhale for 8 seconds. To ensure sufficient impact, we recommended performing the ‘4-7-8’ breathing exercise for a duration of 2 minutes.

Sedentary behavior. Walking offers substantial long-term health benefits, especially after prolonged periods of sitting [40]. As a simple and effortless activity, walking is particularly suitable for knowledge workers in office settings. To accommodate the office environment's constraints, we suggested a goal of minimum of 20 steps per sedentary break.

Inconsistent Hydration. To tackle inconsistent hydration, we requested knowledge workers to drink a cup of water during each intervention. A daily intake of 15.5 cups of water is recommended [11, 15], with each cup being about 250 ml. To achieve this goal during the active sixteen hours of the day (excluding eight hours of sleep), a rate of one cup per hour is advisable, hence our reminder for one cup every 45-minutes.

3.1.2 Conversation Design. When designing the conversational strategy of WorkFit we decided to adhere to the following three principles: (1) User-centered dialogue: Our aim was to consider a user-centered approach and allow knowledge workers to flexibly act upon the interventions. When occupied with a task, it should be possible for the user to decline the given health recommendation and to postpone the recommendations. (2) Reflect: Moreover, our aim was to make users reflect on their unhealthy behavior when they declined a recommendation. During the reflective conversation, WorkFit should highlight the importance of healthy behavior. (3) Motivate: While performing the health recommendation, WorkFit should update the knowledge worker about their progress and give a motivational message regarding their goal.

3.1.3 Smartwatch. We chose to use smartwatches due to their health monitoring and interaction capabilities, driven by factors such as: (1) Availability of physiological sensors: Smartwatches offer real-time physiological measurements through built-in sensors. (2) Proximity: The close proximity of a smartwatch to its user enhances its efficacy in capturing attention and increasing user engagement. (3) Multi-modality: Smartwatches enable multi-modal interactions, including graphical, spoken, and tactile output, allowing users to engage through diverse modalities. For the development and deployment of WorkFit, we selected Samsung Galaxy Watch4¹

3.2.1 Sedentary Behavior Intervention. To automatically track sedentary behavior of the knowledge worker, WorkFit relies on accelerometer sensor readings using class ‘TYPE_STEP_DETECTOR’³. WorkFit monitors activity and inactivity of the knowledge worker based on the readings provided by the step detector. The ‘Last step taken’ timer starts when WorkFit detects user inactivity. This timer increments until a footstep is detected by the step detector (as shown in 1b ). After 60 minutes, WorkFit initiates a voice-based intervention, as described in Table 1. When the user confirms the break, WorkFit prompts the user to take 20 steps (2a ). If a footstep is detected before achieving 60-minute sedentary counter, the timer resets automatically.

Following our conversation design principles, in the sedentary behavior intervention conversation (Table 1), we offered knowledge workers a choice to accept, delay, or decline the health recommendation. If the knowledge worker would like to postpone the break, then WorkFit asks for waiting time (in minutes) before it delivers the intervention again. If the knowledge worker declines the interventions, WorkFit gives information that allows the worker to reflect on their sedentary behavior. It does so by highlighting the importance of taking breaks. If the knowledge worker accepts the intervention and starts performing the health recommendation by walking, then after completing 10 steps, WorkFit VA delivers a motivational message and requests to keep going. After completing the 20 steps, it will update the user about the completion of the exercise.

3.2.2 Hydration Intervention. WorkFit triggers a hydration voice intervention every 45 minutes. It begins with a voice reminder (as shown in Table 2), followed by a request message to confirm after two minutes. If the knowledge worker cannot drink water immediately, perhaps due to an empty bottle or due to being away from their desk. In this case, WorkFit gives them an option to postpone the intervention for a specific time. If the knowledge worker declines to postpone the intervention, it will deliver a reflective message highlighting the importance of drinking water. On the other hand, if the user accepts the intervention and drinks a cup of water, WorkFit acknowledges the user and reports the total water intake for the day (see 2b ).

3.2.3 Stress Intervention. To monitor the knowledge worker's heart rate, we use the heart rate sensor available in the smartwatch. We fetched heart rate reading using the class ‘TYPE_HEART_RATE’⁴. We monitored the heart rate reading every second and displayed it on the interface (see 1b .) A standard resting heart rate for adults typically ranges between 60 and 100 bpm [19, 25]. WorkFit activates the ‘high heart rate’ warning when the heart rate exceeds 105 bpm. Correspondingly, the ‘low heart rate’ warning appears when the heart rate falls below 55 bpm. When any of these warnings are triggered, users are prompted to perform the ‘4-7-8 breathing exercise’ to regulate their heart rate back to a healthy resting range.

Following our conversation design principles, when users confirm their intent to perform the breathing exercise, WorkFit initiates the guidance process (Table 3). It is depicted by an animation consisting of a contracting and expanding bubbles, replicating the inhale and exhale of breathing (Figure 3). Throughout the voice guidance, graphical cues are provided as follows: during the 4-second inhale phase, the label ‘Inhale’ (3a ) is displayed. During the 7-second breath-hold phase, the label ‘Hold’ (3b ) is displayed. While exhaling, the label ‘Exhale’ (3c ) is shown. The complete exercise comprises two cycles of these stages. The whole breathing exercise lasts 120 seconds. WorkFit initiates the conversation concerning the breathing exercise upon the occurrence of high or low heart rate warnings, as illustrated in Table 3. If the knowledge worker declines performing the breathing exercise, it provide the option of postponing the breathing exercise.

Breathing guidance showing, inhale, hold, and exhale.

5.2.1 Pre-Evaluation Questionnaire. The pre-study questionnaire was administered using an online survey to gain insights into participants backgrounds. In total, the pre-study questionnaire consisted of 15 questions focusing on age, profession, gender, work habits, voice interaction familiarity, health priority during work, break frequency, water consumption, experience with fitness tracking tools.

5.2.2 Daily Questionnaire. Participants evaluated WorkFit over five consecutive workdays and completed a daily questionnaire at the end of each day. We created the questionnaire aiming to capture their overall daily experience, encompassing user inputs, impressions, and suggestions. The daily questionnaire facilitated the identification of their preferences, dislikes, inopportune instances, and improvement suggestions. Additionally, participants provided subjective assessments of break interventions, water intake prompts, and breathing exercises. They rated WorkFit's utility, helpfulness, and satisfaction on a scale of 1 to 10.

5.2.3 Post-Evaluation Questionnaire and Interview. After the longitudinal study, participants were asked to complete a post-evaluation survey to capture their overall user experience. The survey inquired about app usage positive and negative experiences with WorkFit as a workplace conversational interface, and subjective evaluations of heart rate warnings, water intake, and break intervention accuracy based on their physiological conditions. Additionally, participants were encouraged to provide suggestions for enhancing the next version of WorkFit. The semi-structured post-evaluation interview centered on participant's insights gathered from their responses to the daily questionnaire. On an average, the interview lasted for 45 minutes. In the interview, the participants were probed with respect to the overall user experience, daily usage, preferred features, factors for voice interaction, behavioral adaptations during interaction and contextual awareness during events

Post-study interview audio recordings were transcribed and coded. The coding process involved generating codes to capture the data's essence and represent its concepts and ideas. Inductive thematic analysis was used to identify emerging themes. Patterns, similarities, and connections between codes were then identified to delineate distinct themes. Each theme was defined and labeled descriptively. To ensure rigor, the second and third authors conducted independent cross-checks on the identified themes [17].

6.1.1 Challenges. Lack of Agency Over Proactive Voice Interventions.

A few participants were annoyed by the lack of agency when the voice interactions were triggered. They desired agency over voice interventions depending on the situation, availability, and people around them. “If your colleagues are not aware of WorkFit, then it is embarrassing. It is like you didn't put your phone on silent mode. This is considered unprofessional.’’ [P05] Participants desired that WorkFit should privately notify them before delivering a voice intervention. They desired a new form of interaction which could give them control and make them aware about an upcoming intervention.

In the study, some of the participants expressed a negative sentiment towards the proactive voice interventions. The instances of these interventions occurring during opportune moments, particularly when there were no scheduled events in the participant's calendar, sometimes resulted in uncomfortable situations. This was attributed to the lack of control over proactive voice interventions to adhere to professional mannerisms in the office setting.

Identifying Inopportune Moments.

Our investigation revealed that relying solely on calendar events falls short in effectively managing inopportune moments. While knowledge workers are typically organized and adhere to calendar events, we identified a range of contextual factors (see section 6.2) that they value for interaction considerations within WorkFit. Although minimizing inconvenient instances is beneficial, the complexity lies in addressing multiple contextual factors for each interaction. Additionally, determining these moments using various contextual factors necessitates continuous monitoring of knowledge worker activities, leading to potential privacy concerns.

Moreover, as we observed calendar events to determine inopportune moments, it might appear technically suitable to intervene the user. However, various factors, including social context, relationships with involved individuals (such as close friends, colleagues, or acquaintances), and the nature of the intervention, make it challenging to accurately predict the user's readiness and comfort for intervention.

Managing Privacy. Although, participants found voice interventions to be effective in gaining attention, they also criticised them with respect to privacy in the office. Some participants were unhappy, because WorkFit delivered indiscreet health warnings in a social setting in their office. The spoken health warning led to a privacy violation. The lack of encoding while delivering a health warning was a common complaint from the participants. “It was a very obvious wording. Everyone around you knows you are stressed. It should not give personal information so obviously. It could rephrase and ask me: How about an oxygen break?” [P09] Further, participant P02 criticised: “If you are with colleagues, you don't want them to know that your heart rate is running high.” [P2]Non-Intrusive for Colleagues.

The primary medium of VA is audio, which easily captures the attention of people present in close proximity. The loud nature of audio often creates problems while interacting in a social setting. Many participants complained that voice interventions in office settings distracted their colleagues. Participants reported that some instances of voice interaction were attention-grabbing for their colleagues. For this reason many participants reduced the volume of the smartwatch. Participant P2 reported that, while working from home volume was at 80%, and while working from the office volume was at 20%. The primary reason for decreasing the volume was that the interventions should only be audible to themselves and not to their colleagues. “To detect inopportune moments it can synchronize with my calendar, but not with my colleague's calendar.” [P14]

6.1.2 Benefits. Effective in Capturing User Attention. Participants noted that voice interventions captured their attention effectively, particularly when they lost track of time. Compared to other notification forms like GUIs and vibration, which participants were familiar with, voice interventions were more proficient at highlighting unhealthy behaviors. This auditory approach not only brought their awareness to these behaviors but also motivated them to follow the recommended activities. As one participant remarked, “The voice interventions stand out in comparison to other notifications. Other notifications are just plain sound and no sentences. Audio sentences cannot be ignored and make you to perform the action.” [P09]Awareness, Self Check, and Reflection on Unhealthy Behavior.

WorkFit conversational interaction enhanced user awareness and encouraged reflection on their work behavior. Statements like “You haven't moved for a while, how about taking a break?” increased participant's awareness of sedentary patterns, motivating them to contemplate their behavior. If users opted to cancel a break, conversational statements reinforced the significance of hydration and breaks. This conversational approach effectively convinced users to adhere to the recommended activities. “Voice interventions not only reminded me, but it also made me reflect and perform a self-check to drink water or take a break. It acted very well as a self-check mechanism. I almost got used to it, today I had to remind myself on my own.” [P10]Dual-Tasking and Eyes-Free Interaction.

Voice notifications should clearly convey the notification's intent, which is crucial for health and well-being notifications that require attention. Unlike GUI notifications, where users need to actively interpret the notification's intent, WorkFit's voice notifications provide the intent and action without demanding continuous monitoring by the user. This enables the knowledge workers to concentrate on their tasks and determine the feasibility of the suggested activity. WorkFit ensures essential attention, minimizing cognitive load for having to check the display. “Unlike other healthcare applications, WorkFit didn't expect me to check the app continuously.” [P03]

6.2.1 Emotional and Cognitive Availability. Serendipitous meetings are spontaneous and could happen at any location in the office setting such as in the office corridor, in the kitchen, at the cubicle. Participants encountered serendipitous meeting scenarios, such as colleagues or their boss reaching them at their desk, coincidental encounters in the corridor, or students or colleagues reaching them. During an on-going engagement with their colleagues, it is difficult for knowledge workers to interact with WorkFit. Therefore, they are cognitively unavailable at those moments. Further, sometimes during those situations knowledge workers could also be going through emotions such as stress, or feeling nervous or tense. “I was in an informal meeting with my colleague at my desk, which was not scheduled in the calendar. I was discussing about problems I have been encountering in my project, so I was bit stressed. Suddenly it said: You seem to be tensed, would you like to perform a breathing exercise? It broke my flow, I was disrupted in my work conversation.” [P01]

6.2.2 Social Context. Participants mentioned that they desired WorkFit to recognize and adapt to social scenarios in their office. Although social moments are informal, participants preferred to refrain from engaging with voice interventions in such moments. “Meetings, lunch break, when I am not at my work station. It should not serve me any voice interventions.” [P09] During the social event, delivering a voice intervention could lead to unintended attention from colleagues. In addition to this, delivering information by voice may lead to privacy issues.

6.2.3 Shared Spaces. Participants stated that they would abstain from having a voice interaction at some locations in their office. These locations are voice sensitive. A voice intervention, even at a low volume, is bound to catch attention of other persons in the vicinity. Participants stated that some locations should be marked as “no voice intervention zones” and mentioned office toilets, laboratories, kitchens, elevators, and office corridors as examples. “When I am at the office kitchen or in the pantry and when I am with my colleagues or at any office common place, WorkFit should not serve me any voice interventions.” [P13]

6.2.4 Embracing Time and Work Flexibility. The COVID-19 pandemic has forced knowledge workers to work from home but enabled them to manage their time according to their convenience. Work from home has offered greater time flexibility and convenience to knowledge workers. Due to flexibility of work time, work from home has enabled knowledge workers to balance between personal and professional time. This convenient work pattern has allowed knowledge workers to opt for relaxing breaks during work. Informal events such as power nap, family time, gym, and playful activities are now considered normal in a work day. “While working from home, I tend to sleep for sometime after lunch. WorkFit should consider my personal time.” [P10].

6.3.1 Hydration Intervention. Overall, WorkFit generated a total of 275 hydration voice interventions. Among these interventions, participants accepted 253 (92%), declined 10 (3.6%), and 12 (4.4%) interventions remained in a state of neither acceptance nor decline (see 5b ). The mean rating for perceived effectiveness of the hydration intervention was 7.9 (as shown in 5a ).

Participants indicated that the hydration interventions encouraged them to hydrate more frequently with small sips of water, in contrast to their prior habit of consuming larger quantities at once or maintaining irregular water consumption patterns. “Earlier I used to drink 400 ml in one go, after a long interval. But WorkFit helped me to divide my water consumption into smaller chunks.” [P13]. Also, participant P04 revealed that hydration interventions were immensely beneficial, because the participant was suffering from the Gilbert syndrome⁵, in which an increase in the level of bilirubin leads to dehydration in the body. “I have a genetic problem of Gilbert syndrome, due to which my liver discharges lots of bil [bilirubin]. So, drinking water frequently is essential. If I don't drink water, I get water cramps [muscle cramps caused by dehydration]. So for me it was really very beneficial” [P04]. Nevertheless, certain participants desired WorkFit to update accurate water consumption rather than assuming a standard consumption of 1 cup of water with every intervention.

6.3.2 Stress Intervention and Breathing Exercise. Participants expressed their appreciation for the stress interventions as it increased their awareness of stressful moments when they experienced stress while working. Also they commended on the ‘4-7-8’ breathing technique as a coping exercise to stabilize their heart rate. Through the WorkFit real-time heart tracking (1b ), the participants noticed that their heart rate returned to a more desirable level while performing the breathing exercise. In total, 56 HHR (high heart rate) warnings were received by 11 participants. In the daily survey questionnaire, participants were asked to rate if they judged that the breathing exercise helped them in ameliorating their heart rate. The question received mean rating of 7.1 with a standard deviation of 2.1.

To determine the impact of the breathing exercise, we recorded the participant's heart rate for 120 seconds, during the breathing exercise. 6a shows the percentile median graph with a 95% confidence interval of heart rate readings across all the participants for each second. The graphs exhibit a constant decline in the heart rate while performing the breathing exercise. To determine if the breathing exercise helped participants to reduce their heart rate significantly, we performed a paired t-test on the percentile value, comparing the initial 5 s (1-5) and the heart rate at 116-120 s. We found that the heart rate was significantly reduced at the end of the breathing exercise (t(52)= 6.1, p < 0.01). We investigated further to determine the minimum duration that the breathing exercise required to significantly reduce the heart rate. We performed paired t-tests comparing the initial 5 s to subsequent five-second blocks (6-10 s, 11-15 s,... 116-120 s). The results show a significant difference at 26-30 s (t(56) = 5.4, p < 0.01) and all further blocks. Hence, the ‘4-7-8’ breathing exercise had a desirable effect on the heart rate after performing it for only 30 s.

WorkFit issued 20 Low Heart Rate (LHR) warnings. Similar to the high heart rate median percentile graph, we calculated a percentile median graph with 95% confidence interval for the low heart rate (6b ). The graph depicts a consistent increase in heart rate while performing a breathing exercise. The paired t-test shows a significant difference at the end of breathing exercise (t(17) = 2.7, p < 0.05). In fact, a significant difference was already found at 6-10 s compared to the initial heart rate (t(19) = 3.0, p < 0.01). Thus after performing the ‘4-7-8’ breathing exercise for only 10 s, the participants were able to improve their heart rate.

6.3.3 Sedentary Intervention. In total WorkFit recommended 43 breaks to all participants, of which 79.1% (34) breaks were accepted, 14.0% (6) were declined, and 6.9% (3) were neither declined nor accepted (as illustrated in 5b ). During the 34 accepted breaks, the participants performed 2438 steps in total, with a mean of 71 steps per break.

Participants valued WorkFit's recommendation of 20 steps during a break, finding it achievable and effective in countering sedentary behavior, whether at home or in the office. “In the office setting, you are not expected to have long walks so I just have to do the bare minimum” [P10]. Moreover, participants appreciated the sensor-based step counting feature, recognizing and valuing WorkFit's ability to not only propose a step count goal but also ensure participant's successful achievement of this goal. “Many times I was lazy to get up, but the step counter was counting steps and displaying ‘0/20’. Because it was sensor-based, it could not be gamed [cheated]. So, I had to get up and walk those steps. The step counter encouraged me to get up and perform those steps.” [P03]. In general, the participants noticed the influence of WorkFit interventions on their sedentary behavior. WorkFit's interventions targeting sedentary behavior motivated them to consistently take breaks and promoted physical activity.

Some participants expressed the view that a 60-minute interval between breaks was excessively long and suggested the option to personalize the time between breaks. Users indicated a preference for customizing the break reminder timer to intervals such as 45 or 50 minutes, aligning better with their individual preferences and behavior.

7.1.1 Redesigning VA: Meeting the Needs of the Workplace. Although voice interaction was helpful in capturing attention and reflection with the proactive nature of WorkFit's voice interventions, the proactive interventions occasionally led to discomfort, as it could not be tailored to professional etiquette norms in the office environment. Participants expressed a desire for agency over voice interactions, particularly in consideration of their surroundings and the presence of colleagues. Many participants viewed the lack of control as a drawback, emphasizing the need for intervention adaptability, based on the situation and the people present. Additionally, voice interventions like heart rate warnings raised privacy concerns among participants. Although these interventions were urgent and required instant attention, some users desired new forms of proactive voice interaction, in order to give them agency and also capture their attention instantly.

We propose adaptive VA, i.e., which refers to a voice interface design that dynamically adjusts its interaction and behavior based on user preferences, such as the user's location, social settings, and cognitive availability. The advantages of an adaptive user interface have been established in previous studies [3, 37], highlighting enhancements in user experience and interface usability. In the case of an adaptive VA, it should dynamically switch between modalities (GUI, vibration and voice) based on factors such as notification type, intent, contextual elements, criticality and urgency. The adaptive VA, has the capability to consider multiple factors and subsequently deliver interventions in a format suitable for the current context. Unlike traditional approaches, the adaptive VA, eliminates the need to explicitly track whether the moment is opportune or inopportune. Instead, it tailors the intervention style based on relevant factors. Investigating the suitability of an adaptive VA, in an office setting holds significant value.

7.1.2 Inopportune Moments in Work Settings. Although it has been reported that participants would like to have health interventions proactively [48], our study found that daily health interventions should also consider opportune and inopportune moments for proactive interaction. Lastly, while working from home, the participants desired WorkFit to be able to recognize personal moments sleep, outdoor walks, and time with their kids. Like Reicherts et al. [31], we identified that participants accepted getting interrupted while at work compared to interrupting an ongoing conversation with their colleague.

In the study, we tried to tackle the problem of interruption during an inopportune moment by synchronizing the user's calendar with WorkFit interventions. Although this approach was successful, there were many inopportune moments, which user desired to be recognized by WorkFit. Considering only events placed in the calendar as inopportune moments and events not placed in the calendar as opportune moments were inadequate for successful user engagement with WorkFit. For some participants events depended on their colleagues availability, which was unpredictable for them. Also, in office setting of knowledge workers it was acceptable to walk into another colleague's cubicle/office if something urgent needed to the discussed. Also, participants desired to avoid having a voice interaction in a social setting which could place them at the center of attention. Furthermore, participants also wanted to consider location while serving a voice intervention. For example, office pantry, common shared places and toilets. Participants also wanted WorkFit to consider the surrounding and the environment they are in.

7.2.1 Empowering User Agency: Dual-Interventions. It is widely advised in various human-computer interaction guidelines to empower users with agency or control during interactions [23, 24, 38]. However, proactive VA has no provision to enable users with agency during the interaction. In proactive interventions, users lack control over voice interactions, which can result in occasional embarrassment and annoyance, ultimately diminishing the overall user experience and making it challenging to maintain sustained user attention and engagement. Even after tracking the user's schedule, it is difficult to predict an opportune moment for the interaction. Tracking their schedule may suggest it is an opportune moment to intervene, there are multiple factors such as the presence of colleagues, location, and cognitive availability, all of which play a crucial role in determining if the moment is opportune or inopportune. In a setting like an office, which consists of formal and social interaction, it is essential to give users control to decide if the particular moment is an opportune or inopportune moment for interaction. Allowing users to decide if the moment is appropriate for interaction will lower the instances of inconvenience. We propose “dual-mode intervention” design, the VA may choose to intervene non-intrusively (ring, auditory icon, vibration, or GUI) as the primary mode, reserving voice as a secondary intervention. For example, the intervention could be presented through a GUI to convey the notification's intent, complemented by a vibration to capture the user's attention. If the user does not respond to the intervention, meaning they neither accept, reject, nor postpone it, the system subsequently may deliver a voice intervention.

7.2.2 Delivering Encoded Voice Interventions for Privacy. Privacy is an important aspect of an interaction. Not disclosing private information during the interaction is essential. The fundamental characteristic of voice interaction involves conveying information through audio. In VA, the inherent nature of information delivery lacks discretion, potentially drawing the attention of those near the user. Consequently, it is imperative for VA to refrain from delivering any private information during voice interventions, as this could lead to a violation of user privacy. In case of WorkFit, users did not flag a privacy concern for sedentary behavior and hydration intervention. However, they raised concerns about unambiguous stress interventions through voice. For instance, stress interventions should be discreet, potentially initiated with a vibration or GUI intervention and further, voice intervention could be in encoded form (e.g., “take some deep breaths”). Also, they were comfortable receiving all notifications when at home in comparison with an office. They highlighted that not all voice interventions can violate privacy, but it depends on the type, nature, and setting in which the information is delivered. Therefore, it is essential to adopt measures to encode a voice interventions considering the setting and nature of intervention.

7.2.3 Prioritization of Interventions: Considering Criticality and Urgency. Previous research has found that users do not mind delivering health and well-being interventions directly without considering opportune moments for delivery [48]. In our experiment, users desired to prioritize interventions. They felt some interventions were urgent while some were not. Some interventions require immediate action while others can be acted upon later. Interventions such as hydration and sedentary reminders possibly can be postponed for some time and still be effective, however, stress interventions may require instant action. Some participants stated that they would like to scale the interventions based on an urgency and criticality parameter. It is essential to consider whether interventions are urgent or not. If the interventions are urgent, then direct voice interventions are appropriate.

7.2.4 Mindful of Setting's Etiquette. When VA is deployed in an office setting, it is essential to consider the rules of the office and interact accordingly. As WorkFit is ubiquitous and not stationary, it is compulsory to follow the etiquette of the place. When delivering interventions it is imperative to consider the fact that the office setting may have informal unscheduled meetings and smalltalk at the corridor. Audio should be kept at a moderate volume and not be disturbing to colleagues during the interaction. Also, users like to refrain from interacting with VA when they are in shared places.

7.2.5 Considering social, location, and user availability factors. Before delivering an intervention, it is important to identify if a particular intervention should be delivered in this setting. Before delivering interventions is essential to consider social, location and availability factors. Taking into account these factors will help to minimize the negative experiences of VA in office settings. We found that participants were comfortable having interaction in the home in comparison to office. The availability of the user is primary factor, as knowledge workers are often working from home. Due to work flexibility, they are switching between leisure activities and work, such as relaxing, and child-care. Hence, it is essential to consider user availability of while delivering intervention.