New AI Team Could Finally Make Your Smartwatch Data Make Sense
Splitting health questions among AI helpers halved computing costs — but it's still early.
Right now, if you ask an AI chatbot something like "how's my sleep lately, and should I worry?", it gets your entire health record plus your messy question all at once. It may answer confidently, but you have no idea whether it actually looked at last Tuesday's data or just guessed. This new paper proposes splitting that one big question into small, labeled jobs — called tasks — and handing each to a specialized AI "agent." One agent fetches your data, one analyzes trends over time, one writes advice, and a manager agent makes sure every part of your question got answered. (An agent is simply AI that can take actions on its own, not just chat.)
The test used 10,000 simulated people, each with one month of fake wearable records — no real patients involved. On 1,500 data-lookup questions, the new setup scored 98.3% correct versus 97.9% for a single AI. Barely better. The real win was cost: computing power per question dropped from 6,869 "tokens" (chunks of text the AI processes) to 3,136 — roughly 54% less. On 180 questions with multiple parts, the system answered every part 100% of the time and matched the expected grouping 94.4% of the time.
Why should you care? Wearable data is a firehose, and your doctor gets about 15 minutes with you. If AI can reliably surface something like "your resting heart rate crept up 8 beats per minute over three weeks," that is genuinely useful. Cheaper processing also means this could run affordably on your phone or in a clinic app rather than on expensive servers. The system also scored higher on transparency — showing which records support each answer — which matters if you're going to trust health advice at all.
The honest catch: everything was tested on synthetic, made-up data, so nobody knows how it performs on real people, real sleep, or real heart conditions. The researchers themselves say that health advice generation and validation on real wearable data remain open challenges. The system also became more trustworthy and transparent, but its "actionability" — whether the advice actually helps you do something — did not consistently improve. So don't hand your health over to your watch yet. Still, the direction is clear: for messy personal data, many small specialized AIs may beat one big one.
- Instead of one AI juggling a whole health record and a messy question, this system splits the job among specialized AI agents for finding data, spotting trends, and giving advice.
- It cut computing power per question by about 54% (from 6,869 to 3,136 tokens) while scoring 98.3% accuracy on data lookups — meaning cheaper, faster health apps.
- All testing used 10,000 fake users on synthetic data, and the quality of actual health advice did not reliably improve — so real-world trust is still unproven.
Why It Matters
Cheaper, clearer AI could turn your smartwatch numbers into trends a doctor can actually use — someday, not today.