A 28-year-old woman, casualty at 2 a.m. Heart rate 142. Temperature 39.4. Blood pressure 88/52. Agitated, sweating, vomiting. The on-call resident is also covering a sepsis admission and a paediatric seizure. She assesses the patient quickly: febrile, tachycardic, hypotensive. Sepsis differential. She orders blood cultures, fluids, paracetamol.

The senior, paged for an unrelated reason, walks past at 2:18 a.m. and looks at the patient. "Is her thyroid status known?"

It is not. The senior asks for a TSH and free T4 STAT. They are not in the standard sepsis workup at this hospital.

The TSH comes back at 4:10 a.m. — undetectable. Free T4 — 7.8 (normal upper limit 1.8). The patient is in thyroid storm. The treatment changes immediately: propylthiouracil, propranolol, hydrocortisone, iodine after one hour.

The patient survives. She would not have, on the sepsis pathway.

This case is not about a bad resident. It is about the trap of the obvious differential. Febrile + tachycardic + hypotensive matches sepsis. It also matches thyroid storm, sympathomimetic toxicity, severe pheochromocytoma, and a few others. The discriminating questions — known thyroid disease, exophthalmos on exam, tremor, recent iodine load — are not in the standard sepsis bundle.

A context-aware AI consult, given the same presentation, returns the broader differential by default. The Medows response for "28F, fever 39.4, HR 142, BP 88/52, agitated" includes thyroid storm in the differential alongside sepsis, with a one-line note on the discriminating exam features and the TSH/T4 order.

This is not because the AI is smarter than the resident. It is because the AI is not also running the paediatric seizure and the sepsis admission in parallel. It has the cognitive bandwidth to consider the second-most-likely differential. The resident, on a fourth-bed-in-thirty-minutes night, does not.

That bandwidth is what an AI workspace adds. Not omniscience. Just the room to consider the case that's not the obvious one.