Write it down before you automate it.

Larry Benz has run dental and physical therapy platforms. In his Substack, The Operator, he describes a back office that now runs on a folder of written procedures and an AI agent that follows them. A daily brief lands at 6 AM. His inboxes get sorted overnight into reply now, know about, and cleared. Scheduling requests get a drafted reply offering three open times. He cancelled a paid scheduling tool and replaced it with a written rulebook of his own preferences.

He watched it for a week before he trusted it. By his count, the agent scanned his mail a little over 800 times, drafted exactly one reply, and stayed away from all 23 threads that touched money. His guardrail is simple. The agent drafts, a human sends. Anything that leaves the building, touches a patient, or involves money stays with a person.

A Journal of Biomechanics study ran 10 people with Parkinson's disease and 10 healthy adults through a marker-based lab system and two commercial camera setups at the same time. Joint angles and gait speed tracked the lab well, with correlations above 0.7. But errors ran from about 5 to 19 degrees depending on the joint, and the smallest change the camera systems could reliably detect went as high as 28.57 degrees. Errors got worse with fewer cameras, simpler body models, and in the Parkinson's group.

A separate systematic review in Cureus looked at randomized trials that used motion capture and wearable sensors in rehab. The authors report better gait and balance outcomes in everything from neuro rehab to cancer rehab. The trials were too different from each other to pool, though.

An npj Digital Medicine meta-synthesis pulled together 13 qualitative studies from nine countries on how health professionals see AI-assisted clinical decisions. They liked the promise of faster diagnosis and easier workflow. The objections were consistent. They couldn't see how the AI reached its answer. Nobody was clear on who's responsible when it's wrong. And they worried about privacy, consent, over-reliance, and losing the human side of care.

A survey of 196 physiotherapy students at three universities in Türkiye, published in Work, found that AI literacy, excitement about AI, trust in AI tools, and believing AI belongs in the curriculum went with more positive attitudes. Higher AI anxiety went with more negative ones, and curiosity went with fewer. Demographics didn't predict either.

A team in Italy built an evaluation framework aligned with the EU AI Act and ran 15 large and 2 small language models through 32 clinical nursing prompts, graded blind against answers from a Delphi expert panel. A safety screen knocked out 11 of the 17 right away. DeepSeek-R1, Perplexity Sonar, Mistral Large 2, and Qwen2.5-14B didn't resist even explicit harmful prompts. Only Anthropic's Sonnet models were rated "recommended" across the board. The paper is in JMIR Medical Informatics.

Pick one task you or your front desk do every week, like the Monday numbers check. Write it out in plain language: what to pull, what counts as a problem, what the finished output looks like, and what the AI is never allowed to do (send, pay, or touch a patient record). Hand it to an AI tool you already use and compare its draft to how you'd do it. The gaps show you what your procedure was missing. Keep patient information out of it unless the tool is covered by a BAA.