The treatment plans that drift.

A patient comes in with acute low back pain. The clinician writes a sensible plan. Pain control for the first week, mobility work the second, progressive loading after that. The patient improves. By week three the pain is gone, the patient is moving well, and the appointments continue. Same exercises. Same routine. Nobody renegotiates.

By week six, the patient has stopped improving. Not because anything is wrong, but because the plan that worked for an acute presentation has quietly become the plan for a return-to-sport problem it was never designed to solve. The clinician feels it. The patient feels it. Nobody quite knows when to call the discharge.

This is what plan drift looks like. It rarely shows up as a failure. It shows up as inertia.

Why plans drift

Treatment plans drift because they are usually written once, on day one, when the dominant complaint is pain. The plan reflects that moment. It contains exercises chosen for the patient who walked in this morning, not for the patient who will walk in next month.

The honest reason plans don't get rewritten is structural. Each appointment runs 30 minutes. The clinician needs to assess, treat, document, schedule the next visit, and try to leave time to use the bathroom. Rewriting the plan, with new exercise selection, new progression criteria, new outcome targets, is a 10-minute task. There is no 10 minutes.

So the plan from day one becomes the plan for week three, week six, week ten. Not because anyone thinks that is right, but because nobody is paid to renegotiate it.

What a plan should actually contain

A plan that does not drift looks structurally different from a plan that does. It has phases (acute, sub-acute, return to function, discharge). Each phase has a goal that is not just “reduce pain”. Each phase has measurable exit criteria so the clinician knows when to move on. And each phase has a scheduled reassessment so the plan gets a chance to evolve.

This is not new. Phased rehabilitation has been in the literature for decades. Every textbook teaches it. What the textbooks don't teach is how to build a phased plan in the seven minutes between two patients on a busy Tuesday afternoon.

That gap, between what the literature recommends and what the clinic allows, is where most plans quietly become weekly exercise lists.

Where AI is useful, and only there

AI is good at structure. Given a diagnosis, a body region, an age, and a clinical goal, a well-built model can draft a phased plan in one pass. It pulls exercises from a library, sequences them across the phases, sets progression criteria, and schedules reassessment points. It does not get tired in the afternoon. It does not skip the reassessment because the patient said they were “fine”.

What it produces is not a finished plan. It is a draft. The clinician reads it, edits it, replaces the manual therapy choice with the technique that fits this patient's tissue feel, drops the exercise the patient can't do because of a knee they didn't mention, and ships the revised version. The AI saved the seven minutes of structural writing. The clinician keeps the judgment.

This is the only honest framing. AI plans the structure. The physiotherapist adapts the content. Anyone selling more than that is selling something that doesn't work in a real clinic.

The part that matters most: reassessment

The single biggest win from AI-assisted planning is not the initial plan. It is the reassessment. A plan with built-in checkpoints will flag, automatically, when measured outcomes are not progressing toward the target. The clinician sees a banner that says “ODI hasn't moved in two weeks, consider re-screening or progressing” instead of having to remember to ask.

This is where the actual difference shows up in patient outcomes. Not in the brilliance of the first plan, but in how reliably it gets revised. Phase 1 plans that quietly extend into phase 4 territory are the silent failure mode of modern physiotherapy. A system that interrupts the inertia is worth more than a perfect day-one prescription.

What this looks like in Healui

We built treatment planning around the things that go wrong, not the things that go right. The five agents that compose a Healui plan are specialised. One handles exercise prescription from a library of 342+ evidence-based exercises. Another handles manual therapy. A third is purely responsible for progression criteria, the criteria that decide when the patient moves from phase to phase. A fourth sequences PROMs across the visit timeline so reassessment is built in. A fifth checks for clinical contraindications.

The five agents collaborate. The output is a phased plan with explicit checkpoints, ready for the clinician to review. The clinician edits and finalises. The plan then re-evaluates itself against measured outcomes at each visit, and surfaces when the patient is not on track.

None of this is glamorous. It is the unglamorous structural work that most clinics don't have time to do by hand. That is exactly why it is the work that benefits most from being done by a model.

The honest closing

A treatment plan is a contract between a clinician and a patient. The patient agrees to do the work. The clinician agrees to know when the plan needs to change. Both sides are usually doing their part. What's missing, more often than not, is the structural support that lets the clinician keep their side of the contract under a busy afternoon's cognitive load.

That is where AI fits. Not in the brilliance, but in the housekeeping. It is not the part of the work that makes for good clinical stories. It is the part that quietly determines how the patient actually does.