Research, normalized for decisions.

Direct evidence is promising, while most studies still emphasize usability and acceptability rather than health outcomes.

Small and heterogeneous literature; only a minority of studies directly measured loneliness.

Use a pilot to test both engagement and validated social-connection outcomes.

Conversational agents show potential mental-health benefits; subgroup findings suggest older adults may benefit, but age-specific certainty is limited.

Interventions, populations, and levels of human support vary substantially.

Treat human support as a design variable and predefine escalation pathways.

Loneliness is a material health concern, but technology should sit within a broader social-connection strategy.

Not an effectiveness trial of AI companions.

Do not position an AI tool as a substitute for community or human contact.

Artificial agents may provide perceived support, but the paper primarily offers a theoretical design framework.

Conceptual evidence cannot establish real-world effectiveness.

Use the work to shape interaction design, not to justify outcome claims.

The intervention landscape is broad, but evidence clusters and gaps need to be separated before product decisions are made.

Maps evidence coverage; it does not itself estimate a single pooled effect.

Use the map to locate comparable interventions and missing outcome evidence.

The field is expanding toward remote interactive support, but study quality and direct outcome evidence remain uneven.

Emerging field with varied agent types and evaluation designs.

Specify the product class and intended mechanism before comparing outcomes.