โŒ

Normal view

There are new articles available, click to refresh the page.
Before yesterdayMain stream

El FEM alerta de que tener arquitecturas de datos obsoletas frena el impacto de la IA en sanidad

16 January 2026 at 04:22

Aunque la IA tiene el potencial de transformar la atenciรณn mรฉdica en todo el mundo, el progreso se estรก topando actualmente con un muro invisible. Los obstรกculos son los sistemas de datos obsoletos. A esta conclusiรณn llega el Foro Econรณmico Mundial (FEM) en el informe publicado en vรญsperas de su reuniรณn anual en Davos, llamado โ€˜La IA puede transformar la asistencia sanitaria si transformamos nuestra arquitectura de datosโ€™.

Segรบn el estudio, dรฉcadas de registros aislados, formatos incompatibles e infraestructuras rรญgidas frenan el progreso. Para que la IA no siga siendo solo una herramienta para tareas especรญficas, sino que se convierta en un sistema autรณnomo y capaz de aprender, el FEM considera que el sector sanitario debe replantearse desde cero su arquitectura de datos.

Urge salir de la trampa del silo

Hasta ahora, las estructuras se basaban a menudo en entradas manuales y actualizaciones diferidas. Sin embargo, el futuro pertenecerรก a un canal de datos inteligente y unificado que limpie la informaciรณn de los sensores y las fuentes automatizadas en tiempo real y la haga directamente legible para la IA. En lugar de almacenarse en rรญgidas bases de datos relacionales, la informaciรณn se almacena cada vez mรกs en bases de datos grรกficas multidimensionales que permiten comprender inmediatamente el contexto y el significado.

El FEM considera que otro gran problema es la investigaciรณn mรฉdica. En este รกmbito, muchos conocimientos valiosos permanecen ocultos en notas o imรกgenes complejas, ya que son difรญciles de encontrar con una bรบsqueda convencional. Aquรญ es donde entra en juego la denominada vectorizaciรณn: los datos multimodales, desde textos hasta secuencias genรณmicas y seรฑales clรญnicas, se convierten en incrustaciones numรฉricas. Esto permite a la IA reconocer relaciones profundas, como comparar sรญntomas con casos anteriores o recuperar resultados de investigaciรณn relevantes con la mรกxima precisiรณn.

Seguridad y confianza

En definitiva, segรบn el FEM, un sistema sanitario moderno necesita un data lakehouse. Es decir, un lugar de almacenamiento centralizado en el que los datos de los laboratorios, los wearables y las aplicaciones de los pacientes confluyan de forma segura y estรฉn disponibles para su anรกlisis. Para que la protecciรณn de datos no se quede en el camino, una fรกbrica de datos inteligente debe garantizar que solo los usuarios autorizados tengan acceso y que la informaciรณn sea coherente.

Para garantizar que las recomendaciones de IA para los mรฉdicos sean comprensibles y fiables, รฉstas deben basarse en conocimientos clรญnicos validados. Los denominados grรกficos del conocimiento podrรญan servir como guรญas para garantizar que los resultados de la IA se ajusten a las directrices mรฉdicas.

Esta transformaciรณn de la IA es mรกs que una simple renovaciรณn tecnolรณgica. Segรบn la valoraciรณn del FEM, para las naciones soberanas, la creaciรณn de una arquitectura de datos preparada para la IA significa considerar la sanidad como un recurso nacional. Y, desde el punto de vista del foro, esta transformaciรณn radical es indispensable. Solo asรญ los paรญses podrรกn garantizar una atenciรณn mejor y personalizada y aprovechar al mรกximo el potencial de una IA con capacidad de autoaprendizaje.

โ€˜์ฑ—GPT ๊ฑด๊ฐ•โ€™ ์„ ๋ณด์ธ ์˜คํ”ˆAI, ์Šคํƒ€ํŠธ์—… ํ† ์น˜ํ—ฌ์Šค ์ธ์ˆ˜ ๋ฐœํ‘œ

14 January 2026 at 02:30

์˜คํ”ˆAI๊ฐ€ ์ƒŒํ”„๋ž€์‹œ์Šค์ฝ” ๊ธฐ๋ฐ˜ ์Šคํƒ€ํŠธ์—… ํ† ์น˜ํ—ฌ์Šค(Torch Health)๋ฅผ ์ธ์ˆ˜ํ–ˆ๋‹ค. ์—…๊ณ„ ๋ถ„์„๊ฐ€๋“ค์€ ์ด๋ฒˆ ์ธ์ˆ˜๊ฐ€ ์ง€๋‚œ์ฃผ ์ถœ์‹œ๋œ ์˜คํ”ˆAI์˜ โ€˜์ฑ—GPT ๊ฑด๊ฐ•(ChatGPT Health)โ€™ ์ด๋‹ˆ์…”ํ‹ฐ๋ธŒ๋ฅผ ๊ฐ•ํ™”ํ•˜๊ธฐ ์œ„ํ•œ ์ „๋žต์  ํ–‰๋ณด๋ผ๊ณ  ํ‰๊ฐ€ํ•˜๊ณ  ์žˆ๋‹ค.

์˜คํ”ˆAI๋Š” ์ง€๋‚œ์ฃผ ๋ธ”๋กœ๊ทธ ๊ฒŒ์‹œ๋ฌผ์„ ํ†ตํ•ด ์ฑ—GPT ๊ฑด๊ฐ•์˜ ๋น„์ „์„ ๊ตฌ์ฒด์ ์œผ๋กœ ์„ค๋ช…ํ–ˆ๋‹ค. ์‚ฌ์šฉ์ž๊ฐ€ ์ž์‹ ์˜ ์˜๋ฃŒ ๊ธฐ๋ก๊ณผ ์• ํ”Œ ํ—ฌ์Šค, ํŽ‘์…˜, ๋งˆ์ดํ”ผํŠธ๋‹ˆ์ŠคํŒ”๊ณผ ๊ฐ™์€ ๊ฑด๊ฐ• ์•ฑ์„ ์—ฐ๊ฒฐํ•ด ์ตœ๊ทผ ๊ฒ€์‚ฌ ๊ฒฐ๊ณผ๋ฅผ ์ดํ•ดํ•˜๊ณ , ๋ณ‘์› ์ง„๋ฃŒ๋ฅผ ์ค€๋น„ํ•˜๊ฑฐ๋‚˜ ๊ฐœ์ธ์˜ ์˜๋ฃŒ ์ด์šฉ ํŒจํ„ด์„ ๋ฐ”ํƒ•์œผ๋กœ ์‹๋‹จ ๊ด€๋ฆฌ ๋ฐฉ๋ฒ•๊ณผ ๋ณดํ—˜ ์˜ต์…˜ ๊ฐ„์˜ ์„ ํƒ์ง€๋ฅผ ํŒŒ์•…ํ•  ์ˆ˜ ์žˆ๋„๋ก ๋•๋Š” ์ฑ—๋ด‡์„ ๊ตฌํ˜„ํ•˜๊ฒ ๋‹ค๋Š” ๊ตฌ์ƒ์ด๋‹ค.

ํ† ์น˜ํ—ฌ์Šค ์—ญ์‹œ ๋น„์Šทํ•œ ๋ฌธ์ œ ํ•ด๊ฒฐ์— ์ฃผ๋ ฅํ•ด ์˜จ ๊ธฐ์—…์ด๋‹ค. ์ฐฝ์—…์ž๋“ค์— ๋”ฐ๋ฅด๋ฉด ํ† ์น˜ํ—ฌ์Šค๋Š” ์ง„๋ฃŒ ๊ธฐ๋ก, ๊ฒ€์‚ฌ ๋ณด๊ณ ์„œ, ์›จ์–ด๋Ÿฌ๋ธ” ๋ฐ์ดํ„ฐ, ์†Œ๋น„์ž ์›ฐ๋‹ˆ์Šค ๊ฒ€์‚ฌ ๊ฒฐ๊ณผ, ๋ถ„์‚ฐ๋œ ์˜๋ฃŒ ํฌํ„ธ์˜ ์ •๋ณด๋ฅผ AI๋กœ ํ†ตํ•ฉํ•ด ๊ฐœ์ธ์ด ์˜๋ฃŒ ๊ธฐ๋ก์„ ๋ณด๋‹ค ์‰ฝ๊ฒŒ ์ดํ•ดํ•  ์ˆ˜ ์žˆ๋„๋ก ๋•๋Š” ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์„ ๊ฐœ๋ฐœํ•ด ์™”๋‹ค.

ํ† ์น˜ํ—ฌ์Šค๋Š” โ€œAI๋ฅผ ์œ„ํ•œ ์˜๋ฃŒ ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ๊ตฌ์ถ•ํ•˜๊ธฐ ์œ„ํ•ด ์‹œ์ž‘ํ–ˆ๋‹ค. ํฉ์–ด์ ธ ์žˆ๋Š” ์˜๋ฃŒ ๊ธฐ๋ก์„ ํ•˜๋‚˜์˜ ๋งฅ๋ฝ ์—”์ง„์œผ๋กœ ํ†ตํ•ฉํ•ด ์ „์ฒด ๊ทธ๋ฆผ์„ ๋ณผ ์ˆ˜ ์žˆ๋„๋ก ํ•˜๊ณ , ์ •๋ณด ๊ฐ„์˜ ์—ฐ๊ฒฐ๊ณ ๋ฆฌ๋ฅผ ์ฐพ์•„ ์ค‘์š”ํ•œ ๋‚ด์šฉ์ด ๋‹ค์‹œ๋Š” ์†Œ์Œ ์†์— ๋ฌปํžˆ์ง€ ์•Š๋„๋ก ๋•๋Š” ๊ฒƒ์ด ๋ชฉํ‘œ์˜€๋‹คโ€๋ผ๊ณ  ์„ค๋ช…ํ–ˆ๋‹ค.

HFS๋ฆฌ์„œ์น˜์˜ ๋ถ€์‹ค๋ฌด์ฑ…์ž„์ž ์•…์ƒคํŠธ ํ‹ฐ์•„๊ธฐ๋Š” ํ† ์น˜ํ—ฌ์Šค์˜ ๊ธฐ์ˆ ์ด ์ฑ—GPT ๊ฑด๊ฐ• ๊ฐœ๋ฐœ์„ ๊ฐ€์†ํ™”ํ•  ๊ฒƒ์ด๋ผ๊ณ  ๋ถ„์„ํ•˜๋ฉด์„œ, ์˜คํ”ˆAI๊ฐ€ ์˜๋ฃŒ ๋ถ„์•ผ์— ํŠนํ™”๋œ ์„œ๋น„์Šค๋ฅผ ์ •๊ตํ™”ํ•˜๋Š” ๋ฐ ๊ธฐ์—ฌํ•  ๊ฒƒ์ด๋ผ๊ณ  ํ‰๊ฐ€ํ–ˆ๋‹ค. ์˜๋ฃŒ ๋ถ„์•ผ๋Š” ์œ„ํ—˜ ๋ฏผ๊ฐ๋„๊ฐ€ ๋†’๊ณ  ๊ทœ์ œ๊ฐ€ ๊ฐ•ํ•œ ๋งŒํผ, ํ‹ฐ์•„๊ธฐ๋Š” ๋ฐ์ดํ„ฐ ํŒŒํŽธํ™” ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๋Š” ๊ฒƒ์ด ๊ฐ€์žฅ ์‹œ๊ธ‰ํ•œ ๊ณผ์ œ๋ผ๊ณ  ์ง€์ ํ–ˆ๋‹ค. ๋ฐ์ดํ„ฐ ํ†ตํ•ฉ์ด ๊ฐ€๋Šฅํ•ด์งˆ ๊ฒฝ์šฐ, ๋ณด๋‹ค ์‹ ๋ขฐ๋„ ๋†’๊ณ  ์ •ํ™•ํ•œ ์‘๋‹ต์„ ์ œ๊ณตํ•˜๋Š” ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์˜ ๊ธฐ๋ฐ˜์ด ๋งˆ๋ จ๋  ์ˆ˜ ์žˆ๋‹ค๊ณ  ๊ทธ๋Š” ์„ค๋ช…ํ–ˆ๋‹ค.

์•คํŠธ๋กœํ”ฝ๋„ ์˜๋ฃŒ AI๋กœ ๋ฒ”์œ„ ๋„“ํ˜€

์˜คํ”ˆAI๋งŒ ์˜๋ฃŒ ์‹œ์žฅ์— ๋›ฐ์–ด๋“  ๊ฒƒ์€ ์•„๋‹ˆ๋‹ค. ์˜คํ”ˆAI์˜ ๊ฒฝ์Ÿ์‚ฌ ์•คํŠธ๋กœํ”ฝ ์—ญ์‹œ AI ์†Œํ”„ํŠธ์›จ์–ดยท์„œ๋น„์Šค ๊ธฐ์—… ๊ฐ€์šด๋ฐ ์˜๋ฃŒ ๋ถ„์•ผ ์ง„์ถœ์— ์†๋„๋ฅผ ๋‚ด๊ณ  ์žˆ๋‹ค.

์•คํŠธ๋กœํ”ฝ์€ ์ตœ๊ทผ ๊ณต๊ฐœํ•œ ๋ธ”๋กœ๊ทธ์—์„œ โ€˜ํด๋กœ๋“œ ํฌ ๋ผ์ดํ”„ ์‚ฌ์ด์–ธ์Šคโ€™ ์ œํ’ˆ๊ตฐ์˜ ๊ธฐ๋Šฅ์„ ๊ฐ•ํ™”ํ•˜๊ณ  โ€˜ํด๋กœ๋“œ ํฌ ํ—ฌ์Šค์ผ€์–ดโ€™๋ฅผ ์ถ”๊ฐ€ํ•œ๋‹ค๊ณ  ๋ฐํ˜”๋‹ค.

์•คํŠธ๋กœํ”ฝ์— ๋”ฐ๋ฅด๋ฉด ํด๋กœ๋“œ ํฌ ํ—ฌ์Šค์ผ€์–ด๋Š” ์˜๋ฃŒ ์ œ๊ณต์ž, ๋ณดํ—˜์‚ฌ, ์†Œ๋น„์ž๊ฐ€ HIPAA(์˜๋ฃŒ์ •๋ณด ๋ณดํ˜ธ๋ฒ•) ์š”๊ฑด์„ ์ถฉ์กฑํ•˜๋Š” ํ™˜๊ฒฝ์—์„œ ์˜๋ฃŒ ๋ฐ์ดํ„ฐ๋ฅผ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ๋„๋ก ์ง€์›ํ•˜๋Š” ๋„๊ตฌ์™€ ๋ฆฌ์†Œ์Šค๋กœ ๊ตฌ์„ฑ๋œ๋‹ค.

์˜คํ”ˆAI ์—ญ์‹œ โ€˜์˜คํ”ˆAI ํฌ ํ—ฌ์Šค์ผ€์–ดโ€™๋ผ๋Š” ์ด๋ฆ„์˜ ์œ ์‚ฌํ•œ ์ œํ’ˆ์„ ์ถœ์‹œํ–ˆ๋‹ค. ์ด ์ œํ’ˆ ์—ญ์‹œ ์˜๋ฃŒ ๋ถ„์•ผ ๋‚ด ๋‹ค์–‘ํ•œ ์ดํ•ด๊ด€๊ณ„์ž๋ฅผ ๊ฒจ๋ƒฅํ•œ ๋„๊ตฌ๋กœ ๊ตฌ์„ฑ๋ผ ์žˆ๋‹ค.

์—…๊ณ„ ๋ถ„์„๊ฐ€๋“ค์€ ์˜คํ”ˆAI์™€ ์•คํŠธ๋กœํ”ฝ์˜ ์ž‡๋”ฐ๋ฅธ ์˜๋ฃŒ ๋ถ„์•ผ ์ง„์ถœ์„ ๋‘๊ณ , LLM ๊ธฐ์ˆ ์ด ์˜๋ฃŒ ๋ถ„์•ผ์—์„œ ์‚ฌ์šฉ๋  ๋งŒํผ ์ถฉ๋ถ„ํžˆ ์„ฑ์ˆ™ ๋‹จ๊ณ„์— ์ด๋ฅด๋ €๋‹ค๋Š” ์‹ ํ˜ธ๋กœ ํ•ด์„ํ•˜๊ณ  ์žˆ๋‹ค. ๋™์‹œ์— ์ด๋Š” ์˜๋ฃŒ ์‹œ์Šคํ…œ ์ „๋ฐ˜์—์„œ ๊ธฐ๋ฐ˜ ์ง€๋Šฅ ๋ ˆ์ด์–ด ์ž๋ฆฌ๋ฅผ ์ฐจ์ง€ํ•˜๊ธฐ ์œ„ํ•œ ๊ฒฝ์Ÿ์ด ๋ณธ๊ฒฉํ™”๋˜๊ณ  ์žˆ์Œ์„ ๋ณด์—ฌ์ค€๋‹ค๋Š” ํ‰๊ฐ€๋‹ค.

ํ‹ฐ์•„๊ธฐ๋Š” โ€œ์ด์ „ ๋ชจ๋ธ์€ ์‹ ๋ขฐ์„ฑ์ด ์ถฉ๋ถ„ํ•˜์ง€ ์•Š์•˜๋‹ค. ํ•˜์ง€๋งŒ ํ˜„์žฌ์˜ ๋ชจ๋ธ์€ ์ž„์ƒ ๊ธฐ๋ก์„ ์š”์•ฝํ•˜๊ณ , ํ™˜์ž์™€์˜ ์ปค๋ฎค๋‹ˆ์ผ€์ด์…˜์„ ์ง€์›ํ•˜๋ฉฐ, ๋ฌธ์„œ ์ž‘์—…์„ ๋ณด์กฐํ•˜๊ณ , ๊ด€๋ จ ์˜๋ฃŒ ๋ฌธํ—Œ์„ ์ฐพ์•„ ์ œ์‹œํ•จ์œผ๋กœ์จ ์˜๋ฃŒ์ง„์˜ ์ˆ˜์ž‘์—… ๋ถ€๋‹ด์„ ์‹ค์ œ๋กœ ์ค„์ผ ์ˆ˜ ์žˆ๋Š” ์ˆ˜์ค€์— ์ด๋ฅด๋ €๋‹คโ€๋ผ๊ณ  ์ง„๋‹จํ–ˆ๋‹ค.

ํ‹ฐ์•„๊ธฐ๋Š” ๋น„์šฉ ์ƒ์Šน๊ณผ ์˜๋ฃŒ์ง„ ๋ฒˆ์•„์›ƒ์ด ์‹ฌํ™”๋˜๋Š” ๊ฐ€์šด๋ฐ, ์˜๋ฃŒ ์‹œ์Šคํ…œ ์ „๋ฐ˜์ด ์ง„๋ฃŒ์˜ ์งˆ์„ ํ›ผ์†ํ•˜์ง€ ์•Š์œผ๋ฉด์„œ ํšจ์œจ์„ฑ์„ ๋†’์ผ ์ˆ˜ ์žˆ๋Š” ํ•ด๋ฒ•์„ ๋ชจ์ƒ‰ํ•˜๊ณ  ์žˆ๋‹ค๊ณ  ๋ถ„์„ํ–ˆ๋‹ค. ์ด๋Ÿฐ ์ƒํ™ฉ์—์„œ ์ž„์ƒ์ด๋‚˜ ํ–‰์ • ์›Œํฌํ”Œ๋กœ์šฐ์— ์กฐ๊ธฐ์— ์ž๋ฆฌ ์žก๋Š” ๊ธฐ์—…์ด ์žฅ๊ธฐ์ ์ธ ์ „๋žต์  ์œ„์น˜๋ฅผ ํ™•๋ณดํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ  ๊ทธ๋Š” ์ง€์ ํ–ˆ๋‹ค. ํ•œ ๋ฒˆ AI ์‹œ์Šคํ…œ์ด ํ•ด๋‹น ํ”„๋กœ์„ธ์Šค ์•ˆ์—์„œ ์‹ ๋ขฐ๋ฅผ ์–ป์œผ๋ฉด, ์ด๋ฅผ ๋‹ค๋ฅธ ์‹œ์Šคํ…œ์œผ๋กœ ๋Œ€์ฒดํ•˜๊ธฐ ๋งค์šฐ ์–ด๋ ค์›Œ์ง„๋‹ค๋Š” ์„ค๋ช…์ด๋‹ค.

์˜คํ”ˆAI์™€ ์•คํŠธ๋กœํ”ฝ์€ ๋ชจ๋‘ ์˜๋ฃŒ์ง„์„ ๋Œ€์ฒดํ•˜๊ฑฐ๋‚˜ ์น˜๋ฃŒ๋ฅผ ์ œ๊ณตํ•˜๋ ค๋Š” ๋ชฉ์ ์€ ์—†์œผ๋ฉฐ, ์˜๋ฃŒ์ง„์„ ๋ณด์กฐํ•˜๋Š” ์—ญํ• ์— ์ง‘์ค‘ํ•˜๊ณ  ์žˆ๋‹ค๊ณ  ๋ฐํ˜”๋‹ค. ์˜คํ”ˆAI๋Š” ์‚ฌ์šฉ์ž ๊ฑด๊ฐ• ๋ฐ์ดํ„ฐ๋ฅผ ๋ชจ๋ธ ํ•™์Šต์— ํ™œ์šฉํ•˜์ง€ ์•Š๊ฒ ๋‹ค๋Š” ๋ฐฉ์นจ๋„ ํ•จ๊ป˜ ์ œ์‹œํ–ˆ๋‹ค.
dl-ciokorea@foundryco.com

The AI readiness gap: Why healthcare and insurance struggle to scale beyond pilots

13 January 2026 at 06:15

When I first began leading AI programs across healthcare and insurance, I kept seeing the same pattern repeat itself. The early stages looked encouraging. Models performed well in controlled environments, small teams reported strong accuracy and executives saw dashboards that suggested meaningful impact. But as soon as we attempted to move those same models into full operational workflows, the results changed. Accuracy dipped, exceptions grew and the expected improvements in cycle time or member experience did not appear.

That moment, the shift from pilot success to production friction, revealed something deeper about AI in regulated industries. These organizations are not struggling with innovation. They are struggling with readiness. The pilot looks promising because it operates inside a narrow, curated world. Scaling requires an ecosystem that is aligned, governed and capable of absorbing new forms of intelligence. Most enterprises are not built for that yet and that gap between possibility and readiness is becoming more visible as AI moves from experimentation to real operations.

Letโ€™s continue the narrative I built in my earlier CIO pieces on CRM, AI and the healthcare experience. Weโ€™ve focused on how AI and CRM help organizations move beyond transactional processes and toward more proactive models of engagement. The next step in that journey is understanding why so many AI initiatives stall on the path from promise to performance and what CIOs can do to close that gap.

Why pilots create a false sense of confidence

AI pilots succeed because they avoid real-world conditions. They operate on clean datasets, constrained workflows and a level of manual support that no enterprise can sustain. In one healthcare program I led, a risk prediction model delivered strong accuracy during testing. Once we connected it to multiple clinical, claims and eligibility systems, the model behaved differently. The issue was not the algorithm. It was the environment around it.

Pilots provide clarity because they filter out everything that makes healthcare and insurance difficult. Production systems reintroduce the complexity that pilots deliberately remove. Data becomes inconsistent. Workflows expand. Roles multiply. Compliance teams ask new questions. What appeared efficient in a contained environment suddenly feels fragile and incomplete.

This pattern is not just something I have seen in individual programs. External analyses show the same thing. McKinsey, for example, has documented how many payers remain stuck in pilot mode because their data, processes and operating models are not ready for AI at scale.

I began seeing the same dynamic in other regulated sectors as well. In a manufacturing program I supported, an equipment failure prediction model performed well in engineering pilots but struggled once connected to maintenance workflows, supplier data and plant-floor operations. In banking, a fraud-risk model delivered strong early accuracy but failed to scale because the surrounding compliance reviews and case management systems were not designed to absorb algorithmic decisions. These industries differed in context, but the readiness gap appeared for the same reason: the supporting environment could not sustain the weight of enterprise AI.

Where AI breaks when organizations try to scale

Across healthcare and insurance, the breakdown tends to happen in the same places. The first is data fragmentation. Clinical information lives in electronic records. Claims data lives in adjudication systems. Member interactions live inside CRM platforms. Pharmacy data, care management notes, eligibility information and provider relationships each have their own systems. A model trained on one dataset cannot handle the reality of workflows that cross 10 or more environments.

The second breakdown happens at the workflow layer. Pilots isolate a decision. Production requires that decision to move through people, systems and documentation requirements. A predicted risk score means nothing if it cannot be routed to a nurse, documented for compliance, recorded in CRM and tracked for audit purposes. Many organizations reach this point and realize they lack the operational foundation to support AI-driven decisions at scale.

The third breakdown is contextual. Humans interpret data through policy, history, clinical appropriateness, operational nuance and lived experience. AI does not have that instinct unless it is trained, governed and monitored in a way that reflects actual decision-making. In pilots, analysts bridge the gap manually. In production, the absence of context becomes a source of friction.

The final breakdown involves compliance. Healthcare and insurance operate under strict oversight. AI-driven decisions must be explainable, traceable and ethically defensible. A system that cannot demonstrate why it decided or how it treated different populations will not pass regulatory review. This does not slow innovation. It slows ungoverned innovation, which is exactly the concern behind emerging frameworks such as the EU Artificial Intelligence Act and the U.S. Algorithmic Accountability Act of 2023.

The cultural readiness gap

Technology gaps can be addressed with time and investment. Cultural gaps take longer. Many organizations still treat AI as a project inside data science or analytics teams. They celebrate proofs-of-concept but do not build the operational or governance environment required to support continuous learning and deployment.

In one health plan I worked with, a model predicting medication nonadherence delivered accurate insights, but adoption was low. Care coordinators did not understand how the model generated recommendations, so they distrusted its output. When we introduced transparent explanations, training sessions and role-based views, adoption increased dramatically.

This experience reinforced an important reality: People do not adopt what they cannot trust. And trust is not created through accuracy metrics. It is created through clarity, collaboration and visibility. These breakdowns reveal a clear pattern in which AI has matured faster than the operational and governance structures required to support it. This is why the CIOโ€™s role is shifting from system integration to readiness orchestration.

The CIOโ€™s role in closing the readiness gap

CIOs are uniquely positioned to bridge the gap between technical possibility and operational reality. They sit at the intersections of data, governance, compliance, workflow design and enterprise leadership. AI cannot scale until these elements come together in a structured and predictable way.

The first area CIOs must focus on is data readiness. Healthcare and insurance do not need a single consolidated dataset. They need aligned definitions, lineage and quality standards that allow models to behave consistently across workflows. This requires collaboration between technology, clinical, claims and service teams. Without that alignment, AI produces insights that break as soon as they cross departmental boundaries.

The second area is operational readiness. AI must be integrated into the systems teams already use. A model has little value if it only produces a score. The real value appears when that score routes into a CRM console, triggers a task, enters a case management queue or initiates proactive outreach. This integration turns AI from an analytical tool into an operational capability.

The third area is governance. AI in regulated industries must be explainable, testable and monitored continuously. A responsible AI framework ensures that models meet fairness expectations, documentation requirements and audit standards. Governance should not be a checkpoint at the end of deployment. It should be embedded into the design. Much of this is now being framed as a digital trust challenge. Deloitte, for example, highlights how enterprises that invest in governance, transparency and accountability build an advantage in โ€œEarning digital trust.โ€

The fourth area is measurement. Pilots often focus on accuracy metrics. Enterprises care about impact. CIOs must redefine success through operational outcomes such as reduced cycle time, improved member satisfaction, lower rework and stronger compliance posture. This shift in measurement helps organizations focus on what matters most.

Finally, organizations must redesign processes around intelligence. AI changes how the work flows: Decisions move earlier in the process. Exceptions become clearer. Proactive outreach becomes possible. CIOs must help teams rethink workflows so AI becomes a structural part of operations rather than a tool sitting beside them.

The CIO is now the connective leader who brings data, compliance, clinical insight, claims operations and customer experience together under a single readiness model. That responsibility goes far beyond technical implementation. It involves shaping behaviors, redesigning workflows, establishing shared definitions and ensuring that every algorithm introduced into the enterprise is explainable, traceable and actionable. Without this cross-functional alignment, even the best models will fail to scale.

Moving from experimentation to enterprise value

Healthcare and insurance organizations are facing a moment where the limitations of pilot-driven innovation are becoming clear. They do not lack ideas or algorithms. They lack readiness. And readiness is not about technology. It is about leadership, design and alignment.

The organizations that scale AI successfully do not treat it as a project. They treat it as a capability that requires shared ownership. They invest in data alignment, operational integration, governance visibility and behavioral readiness. They understand that AI becomes powerful only when it becomes part of how the enterprise thinks, acts and learns.

As I reflect on the organizations that have successfully scaled AI, one lesson stands out. Transformation does not come from the model; it comes from the readiness of the enterprise around it. Technology alone has never changed healthcare or insurance. Alignment, trust and disciplined execution have. When CIOs focus on readiness as much as innovation, AI stops being an experiment and becomes a structural capability that improves outcomes, strengthens compliance and makes complex systems feel more human rather than more technical.

This article is published as part of the Foundry Expert Contributor Network.
Want to join?

AI ์˜๋ฃŒ ํ™œ์šฉ ์ž‡๋‹จ ํ™•๋Œ€โ€ฆ์˜คํ”ˆAI ์ด์–ด ์•คํŠธ๋กœํ”ฝ โ€˜ํด๋กœ๋“œ ํฌ ํ—ฌ์Šค์ผ€์–ดโ€™ ๊ณต๊ฐœ

13 January 2026 at 03:15

์ด๋ฒˆ์— ๊ณต๊ฐœ๋œ ํด๋กœ๋“œ ํฌ ํ—ฌ์Šค์ผ€์–ด๋Š” ์˜๋ฃŒ ์ œ๊ณต์ž์™€ ๋ณดํ—˜์‚ฌ, ํ™˜์ž๋“ค์ด HIPAA(์˜๋ฃŒ์ •๋ณด ๋ณดํ˜ธ๋ฒ•) ๊ธฐ์ค€์„ ์ถฉ์กฑํ•˜๋Š” ํ™˜๊ฒฝ์—์„œ ์˜๋ฃŒ ๋ฐ์ดํ„ฐ๋ฅผ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ๋„๋ก ์„ค๊ณ„๋๋‹ค. ํ•ต์‹ฌ ํŠน์ง•์€ ์˜๋ฃŒ ๋ฐ์ดํ„ฐ ์ ‘๊ทผ์„ฑ๊ณผ ํ–‰์ • ์—…๋ฌด ์ž๋™ํ™”๋ฅผ ๋™์‹œ์— ๊ฐ•ํ™”ํ–ˆ๋‹ค๋Š” ์ ์ด๋‹ค. ํŠนํžˆ ๋ฏธ๊ตญ ์˜๋ฃŒ ์‹œ์Šคํ…œ์—์„œ ์‚ฌ์šฉํ•˜๋Š” ๋ฏธ๊ตญ ๋ฉ”๋””์ผ€์–ดยท๋ฉ”๋””์ผ€์ด๋“œ ์„œ๋น„์Šค์„ผํ„ฐ(CMS) ๋ณด์žฅ ๋ฐ์ดํ„ฐ๋ฒ ์ด์Šค, ๊ตญ์ œ์งˆ๋ณ‘๋ถ„๋ฅ˜(ICD-10) ์ฝ”๋“œ, ๊ตญ๊ฐ€ ์˜๋ฃŒ์ธ ์‹๋ณ„ ๋“ฑ๋ก๋ถ€(NPI Registry) ๋“ฑ ์˜๋ฃŒ ๋ฐ์ดํ„ฐ๋ฒ ์ด์Šค์™€ ์—ฐ๋™๋œ๋‹ค.

์•คํŠธ๋กœํ”ฝ์€ ๊ณต์‹ ๋ธ”๋กœ๊ทธ๋ฅผ ํ†ตํ•ด โ€œํด๋กœ๋“œ ํฌ ํ—ฌ์Šค์ผ€์–ด๋กœ ์‚ฌ์ „์Šน์ธ ๊ฒ€ํ†  ํ”„๋กœ์„ธ์Šค๋ฅผ ๊ฐ€์†ํ™”ํ•˜๊ณ , ๋ณดํ—˜ ์ฒญ๊ตฌ ์ด์˜์‹ ์ฒญ ์ž๋ฃŒ๋ฅผ ์ž‘์„ฑํ•˜๋ฉฐ, ํ™˜์ž ๋ฉ”์‹œ์ง€๋ฅผ ๋ถ„๋ฅ˜ํ•ด ๊ธด๊ธ‰ํ•œ ์‚ฌ์•ˆ์„ ์‹๋ณ„ํ•˜๋Š” ๋“ฑ์˜ ์—…๋ฌด๋ฅผ ์ง€์›ํ•  ์ˆ˜ ์žˆ๋‹คโ€๋ผ๊ณ  ์„ค๋ช…ํ–ˆ๋‹ค.

์•คํŠธ๋กœํ”ฝ์€ ์ด๋ฒˆ ์„œ๋น„์Šค๋ฅผ ํ†ตํ•ด ๋ณด์žฅ ๊ธฐ์ค€๊ณผ ์ž„์ƒ ๊ฐ€์ด๋“œ๋ผ์ธ, ํ™˜์ž ๊ธฐ๋ก์ด ์—ฌ๋Ÿฌ ์‹œ์Šคํ…œ์— ๋ถ„์‚ฐ๋ผ ์ˆ˜์ž‘์—… ํ™•์ธ์ด ํ•„์š”ํ–ˆ๋˜ ๊ธฐ์กด ์‚ฌ์ „ ์Šน์ธ ์ ˆ์ฐจ๋ฅผ ๋ณด์™„ํ•  ์ˆ˜ ์žˆ์„ ๊ฒƒ์œผ๋กœ ๋ณด๊ณ  ์žˆ๋‹ค. ํด๋กœ๋“œ๋Š” ๊ด€๋ จ ์ •๋ณด๋ฅผ ์ข…ํ•ฉํ•ด ์Šน์ธ ํŒ๋‹จ์— ์ฐธ๊ณ ํ•  ์ˆ˜ ์žˆ๋Š” ์ดˆ์•ˆ๊ณผ ๊ทผ๊ฑฐ ์ž๋ฃŒ๋ฅผ ์ œ์‹œํ•˜๋„๋ก ์„ค๊ณ„๋๋‹ค๋Š” ์„ค๋ช…์ด๋‹ค. ์ด์™€ ํ•จ๊ป˜ ํ™˜์ž ๋ฉ”์‹œ์ง€ ๋ถ„๋ฅ˜, ์ง„๋ฃŒ ํ˜‘์—… ์ง€์›, ์˜๋ฃŒ ์Šคํƒ€ํŠธ์—…์„ ์œ„ํ•œ ์„œ๋น„์Šค ๊ฐœ๋ฐœ ๋“ฑ ๋‹ค์–‘ํ•œ ์˜์—ญ์—์„œ๋„ ํ™œ์šฉ๋  ์ˆ˜ ์žˆ์„ ๊ฒƒ์œผ๋กœ ๋‚ด๋‹ค๋ดค๋‹ค.

์ƒ๋ช…๊ณผํ•™ ๋ถ„์•ผ์—์„œ๋„ ํด๋กœ๋“œ์˜ ํ™œ์šฉ ๋ฒ”์œ„๊ฐ€ ํ™•๋Œ€๋˜๊ณ  ์žˆ๋‹ค. ์ž„์ƒ์‹œํ—˜ ์šด์˜๊ณผ ๊ทœ์ œ ๋Œ€์‘ ๋‹จ๊ณ„๊นŒ์ง€ ์ง€์› ๋ฒ”์œ„๊ฐ€ ๋„“์–ด์กŒ์œผ๋ฉฐ, ๋ฉ”๋””๋ฐ์ดํ„ฐ(Medidata), ํด๋ฆฌ๋‹ˆ์ปฌํŠธ๋ผ์ด์–ผ์Šค๋‹ท๊ฑฐ๋ธŒ(ClinicalTrials.gov), ๋ฐ”์ด์˜ค์•„์นด์ด๋ธŒยท๋ฉ”๋“œ์•„์นด์ด๋ธŒ(bioRxivยทmedRxiv), ์˜คํ”ˆํƒ€๊นƒ์ธ (Open Targets), ์ผ๋ธ”(ChEMBL) ๋“ฑ ์ฃผ์š” ์—ฐ๊ตฌยท์ž„์ƒ ๋ฐ์ดํ„ฐ ํ”Œ๋žซํผ๊ณผ์˜ ์—ฐ๋™์ด ์ถ”๊ฐ€๋๋‹ค. ์•คํŠธ๋กœํ”ฝ์€ ํ•ด๋‹น ๊ธฐ๋Šฅ์„ ํ†ตํ•ด ์—ฐ๊ตฌ์ž๋“ค์ด ์ž„์ƒ์‹œํ—˜ ํ”„๋กœํ† ์ฝœ ์ดˆ์•ˆ ์ž‘์„ฑ, ๋Œ€์ƒ์ž ๋ชจ์ง‘ ์ „๋žต ์ˆ˜๋ฆฝ, ๊ทœ์ œ ๋ฌธ์„œ ์ค€๋น„ ๋“ฑ์˜ ์—…๋ฌด๋ฅผ ํšจ์œจํ™”ํ•  ์ˆ˜ ์žˆ์„ ๊ฒƒ์œผ๋กœ ์„ค๋ช…ํ–ˆ๋‹ค.

ํ•œํŽธ ์ตœ๊ทผ AI ์—…๊ณ„ ์ „๋ฐ˜์€ ํ—ฌ์Šค์ผ€์–ด ๊ด€๋ จ ์ž‘์—…์— ์ง‘์ค‘ํ•˜๋Š” ํ๋ฆ„์„ ๋ณด์ด๊ณ  ์žˆ๋‹ค. ์˜คํ”ˆAI๋Š” ๊ฒ€์ง„ ๊ฒฐ๊ณผ ๋ถ„์„์ด๋‚˜ ์‹๋‹จ ๊ด€๋ฆฌ ๋“ฑ ๊ฑด๊ฐ• ๊ด€๋ฆฌ์— ๋„์›€์ด ๋˜๋Š” ์ •๋ณด๋ฅผ ์ œ๊ณตํ•˜๋Š” โ€˜์ฑ—GPT ๊ฑด๊ฐ•โ€˜์„ 8์ผ ๊ณต๊ฐœํ–ˆ๋‹ค. ๊ตฌ๊ธ€ ์—ญ์‹œ AI ์˜ค๋ฒ„๋ทฐ๋ฅผ ํ†ตํ•ด ๊ฑด๊ฐ• ์ •๋ณด๋ฅผ ์ œ๊ณตํ•ด ์™”์œผ๋‚˜, ์ผ๋ถ€ ๋ถ€์ •ํ™•ํ•œ ์˜๋ฃŒ ์ •๋ณด๊ฐ€ ํฌํ•จ๋œ ์‚ฌ์‹ค์„ ํ™•์ธํ•˜๊ณ  ๊ด€๋ จ ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ ๋…ธ์ถœ์„ ์ค„์ธ ๊ฒƒ์œผ๋กœ ์•Œ๋ ค์กŒ๋‹ค.
jihyun.lee@foundryco.com

2026๋…„ ์˜๋ฃŒ AI, ๋‹จ์ผ LLM ๋„˜์–ด ๋ฉ€ํ‹ฐ ์—์ด์ „ํŠธยท๋„๋ฉ”์ธ ํŠนํ™” ๋ชจ๋ธ๋กœ ๋‚˜์•„๊ฐˆ ์ด์œ 

12 January 2026 at 02:38

์ดˆ๊ธฐ ์ƒ์„ฑํ˜• AI ๋„์ž… ๋ฐฉ์‹์€ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ๋ชจ๋“  ๋ฐ์ดํ„ฐ๋ฅผ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์— ์ž…๋ ฅํ•œ ๋’ค โ€˜์ถ”๋ก โ€™์„ ์š”๊ตฌํ•˜๋Š” ํ˜•ํƒœ์˜€๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด ์ ‘๊ทผ๋ฒ•์€ ์ ์  ์ง€์†๋˜๊ธฐ ์–ด๋ ค์›Œ์ง€๊ณ  ์žˆ๋‹ค. ๋น„์šฉ์€ ๊ธ‰๊ฒฉํžˆ ๋ถˆ์–ด๋‚˜๊ณ  ์ •ํ™•๋„๋Š” ํ”๋“ค๋ฆฌ๋ฉฐ, ๊ทœ์ œ ์ค€์ˆ˜๋Š” ๊ด€๋ฆฌํ•˜๊ธฐ ์–ด๋ ค์šด ์ˆ˜์ค€์— ์ด๋ฅด๊ณ  ์žˆ๋‹ค. ๋ถ€๊ฐ€๊ฐ€์น˜๊ฐ€ ์œ ๋งํ–ˆ๋˜ ์ƒ์„ฑํ˜• AI๊ฐ€ ์—”ํ„ฐํ”„๋ผ์ด์ฆˆ ๊ทœ๋ชจ๋กœ ํ™•์‚ฐ๋˜๋Š” ๊ณผ์ •์—์„œ ํ˜„์‹ค์ ์ธ ์ œ์•ฝ์— ๋ถ€๋”ชํžˆ๊ณ  ์žˆ๋‹ค.

์ƒˆํ•ด๋ฅผ ๋งž์•„ ์ƒ์„ฑํ˜• AI์˜ ๋‹ค์Œ ๊ตญ๋ฉด์ด ์‹œ์ž‘๋˜๊ณ  ์žˆ๋‹ค. ๊ทธ ๋ณ€ํ™”๋Š” ๋” ํฐ ๋ชจ๋ธ๋กœ ์ •์˜๋˜์ง€ ์•Š๋Š”๋‹ค. ๊ณผ๊ฑฐ์™€ ๋‹ฌ๋ฆฌ 2026๋…„์€ ๋ชจ๋“ˆํ˜• ๊ตฌ์กฐ๋ฅผ ๊ฐ–์ถ”๊ณ , ๋„๋ฉ”์ธ์— ํŠนํ™”๋˜๋ฉฐ, ๊ฑฐ๋ฒ„๋„Œ์Šค๋ฅผ ๋‚ด์žฌํ™”ํ•œ AI ์‹œ์Šคํ…œ์œผ๋กœ ์ „ํ™˜๋˜๋Š” ์‹œ์ ์ด ๋  ์ „๋ง์ด๋‹ค. ์ด๋Ÿฐ ์•„ํ‚คํ…์ฒ˜๋Š” ์‹ค์งˆ์ ์ด๊ณ  ์ธก์ • ๊ฐ€๋Šฅํ•œ ๋น„์ฆˆ๋‹ˆ์Šค ๊ฐ€์น˜๋ฅผ ์ œ๊ณตํ•˜๋Š” ๋ฐ ์ดˆ์ ์„ ๋งž์ถ˜๋‹ค.

๋‹จ์ผ ๋ชจ๋ธ ์ค‘์‹ฌ AI ์ „๋žต์˜ ํ•œ๊ณ„

์ง€๋‚œ 2๋…„๊ฐ„ ๋งŽ์€ ๊ธฐ์—…์ด ์ƒ์„ฑํ˜• AI๋ฅผ ๋ฌด์ฐจ๋ณ„์ ์œผ๋กœ ๋„์ž…ํ•˜๋Š” ๋ฐฉ์‹์„ ํƒํ–ˆ๋‹ค. ์˜๋ฃŒ ๋ถ„์•ผ๋ฅผ ๋ณด๋ฉด, ๋ชจ๋“  ์ง„๋ฃŒ ๊ธฐ๋ก๊ณผ ๊ฒ€์‚ฌ ๊ฒฐ๊ณผ, ์ž„์ƒ ๋…ธํŠธ๋ฅผ ํ•˜๋‚˜์˜ LLM์— ์ž…๋ ฅํ•œ ๋’ค ์š”์•ฝ์ด๋‚˜ ์˜ˆ์ธก์„ ์š”๊ตฌํ•˜๋Š” ๋ฐฉ์‹์ด์—ˆ๋‹ค. ์ด๋Š” ํŒŒ์ผ๋Ÿฟ ํ”„๋กœ์ ํŠธ ๋‹จ๊ณ„์—์„œ ๋น ๋ฅธ ์„ฑ๊ณผ๋ฅผ ๋ƒˆ์ง€๋งŒ, ๊ณง ๋ฌธ๋งฅ ์ฒ˜๋ฆฌ ํ•œ๊ณ„์— ๋ถ€๋”ชํ˜”๊ณ  ์ถ”๋ก  ๋น„์šฉ์€ ๊ธ‰๊ฒฉํžˆ ์น˜์†Ÿ์•˜์œผ๋ฉฐ, ๊ฒฐ๊ณผ๋ฌผ์€ ์ž„์ƒ ํ˜„์žฅ์—์„œ ์š”๊ตฌ๋˜๋Š” ์ˆ˜์ค€์˜ ์ •ํ™•์„ฑ์„ ํ™•๋ณดํ•˜๊ธฐ ์–ด๋ ค์› ๋‹ค.

์Šคํƒ ํผ๋“œ๋Œ€ํ•™๊ต์˜ AI ์ธ๋ฑ์Šค ๋ณด๊ณ ์„œ์— ๋”ฐ๋ฅด๋ฉด, AI ๋ชจ๋ธ์€ ์ˆ˜ํ•™ ๋ฌธ์ œ๋‚˜ ํ‘œ์ค€ํ™”๋œ ์‹œํ—˜์ฒ˜๋Ÿผ ์ •๋‹ต์ด ๋ช…ํ™•ํ•œ ๊ณผ์ œ์—์„œ๋Š” ๊ฐ•์ ์„ ๋ณด์ด์ง€๋งŒ, ๋ณต์žกํ•œ ์ถ”๋ก ์ด ์š”๊ตฌ๋˜๋Š” ๋ฒค์น˜๋งˆํฌ์—์„œ๋Š” ์—ฌ์ „ํžˆ ์–ด๋ ค์›€์„ ๊ฒช๊ณ  ์žˆ๋‹ค. ๋…ผ๋ฆฌ์  ๊ณผ์ œ๋ฅผ ์•ˆ์ •์ ์œผ๋กœ ํ•ด๊ฒฐํ•˜์ง€ ๋ชปํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์•„, ์ •๋ฐ€๋„๊ฐ€ ํ•ต์‹ฌ์ธ ๊ณ ์œ„ํ—˜ ํ™˜๊ฒฝ์—์„œ๋Š” ํ™œ์šฉ ํšจ๊ณผ๊ฐ€ ์ œํ•œ์ ์ด๋‹ค.

์•ž์„  ์กฐ์ง์€ ์ด๋Ÿฌํ•œ ํ•œ๊ณ„๋ฅผ ๋ณด์™„ํ•˜๊ธฐ ์œ„ํ•ด ๋ชจ๋“ˆํ˜• ํŒŒ์ดํ”„๋ผ์ธ์œผ๋กœ ์ „ํ™˜ํ•˜๊ณ  ์žˆ๋‹ค. ์ด ํŒŒ์ดํ”„๋ผ์ธ์€ ์ •๋ณด ์ถ”์ถœ, ์ถ”๋ก , ๋Œ€ํ™”๋ฅผ ๊ฐ๊ฐ ์ตœ์ ํ™”๋œ ๋‹จ๊ณ„๋กœ ๋ถ„๋ฆฌํ•œ๋‹ค. ํ•˜๋‚˜์˜ ๋ชจ๋ธ์€ ์ž์œ  ํ˜•์‹์˜ ์ž„์ƒ ๋…ธํŠธ์—์„œ ํ•ต์‹ฌ ์ž„์ƒ ๊ฐœ์ฒด๋ฅผ ์ถ”์ถœํ•˜๊ณ , ๋‹ค๋ฅธ ๋ชจ๋ธ์€ ํ•ด๋‹น ๋ฐ์ดํ„ฐ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ตฌ์กฐํ™”๋œ ์ถ”๋ก ์„ ์ˆ˜ํ–‰ํ•˜๋ฉฐ, ๋˜ ๋‹ค๋ฅธ ๋ชจ๋ธ์€ ์ž์—ฐ์–ด ์ธํ„ฐํŽ˜์ด์Šค๋ฅผ ํ†ตํ•ด ๊ฒฐ๊ณผ๋ฅผ ์ „๋‹ฌํ•œ๋‹ค. ๊ฐ ๋ชจ๋“ˆ์€ ๋…๋ฆฝ์ ์œผ๋กœ ์กฐ์ •ํ•˜๊ณ  ๊ฐ์‚ฌํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ํ•„์š”์— ๋”ฐ๋ผ ๊ฐœ์„ ์ด ๊ฐ€๋Šฅํ•˜๋‹ค.

์ด์ฒ˜๋Ÿผ ์ž‘์—…์— ๋งž๋Š” ๋„๊ตฌ๋ฅผ ์ •ํ™•ํžˆ ๋ฐฐ์น˜ํ•˜๋Š” ์ ‘๊ทผ๋ฒ•์€ ์‹œ์Šคํ…œ์˜ ์†๋„์™€ ์•ˆ์ „์„ฑ์„ ๋†’์ด๊ณ , ํˆฌ๋ช…์„ฑ๋„ ํฌ๊ฒŒ ๊ฐ•ํ™”ํ•œ๋‹ค. ์ด๋Š” ์˜๋ฃŒ์ฒ˜๋Ÿผ ๊ทœ์ œ๊ฐ€ ์—„๊ฒฉํ•œ ์‚ฐ์—…์—์„œ AI ๊ฒฐ๊ณผ๋ฌผ์ด ์‹ค์ œ๋กœ ํ™œ์šฉ๋˜๊ธฐ ์œ„ํ•ด ๋ฐ˜๋“œ์‹œ ์ถฉ์กฑ๋ผ์•ผ ํ•  ํ•ต์‹ฌ ์š”๊ฑด์ด๋‹ค.

๋ฉ€ํ‹ฐ ์—์ด์ „ํŠธ ํ˜‘์—… ์‹œ์Šคํ…œ์˜ ๋ถ€์ƒ

๋‹ค์Œ ๋‹จ๊ณ„์˜ ์ฃผ์š” ๋ณ€ํ™”๋Š” ๋ฉ€ํ‹ฐ ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ์—์„œ ๋‚˜ํƒ€๋‚  ์ „๋ง์ด๋‹ค. ์ด๋Š” ์†Œ๊ทœ๋ชจ ์ „๋ฌธ AI ๋ชจ๋ธ์ด ์—ฌ๋Ÿฌ ์ž‘์—…์— ๊ฑธ์ณ ํ˜‘๋ ฅํ•˜๋Š” ๋„คํŠธ์›Œํฌ๋กœ, ๋””์ง€ํ„ธ ํŒ€์— ๋น„์œ ํ•  ์ˆ˜ ์žˆ๋‹ค. ์˜๋ฃŒ ๋ถ„์•ผ๋ฅผ ๊ธฐ์ค€์œผ๋กœ ์‚ดํŽด๋ณด๋ฉด, ํ•˜๋‚˜์˜ ์—์ด์ „ํŠธ๋Š” ๊ฒ€์‚ฌ ์ˆ˜์น˜์˜ ๋ณ€ํ™”๋ฅผ ๋ชจ๋‹ˆํ„ฐ๋งํ•˜๊ณ , ๋‹ค๋ฅธ ์—์ด์ „ํŠธ๋Š” ์•ฝ๋ฌผ ๊ฐ„ ์ถฉ๋Œ ์—ฌ๋ถ€๋ฅผ ์ ๊ฒ€ํ•˜๋ฉฐ, ๋˜ ๋‹ค๋ฅธ ์—์ด์ „ํŠธ๋Š” ์˜๋ฃŒ์ง„ ๊ฒ€ํ† ๋ฅผ ์œ„ํ•œ ํ™˜์ž ์š”์•ฝ์„ ์ž‘์„ฑํ•œ๋‹ค.

์ตœ๊ทผ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด ๋ฉ€ํ‹ฐ ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ์€ ์ถ”๋ก ๊ณผ ์˜์‚ฌ๊ฒฐ์ • ๋ฒค์น˜๋งˆํฌ์—์„œ ๋‹จ์ผ LLM๋ณด๋‹ค ๋›ฐ์–ด๋‚œ ์„ฑ๊ณผ๋ฅผ ๋‚ด๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์œผ๋ฉฐ, ์—ฐ์‚ฐ ๋น„์šฉ๋„ ๋” ๋‚ฎ์€ ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ๋‹ค. ์˜๋ฃŒ ๋ถ„์•ผ์—์„œ๋Š” ์—์ด์ „ํŠธ ๊ฐ„ ์—ญํ• ์„ ๋ถ„๋ฆฌํ•ด ์„œ๋กœ ๊ฒฐ๊ณผ๋ฅผ ๊ต์ฐจ ๊ฒ€์ฆํ•˜๋Š” ๊ตฌ์กฐ๋ฅผ ๊ฐ–์ถœ ์ˆ˜ ์žˆ๋‹ค. ์ด๋Š” ๊ฐ ์—์ด์ „ํŠธ์˜ ์ฑ…์ž„ ๋ฒ”์œ„๊ฐ€ ๋ช…ํ™•ํ•˜๊ฒŒ ์ •์˜๋ผ ์žˆ์–ด, ํ•œ ์ง€์ ์˜ ํŒ๋‹จ ์˜ค๋ฅ˜๊ฐ€ ๋‹ค์Œ ๋‹จ๊ณ„๋กœ ๊ทธ๋Œ€๋กœ ์ „์ด๋˜๋ฉฐ ์ปค์ง€๋Š” ์œ„ํ—˜์„ ์ค„์ธ๋‹ค.

์ž„์ƒ ์˜์‚ฌ๊ฒฐ์ • ์ง€์›, ๋ถ„๋ฅ˜ ์ž๋™ํ™”, ํ™˜์ž ์ฐธ์—ฌ ์˜์—ญ์—์„œ๋Š” ๋ฉ€ํ‹ฐ ์—์ด์ „ํŠธ ์•„ํ‚คํ…์ฒ˜๊ฐ€ ํ‘œ์ค€์œผ๋กœ ์ž๋ฆฌ ์žก์„ ๊ฐ€๋Šฅ์„ฑ์ด ํฌ๋‹ค. ์ด๋Š” ๋‹จ์ผํ•œ ๋ชจ๋ธ์— ์˜์กดํ•˜๋Š” ๋Œ€์‹  ์—ฌ๋Ÿฌ ์ „๋ฌธ ์ธ๋ ฅ์ด ํ˜‘์—…ํ•˜๋Š” ์‹ค์ œ ์ž„์ƒ ํ˜„์žฅ์˜ ์šด์˜ ๋ฐฉ์‹๊ณผ ๊ตฌ์กฐ์ ์œผ๋กœ ๋‹ฎ์•„์žˆ๊ธฐ ๋•Œ๋ฌธ์ด๋‹ค.

๋„๋ฉ”์ธ ํŠนํ™” ๋ชจ๋ธ์˜ ํ•„์š”์„ฑ

GPT-5๋‚˜ ํด๋กœ๋“œ ๊ฐ™์€ LLM์€ ๊ฐ•๋ ฅํ•˜์ง€๋งŒ, ์˜๋ฃŒ ๋ถ„์•ผ์—์„œ๋Š” ๋„๋ฉ”์ธ์— ํŠนํ™”๋œ ์ˆ˜์ค€์˜ ์ •ํ™•์„ฑ๊ณผ ์„ค๋ช… ๊ฐ€๋Šฅ์„ฑ์ด ์š”๊ตฌ๋œ๋‹ค. ์—ฐ๊ตฌ ๊ฒฐ๊ณผ๋ฅผ ๋ณด๋ฉด, ์ƒ์˜ํ•™ ๋ฐ์ดํ„ฐ์™€ ๋ถ„๋ฅ˜ ์ฒด๊ณ„(์˜จํ†จ๋กœ์ง€), ์ž„์ƒ ์›Œํฌํ”Œ๋กœ์šฐ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•™์Šต๋œ ์ „๋ฌธ ๋ชจ๋ธ์ด ์•ˆ์ „์„ฑ๊ณผ ๊ด€๋ จ์„ฑ ์ธก๋ฉด์—์„œ ๋ฒ”์šฉ ๋ชจ๋ธ์„ ์ผ๊ด€๋˜๊ฒŒ ์•ž์„œ๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚˜๊ณ  ์žˆ๋‹ค.

ํŠน์ • ์˜๋ฃŒ ๋ถ„์•ผ์— ๋งž์ถฐ ์กฐ์ •๋œ AI๋Š” ์ด๋ฏธ ์ž„์ƒ ๋ฌธ์„œํ™”๋‚˜ ์‹ ์•ฝ ๊ฐœ๋ฐœ๊ณผ ๊ฐ™์€ ์˜์—ญ์—์„œ ๋ฒ”์šฉ ๋ชจ๋ธ๋ณด๋‹ค ๋” ๋‚˜์€ ์„ฑ๊ณผ๋ฅผ ๋‚ด๊ณ  ์žˆ๋‹ค. ์ด๋“ค ์‹œ์Šคํ…œ์€ ์˜๋ฃŒ ์ „๋ฌธ ์šฉ์–ด๋ฅผ ์ดํ•ดํ•˜๊ณ , FHIR์™€ ๊ฐ™์€ ์ „์ž์˜๋ฌด๊ธฐ๋ก(EHR) ํ‘œ์ค€๊ณผ ์ง์ ‘ ์—ฐ๋™๋˜๋ฉฐ, ์•ฝ๋ฌผ ์šฉ๋Ÿ‰ ๊ธฐ์ค€์ด๋‚˜ ์ž„์ƒ ์ง„๋ฃŒ ์ง€์นจ ๊ฐ™์€ ์˜๋ฃŒ ๊ทœ์น™์„ ๋ฐ˜์˜ํ•ด ์ž‘๋™ํ•œ๋‹ค.

๊ทœ์ œ๊ฐ€ ๊ฐ•ํ™”๋ ์ˆ˜๋ก, ํ™˜์ž ๋ฐ์ดํ„ฐ๋ฅผ ๋‹ค๋ฃจ๋Š” ์˜๋ฃŒ ๊ธฐ๊ด€์—์„œ๋Š” ๋„๋ฉ”์ธ ํŠนํ™” AI๊ฐ€ ์‚ฌ์‹ค์ƒ ์œ ์ผํ•œ ์„ ํƒ์ง€๊ฐ€ ๋  ์ˆ˜ ์žˆ๋‹ค. 2026๋…„์—๋Š” ์˜๋ฃŒ๋ฅผ ๋„˜์–ด ๊ธˆ์œต๊ณผ ๋ฒ•๋ฅ  ๋“ฑ ๊ทœ์ œ ์‚ฐ์—… ์ „๋ฐ˜์—์„œ ๋ถ„์•ผ๋ณ„ ํŠนํ™” ๋ชจ๋ธ์ด ์ฃผ๋„๊ถŒ์„ ์žก๊ณ , ๋ฒ”์šฉ LLM์€ ์ €์œ„ํ—˜ ํ–‰์ • ์—…๋ฌด๋‚˜ ์†Œ๋น„์ž์šฉ ๊ณผ์ œ์— ํ•œ์ •๋  ๊ฐ€๋Šฅ์„ฑ๋„ ์žˆ๋‹ค.

ํ•ต์‹ฌ ์ธํ”„๋ผ๋กœ ์ž๋ฆฌ ์žก๋Š” ๊ฑฐ๋ฒ„๋„Œ์Šค์™€ ์‹ ๋ขฐ

AI ์‹œ์Šคํ…œ์ด ์ ์  ๋ณต์žกํ•ด์ง€๋ฉด์„œ, ๊ฑฐ๋ฒ„๋„Œ์Šค๋Š” ๋” ์ด์ƒ ๊ทœ์ œ ๋Œ€์‘์„ ์œ„ํ•œ ์ฒดํฌ๋ฆฌ์ŠคํŠธ์— ๋จธ๋ฌผ์ง€ ์•Š๋Š”๋‹ค. ์ด์ œ ๊ฑฐ๋ฒ„๋„Œ์Šค๋Š” ์•„ํ‚คํ…์ฒ˜ ์ž์ฒด์˜ ์ผ๋ถ€๋กœ ์ž๋ฆฌ ์žก๊ณ  ์žˆ๋‹ค. ๋”œ๋กœ์ดํŠธ๊ฐ€ ์กฐ์‚ฌํ•œ ๊ฒฐ๊ณผ์— ๋”ฐ๋ฅด๋ฉด, ์˜๋ฃŒ ๊ธฐ์—… ๊ฒฝ์˜์ง„์€ 2025๋…„ AI ๋„์ž…์—์„œ ๊ฑฐ๋ฒ„๋„Œ์Šค์™€ ๋ฆฌ์Šคํฌ ๊ด€๋ฆฌ๋ฅผ ์ตœ์šฐ์„  ๊ณผ์ œ๋กœ ๊ผฝ์•˜๋‹ค. ์ด๋Ÿฐ ํ๋ฆ„์€ 2026๋…„์— ๋”์šฑ ๋šœ๋ ทํ•ด์งˆ ์ „๋ง์ด๋‹ค.

์ •๋ณด ์ถ”์ถœ ์—”์ง„์ด๋“  ๋Œ€ํ™”ํ˜• ์ธํ„ฐํŽ˜์ด์Šค๋“ , ๋ชจ๋“  AI ๋ชจ๋“ˆ์€ ๋ˆ„๊ฐ€ ์–ด๋–ค ๋ฐ์ดํ„ฐ๋กœ ํ•™์Šตํ–ˆ๊ณ  ์–ด๋–ค ๊ฒ€์ฆ ์ง€ํ‘œ๋ฅผ ๊ฑฐ์ณค๋Š”์ง€๋ฅผ ์ฆ๋ช…ํ•  ์ˆ˜ ์žˆ์–ด์•ผ ํ•œ๋‹ค. ์ด๋•Œ ๋ฐ์ดํ„ฐ ์ถœ์ฒ˜์™€ ์„ค๋ช… ๊ฐ€๋Šฅ์„ฑ์€ ์„ ํƒ ์‚ฌํ•ญ์ด ์•„๋‹ˆ๋ผ ํ•„์ˆ˜ ์š”๊ฑด์ด ๋œ๋‹ค. ๊ธฐ์—…์€ ๋ชจ๋ธ์ด ์‹ค์ œ ์šด์˜ ๋ฐ์ดํ„ฐ์— ์ ์šฉ๋˜๊ธฐ ์ „์— ํŽธํ–ฅ ์—ฌ๋ถ€, ์„ฑ๋Šฅ ๋ณ€ํ™”, ์•ˆ์ •์„ฑ์„ ์ ๊ฒ€ํ•˜๊ธฐ ์œ„ํ•œ ๋‚ด๋ถ€ ๋ ˆ๋“œํŒ€์„ ์šด์˜ํ•˜๊ฒŒ ๋  ๊ฒƒ์ด๋‹ค.

์ด๋Ÿฐ ๋ณ€ํ™”๋Š” ์ƒ์„ฑํ˜• AI๋ฅผ ์‹คํ—˜์  ๊ธฐ์ˆ ์—์„œ ๋ฒ—์–ด๋‚˜, ๊ธฐ๋ก๊ณผ ์ฑ…์ž„์„ ๊ณต์‹ ๊ด€๋ฆฌํ•˜๋Š” ์‹œ์Šคํ…œ์œผ๋กœ ์ „ํ™˜์‹œํ‚ค๊ณ  ์žˆ๋‹ค. ์•ž์„  ์˜๋ฃŒ ๊ธฐ๊ด€์€ ์ด๋ฏธ ์Šน์ธ๋œ ๋ชจ๋ธ๊ณผ ๋ฐ์ดํ„ฐ ์†Œ์Šค, ๊ฑฐ๋ฒ„๋„Œ์Šค ์ฑ…์ž„์ž๋ฅผ ์ •๋ฆฌํ•œ AI ๋ ˆ์ง€์ŠคํŠธ๋ฆฌ๋ฅผ ์šด์˜ํ•˜๊ณ  ์žˆ๋‹ค. ์ด๋Š” ์†Œํ”„ํŠธ์›จ์–ด ์ž์žฌ ๋ช…์„ธ์„œ์™€ ์œ ์‚ฌํ•œ ๊ฐœ๋…์ด๋‹ค. ๋‚ด๋…„์ด๋ฉด ์ด๋Ÿฌํ•œ ๊ด€๋ฆฌ ๋ฐฉ์‹์ด ์—…๊ณ„ ์ „๋ฐ˜์—์„œ ํ‘œ์ค€์œผ๋กœ ์ž๋ฆฌ ์žก๊ฑฐ๋‚˜, ์ตœ์†Œํ•œ ๊ทธ์— ์ค€ํ•˜๋Š” ๋‹จ๊ณ„์— ์ด๋ฅผ ๊ฒƒ์œผ๋กœ ์˜ˆ์ƒ๋˜๊ณ  ์žˆ๋‹ค.

์‹ค์ œ ์ ์šฉ ์‚ฌ๋ก€

๋‹น๋‡จ๋ณ‘์ด๋‚˜ ์‹ฌ๋ถ€์ „๊ณผ ๊ฐ™์€ ๋งŒ์„ฑ ์งˆํ™˜์„ ํ•จ๊ป˜ ์•“๋Š” ํ™˜์ž๋ฅผ ๊ด€๋ฆฌํ•˜๋Š” ์ƒํ™ฉ์„ ๋– ์˜ฌ๋ ค ๋ณผ ์ˆ˜ ์žˆ๋‹ค. ์ด ํ™˜์ž์˜ ๋ฐ์ดํ„ฐ๋Š” ์ˆ˜๋…„์— ๊ฑธ์นœ ๊ฒ€์‚ฌ ๊ฒฐ๊ณผ์™€ ์˜์ƒ ์ž๋ฃŒ, ์ฒ˜๋ฐฉ ๊ธฐ๋ก, ์ž„์ƒ ๋…ธํŠธ๊ฐ€ ์—ฌ๋Ÿฌ EHR ์‹œ์Šคํ…œ์— ํฉ์–ด์ ธ ์žˆ๋‹ค. ๊ธฐ์กด ๋ฐฉ์‹์ด๋ผ๋ฉด ์ด ๋ชจ๋“  ๊ธฐ๋ก์„ ํ•˜๋‚˜์˜ LLM์— ์ž…๋ ฅํ•œ ๋’ค โ€œ๋‹ค์Œ ๋‹จ๊ณ„๋กœ ๋ฌด์—‡์„ ํ•ด์•ผ ํ•˜๋Š”๊ฐ€โ€๋ผ๊ณ  ๋ฌป๋Š” ์‹์ด์—ˆ๋‹ค.

๋ชจ๋“ˆํ˜• ๋ฉ€ํ‹ฐ ์—์ด์ „ํŠธ ์ ‘๊ทผ๋ฒ•์€ ์ „ํ˜€ ๋‹ค๋ฅด๊ฒŒ ์ž‘๋™ํ•œ๋‹ค. ์ •๋ณด ์ถ”์ถœ ์—์ด์ „ํŠธ๊ฐ€ ํ™˜์ž์˜ ๋ณ‘๋ ฅ์„ ๊ตฌ์กฐํ™”ํ•˜๊ณ , ์ถ”๋ก  ์—์ด์ „ํŠธ๊ฐ€ ์œ„ํ—˜ ํŒจํ„ด์„ ์‹๋ณ„ํ•˜๋ฉฐ, ์•ฝ๋ฌผ ๊ฒ€ํ†  ์—์ด์ „ํŠธ๊ฐ€ ๊ธˆ๊ธฐ ์‚ฌํ•ญ์„ ํ‘œ์‹œํ•œ๋‹ค. ๋Œ€ํ™”ํ˜• ์—์ด์ „ํŠธ๋Š” ์ด๋Ÿฌํ•œ ๋ถ„์„ ๊ฒฐ๊ณผ๋ฅผ ์˜๋ฃŒ์ง„์ด ์ดํ•ดํ•˜๊ธฐ ์‰ฌ์šด ์–ธ์–ด๋กœ ์„ค๋ช…ํ•œ๋‹ค. ์—ฌ๊ธฐ์— ๊ฑฐ๋ฒ„๋„Œ์Šค ๋ ˆ์ด์–ด๊ฐ€ ๋ชจ๋“  ์ถ”๋ก  ๊ณผ์ •์„ ์ถ”์ ํ•ด ํˆฌ๋ช…์„ฑ๊ณผ ๊ฐ์‚ฌ ๊ฐ€๋Šฅ์„ฑ์„ ํ™•๋ณดํ•œ๋‹ค.

์ด๋Š” ์„ค๊ณ„ ๋‹จ๊ณ„๋ถ€ํ„ฐ ์„ค๋ช… ๊ฐ€๋Šฅ์„ฑ์„ ์ „์ œ๋กœ ํ•˜๋ฉฐ, ๊ทœ์ œ ๊ฒ€์ฆ์—๋„ ์œ ์—ฐํ•˜๊ฒŒ ๋Œ€์‘ํ•  ์ˆ˜ ์žˆ๋‹ค. ๋™์‹œ์— ์‹ค์ œ ์˜๋ฃŒํŒ€์ด ํ˜‘์—…ํ•˜๋Š” ๋ฐฉ์‹๊ณผ๋„ ๊ตฌ์กฐ์ ์œผ๋กœ ๋งž๋‹ฟ์•„ ์žˆ๋‹ค. ์ •๋ฐ€๋„์™€ ์ฑ…์ž„์ด ์š”๊ตฌ๋˜๋Š” ์žฅ๊ธฐ๊ฐ„์˜ ํ™˜์ž ์น˜๋ฃŒ ๊ณผ์ • ๋ถ„์„์—๋Š” ๋ฉ€ํ‹ฐ ์—์ด์ „ํŠธ ๊ธฐ๋ฐ˜์˜ ๋„๋ฉ”์ธ ํŠนํ™” ํ”„๋ ˆ์ž„์›Œํฌ๊ฐ€ ํ›จ์”ฌ ๋” ํšจ๊ณผ์ ์ด๋‹ค.

์ฐจ์„ธ๋Œ€ ์ƒ์„ฑํ˜• AI์˜ ์ฃผ๋ชฉ๋ฐ›๋Š” ์„ฑ๊ณต ์‚ฌ๋ก€๋Š” ๊ฐ€์žฅ ํฐ ๋ชจ๋ธ์„ ๋„์ž…ํ•œ ๊ธฐ์—…์ด ์•„๋‹ˆ๋ผ, ๊ฐ€์žฅ ํšจ์œจ์ ์ด๊ณ  ํˆฌ๋ช…ํ•˜๋ฉฐ ๋„๋ฉ”์ธ์— ๋งž๊ฒŒ ์„ค๊ณ„๋œ ์‹œ์Šคํ…œ์„ ๊ตฌ์ถ•ํ•œ ๊ณณ์ด ๋  ๊ฒƒ์ด๋‹ค. ์˜๋ฃŒ ๋ถ„์•ผ ๋ฆฌ๋”์—๊ฒŒ ์ด์ œ ํ•ต์‹ฌ ์งˆ๋ฌธ์€ โ€œ์–ด๋–ค LLM์„ ๋„์ž…ํ•  ๊ฒƒ์ธ๊ฐ€โ€๊ฐ€ ์•„๋‹ˆ๋ผ โ€œAI ์‹œ์Šคํ…œ์„ ์–ด๋–ป๊ฒŒ ํ˜‘์—…์‹œํ‚ค๊ณ , ๊ฑฐ๋ฒ„๋„Œ์Šค๋ฅผ ์œ ์ง€ํ•˜๋ฉฐ, ํ•จ๊ป˜ ํ™•์žฅํ•  ๊ฒƒ์ธ๊ฐ€โ€๋กœ ๋ฐ”๋€Œ๊ณ  ์žˆ๋‹ค. ์ด๊ฒƒ์ด ์•ˆ์ „ํ•˜๊ณ  ์„ค๋ช… ๊ฐ€๋Šฅํ•œ AI๋กœ ๋‚˜์•„๊ฐ€๋Š” ์ƒˆ๋กœ์šด ๋ฐฉ์‹์ด๋‹ค.
dl-ciokorea@foundryco.com

Multi-agent, domain-specific and governed models will define healthcare genAI in 2026

9 January 2026 at 11:23

The early adoption pattern of generative AI โ€” dumping all available data into a large language model (LLM) and asking it to โ€œreasonโ€ โ€” is proving unsustainable. Costs are ballooning, accuracy is wavering and compliance is becoming unmanageable. What began as a promising value-add is now fighting against the realities of enterprise-scale deployment.

As we start a new year, the next chapter of genAI is upon us and it wonโ€™t be defined by bigger models. Instead of years past, 2026 will mark the transition to truly smarter systems: modular, domain-specific and governed architectures that deliver measurable business value.

Hereโ€™s whyโ€ฆ

Itโ€™s the end of one-size-fits-all AI

Over the past two years, many organizations took a brute-force approach to genAI. In healthcare, for example, teams fed every chart, lab and note into a single LLM, then asked it to summarize or predict. It was fast to prototype, but models hit context limits, inference costs skyrocketed and outputs often lacked clinical-grade accuracy.

AI models excel at black and white tasks like mathematical problems or standardized tests, but still struggle with complex reasoning benchmarks, according to Stanford Universityโ€™s AI Index Report. They often fail to reliably solve logic tasks, limiting their effectiveness in high-stakes settings where precision is critical.

To mitigate this, smart organizations will adopt modular pipelines instead. These pipelines separate information extraction, reasoning and conversation into distinct, optimized buckets. One model extracts clinical entities from free-text notes; another performs structured reasoning over that data; a third delivers results via a natural-language interface. Each module can be tuned, audited and improved independently.

This right-tool-for-the-right-job approach makes systems faster, safer and far more transparent. This is a critical requirement when AI outputs operate in highly regulated industries like medicine.

The rise of multi-agent collaboration teams

The next major evolution will come from multi-agent systems โ€” networks of smaller, specialized AI models that coordinate across tasks. Think of them as digital teams. Keeping with the healthcare theme: one agent monitors lab trends, another checks for medication conflicts and a third drafts a patient summary for clinician review.

Recent studies show that multi-agent systems outperform monolithic LLMs on reasoning and decision-making benchmarks, often with lower computational costs. In healthcare, they also bring built-in checks and balances. Each agentโ€™s scope is clearly defined, reducing the risk of compounding errors.

Expect multi-agent architectures to become the standard pattern for clinical decision support, triage automation and patient engagement. Why? Because it reflects how real-world clinical settings already operate โ€” through collaboration among specialists, rather than a single all-knowing model.

Domain-specific models leap ahead

General-purpose LLMs like GPT-5 and Claude are powerful, but healthcare demands domain-specific accuracy and explainability. Itโ€™s in the research: specialized models trained on biomedical data, ontologies and clinical workflows consistently outperform general models in safety and relevance.

AI tuned for specific medical subfields is already outperforming general models in tasks like clinical documentation and drug discovery. These systems know medical vocabulary, integrate directly with electronic health record (EHR) standards like FHIR and encode domain constraints such as dosage limits and clinical guidelines.

As regulatory expectations tighten, domain-specific AI will become the only viable option for healthcare organizations handling patient data. In 2026, weโ€™ll see specialty-specific models dominate regulated environments from healthcare to finance and law, while general LLMs remain limited to low-risk administrative or consumer tasks.

Governance and trust as core infrastructure

As AI systems grow more complex, governance is no longer a compliance checkbox โ€” itโ€™s part of the architecture itself. Healthcare executives surveyed by Deloitte ranked governance and risk management as top priorities for AI adoption in 2025. That emphasis will deepen in 2026.

Each AI module, whether an extraction engine or conversational layer, must have a documented lineage proving who trained it, on what data and with what validation metrics. Provenance and explainability will become mandatory features, not optional add-ons. Organizations will deploy internal red-teaming to test bias, drift and robustness before models touch production data.

This shift is transforming genAI from an experimental capability into an auditable system of record. The most forward-looking health systems already maintain AI registries (similar to software bills of materials) listing approved models, data sources and governance owners. By next year, that practice will be standardized or well on its way.

How this looks in the real world

Consider the challenge of managing a patient with chronic conditions such as diabetes and heart failure. Their data spans years of lab results, imaging, prescriptions and clinical notes scattered across multiple EHRs. The old approach would be to dump the entire record into an LLM and ask, โ€œWhat should happen next?โ€

A modular, multi-agent approach works differently. An extraction agent structures the patientโ€™s history, a reasoning agent identifies risk patterns, a medication-review agent flags contraindications, and a conversational agent explains the findings to clinicians in plain language. A governance layer tracks every inference, ensuring transparency and auditability.

This second architecture is explainable by design, adapts to regulatory scrutiny and mirrors how care teams collaborate in reality. For longitudinal patient-journey analysis, which requires precision and accountability, a multi-agent, domain-specific framework will fare far better. Which would you prefer as the patient?

The success stories in the next phase of genAI wonโ€™t be the ones deploying the largest models, but the ones that engineer the most efficient, transparent and domain-tuned systems. For healthcare leaders, the key question is no longer โ€œWhich LLM should we buy?โ€ but โ€œHow do our AI systems collaborate, govern and scale together?โ€ This is the new way for safe, responsible and explainable AI.

This article is published as part of the Foundry Expert Contributor Network.
Want to join?

โŒ
โŒ