We chose to use qualitative semistructured individual interviews to explore stakeholder perspectives. This project is part of a larger study in which we also explored stakeholder perspectives on priority areas for AI applications, perceived benefits of and concerns about AI, and educational needs; these results have been reported previously [23,24]. In this paper, we report stakeholder perspectives on decision-making on the development, financing, and adoption of AI applications in the care of older adults. This study focused on stakeholders, including older adults (defined as those aged 60 years or older) and care partners for older adults (via self-report); clinicians; health system or health insurance leaders involved in technology adoption, such as the chief information officer (hereafter referred to as “payers”); and investors and technology developers. We combined older adults and care partners in the same group as many older adults were or had been care partners.

This project was supported by the Johns Hopkins Artificial Intelligence and Technology Collaboratory (JHU-AITC) for Aging Research [25]. The Stakeholder Engagement Core within the JHU-AITC engages longitudinally with a stakeholder council of older adults, care partners, and clinicians from Maryland and Iowa. We recruited older adults, care partners, and clinicians via referrals from the study team and council members; we recruited payers, investors, and technology developers via referrals from the study team, JHU-AITC colleagues, and the coordinating center that supports artificial intelligence and technology collaboratories. We also used snowball sampling to recruit contacts of existing participants. Recruitment continued until information saturation was reached.

All participants needed to be English speaking and be able to provide informed consent. Each participant was offered a US $50 gift card.

The interview guide was developed by the study team, which included expertise from engineering, business development, patient engagement, geriatric medicine, and social and behavioral science. We defined AI as “technologies where machines try to do what human beings do” and described that there were many types of AI applications, such as alerts, sensors, prediction algorithms, or robots. For older adults, care partners, clinicians, and payers, we asked how they decided whether to use an AI product; for investors, we asked how they decided to invest in an AI product; for developers, we asked how they made decisions on what type of AI product to work on or move forward with in development. The full interview guide is available in Multimedia Appendix 1.

Three team members (NLS, KC, and JM) conducted interviews in person (at private workspaces) or virtually from May 2023 to January 2024. All 3 team members had prior qualitative research experience. One team member (NLS) had prior working relationships with some of the clinician participants.

Interviews were audio recorded and transcribed verbatim; the average duration of the interviews was 34 (SD 11) minutes. We collected demographic information from all participants. For older adults and care partners, we collected information about education, financial strain, health literacy, self-rated health, and duration of caregiving. For clinicians, we asked about their practice setting. For all other participants, we asked about their professional roles and the duration working in that role.

We analyzed interview transcripts using reflexive thematic analysis informed by the approach by Braun and Clarke [26] to identifying, analyzing, and reporting patterns of meaning across qualitative data. This method was appropriate because our study aimed to examine how different stakeholder groups described decision-making on the development, financing, and adoption of AI health technologies for older adults. Consistent with the guidance from Braun and Clarke [26] on good practice in thematic analysis, we used an iterative analytic process that emphasized methodological clarity; reflexive engagement with the data; and alignment among the research questions, analytic procedures, and resulting themes.

To code the transcripts, we used thematic analysis with an integrated approach that used both inductive coding and a deductive organizing framework that drew conceptually from the TAM [16,20]. Specifically, a codebook was developed based on the interview guide and revised iteratively to add new codes generated inductively based on the review of 9 randomly selected transcripts. We organized the codes, when applicable, by domains in the TAM and compared our findings to the TAM during the interpretation of the results. Each transcript was independently coded using the finalized codebook by at least 2 investigators (NLS, KC, and JM), and differences were reconciled through consensus. The investigators involved in the analysis had training in clinical general or geriatric medicine, health service research, and/or qualitative research. Results were summarized into major themes and illustrated via representative quotes. Analysis used the ATLAS.ti software (version 24; ATLAS.ti Scientific Software Development GmbH).

Reflexivity was addressed throughout data collection and analysis. The study team included members with expertise in engineering, business development, patient engagement, geriatric medicine, and social and behavioral science, which allowed for multiple disciplinary perspectives to inform interpretation. At the same time, our team recognized that these professional backgrounds, as well as prior relationships with some participants, could shape data collection and interpretation. To address this, the team discussed emerging interpretations, considered alternative explanations, and examined how researchers’ assumptions and disciplinary perspectives may have influenced theme development.

Several strategies were used to enhance trustworthiness. Credibility was supported through independent coding by multiple investigators and consensus discussions using transcripts. Dependability was strengthened by maintaining documentation of analytic decisions, including codebook revisions, coding discussions, and theme refinement. Confirmability was supported through team-based reflexive discussion and by grounding interpretations in the data rather than a priori assumptions. Transferability was supported by describing the study sample, stakeholder groups, recruitment approach, interview procedures, and analytic process in sufficient detail to allow readers to assess the relevance of the findings to other settings and populations.

This project was approved by the Johns Hopkins Medicine Institutional Review Board (IRB00356437). The reporting of this study followed the COREQ (Consolidated Criteria for Reporting Qualitative Research) reporting guidelines [27]. Consent to participate was obtained for this project, and it adheres to human ethics guidelines. This study used deidentified data and posed minimal risk to participants.

The 49 participants included 15 (30.6%) older adults or care partners, 15 (30.6%) clinicians, 8 (16.3%) payers, 5 (10.2%) investors, and 6 (12.2%) technology developers (Table 1). The older adult group had a mean age of 71.3 years; 60% (9/15) of the older adults reported caregiving experience. Clinicians comprised physicians, nurses, pharmacists, rehabilitation therapists, and a dentist. Most of the older adult or care partner participants (14/15, 93.3%) and all the clinician participants (15/15, 100%) were from Maryland and Iowa, whereas payer, investor, and technology developer participants were from 9 different states.

Three themes emerged from the interviews. First, stakeholders across groups identified cost, usability, and value as key considerations when making decisions about AI health technologies for older adults, although the meaning of these concepts differed by stakeholder group. Second, these differing interpretations contributed to tensions between end users (older adults, care partners, and clinicians) and those responsible for developing and financing technologies (investors and developers), particularly regarding affordability and the design process. Third, participants described several potential strategies to better align stakeholder priorities and support the adoption of AI technologies for older adults. We describe each of the 3 themes below in detail.

We observed tension between the decision-making priorities of the end users (eg, older adults, care partners, and clinicians) and those of the developers and investors. The decisional factors that were important to end users were at times in conflict with those that were important to investors and developers. Two areas of tension included the affordability of the technology and the design process (Figure 1).

As stated above, older adults, care partners, and clinicians valued low cost (ie, “costs should be kept as minimal as possible” [older adult 3]). On the other hand, investors and developers commented on multiple considerations in the development process that drive up cost due to the significant risk and the long time horizon involved in developing AI health technologies. Both investors and developers commented that direct-to-consumer products are much higher risk:

However, the alternative involves more regulatory cost and time:

A developer commented the following on a recent publication looking at novel medical devices:

The investment risks and the long wait, in turn, drive up cost:

A second area of tension was regarding the design process for novel technologies. All stakeholder groups agreed that AI health applications should be designed after first understanding the target users and the health problem. However, many commented that, often, they saw the reverse—that existing technologies were offered to older adults even when they were not the most suitable solutions. One older adult said the following about health technology:

One engineer said the following:

Another AI researcher commented on the incentive structure being a barrier to problem-driven design:

This is the first study, to our knowledge, to examine decision-making related to AI technology development, financing, and adoption across diverse stakeholder groups. While previous research and frameworks such as the TAM have broadly acknowledged that cost, usability, and perceived value are significant determinants of technology acceptance by end users [28], this study adds depth by demonstrating how these concepts hold varied and nuanced meanings among stakeholder groups. In general, older adults and clinicians shared similar perspectives as end users, whereas investors and technology developers shared similar perspectives as those responsible for bringing about innovation; payers tended to have distinct perspectives from those of the other groups as they focused on broader, system-level concerns. The findings on the tension areas between decision-making priorities across stakeholder groups are also novel, drawing attention to factors in the existing technology development process that contribute to competing priorities.

How the different stakeholders viewed cost, value, and usability is consistent with what has been reported regarding other health technologies, but these have usually been reported separately and in different fields of study. For example, older adults, care partners, and clinicians primarily interpreted cost in terms of affordability, which has also been shown in studies on patient behavior regarding adoption of health care technologies [17,29]. Payers framed cost and value in terms of return on investment through reduced costs associated with high-frequency or high-severity clinical outcomes, a viewpoint common in the health care economics literature [30,31]. Investors and technology developers interpreted cost and value through the economic lens of market viability and profitability [32]. Our study design offering a juxtaposition of perspectives is a strength. A comparison of stakeholder perspectives has been described in the adoption of other technologies, such as patient portal use, but has focused on implementation without including the perspectives of developers or investors [18,33].

Recognizing that the same words and labels (eg, “cost” or “usability”) represent different concepts to different stakeholders can help entrepreneurs and technology developers be more considerate and specific in the design process to meet stakeholder needs. More specific descriptions of these considerations may help highlight their importance. For example, instead of cost and usability, it may be more helpful for developers to recognize that older adults value products that are affordable and easy to use for individual users, whereas payers value technology that reduces health care costs and is easy to implement within an existing health care system. Similarly, funders of AI health technology research can evaluate proposals with metrics that are inclusive of different stakeholder viewpoints, for example, ensuring that a project addresses both individual usability and ease of health system–level implementation.

AI is unique in that the same underlying product can be used in multiple contexts. For example, large language models can be a direct-to-consumer tool (eg, ChatGPT) or a clinician-facing decision support system (eg, OpenEvidence). In contrast, prior health technologies such as patient portals have much more defined and stable target users. AI’s adaptability means that the target user can shift after a product has already been developed and deployed. This has important implications for stakeholder alignment: the cost, usability requirements, accountability structures, and regulatory pathways may differ substantially depending on the target user. This flexibility creates both opportunity, allowing products to find their highest-value use case, and risk as it can obscure whose needs the technology was actually designed to serve.

Considering multiple perspectives together revealed critical tensions between end users (older adults, care partners, and clinicians) and technology creators (developers and investors). These tensions, particularly regarding affordability and design processes, signal significant barriers to wider adoption of patient-centered, impactful AI technologies for older adults. AI operates within a uniquely heterogeneous regulatory landscape where AI health applications can enter the market through multiple channels [34]. Some can be marketed directly to consumers as wellness products, which are subject to minimal regulatory oversight, whereas others may be considered medical devices that need to meet rigorous requirements before approval by the Food and Drug Administration [35]. Investors and developers highlighted factors in both pathways that contributed to tensions regarding affordability and design of AI health applications, including that direct consumer-facing technologies posed substantial financial risks whereas regulated products faced significant regulatory hurdles and lengthy approval timelines. The related incentive misalignments favor incremental enhancements to existing technologies over exploratory problem-driven innovations addressing complex issues in older populations, which is a phenomenon that has also been reported in more traditional medical device development [35]. Indeed, the current health technology development infrastructure may very well inadvertently motivate high-cost products that are “solutions in search of a problem,” and our results highlight the need for change.

Potential changes may include a revised or new regulatory and approval pathway for AI health technologies that is better tailored to the specific risks involved. Another change may involve altering the incentive structures, at least in research environments with a mission of discovery, to motivate problem-driven design and exploration of new ideas over honing existing expertise. We also identified actionable suggestions to facilitate adoption and align decision-making priorities from participants. Closer engagement across stakeholder groups throughout the development process was suggested by participants as one key solution to problem-driven design, which has also been reported in prior literature [12]. However, for older adults and clinicians to be interested and willing to engage, there first needs to be better public education and awareness campaigns to demystify AI and make clear its potential benefits. Participants suggested specific approaches for educating older adults about AI, such as through health care providers as well as media and cultural norms. Participants also suggested including stakeholder engagement as an explicit criterion for investment or funding, which can enhance accountability. Another suggestion by participants identified financial models such as public-private partnerships to balance cost considerations across stakeholders. The JHU-AITC is an example of one such partnership where public funding is awarded to promising pilot AI projects in very early development to help them advance to a stage where they can successfully attract private investment. Finally, it may be important to learn from other countries and health care models in terms of how they approach stakeholder engagement in AI health application development.

We acknowledge that our study has limitations. First, our participants were only restricted to English speakers, which may limit the generalizability of our findings to non-English speakers. Second, most older adult, caregiver, and clinician participants were based in Maryland and Iowa in the United States, which may limit geographic transferability. However, the inclusion of Iowa residents allowed us to obtain perspectives from rural residents and clinicians, where technology adoption is often delayed or limited. Third, several stakeholder groups lacked gender and racial diversity. For instance, all investor and technology developer participants were male. We did not assess prior experiences and familiarity with AI among our participants. Our future work should build on these findings through larger surveys with more diverse and representative samples across stakeholder groups.

In sum, we found that, across diverse stakeholder groups, the concepts of cost, usability, and value were considered universally important but had different meanings across different groups. This finding can help inform future research and technology development by highlighting that different elements of cost, usability, and value need to be separately considered and evaluated to meet the needs of diverse stakeholders. We also found competing decision-making priorities among different stakeholder groups when considering AI health applications for older adults, especially between those responsible for making new AI technologies available and the end users of such technologies. Implementing participant suggestions and other strategies to better align these priorities is essential for ensuring the widespread adoption of patient-centered, impactful AI solutions for older adults.