Gabe Wilson MD

AI in 2026 cannot palpate an abdomen, intubate a patient, feel a thyroid nodule, test a patellar reflex, reduce a dislocated shoulder, perform a colonoscopy, or deliver a baby. That is not a temporary limitation. It is structural. When we scored AI capability across seven clinical dimensions for 240 visit reasons in 20 specialties, the physical/procedural dimension averaged 1.5 out of 5. The cognitive dimensions averaged 3.0 to 4.1. No specialty broke 2.0 on procedure. Not one. History-taking averaged 4.1 — approaching specialist level. Patient communication 3.6. Follow-up management 3.5. Documentation, which runs through every workflow component, is arguably where AI already outperforms most physicians in speed and completeness. The 2.6-point gap between the cognitive ceiling and the procedural wall is not closing with larger language models. Language models do not have hands. Closing that gap requires robotics, haptic sensing, and physical infrastructure at clinical scale — none of which exists beyond narrow research applications. This matters for how we think about workforce planning. The specialties in Tier 3 of our ranking — Ophthalmology, General Surgery, ENT, Emergency Medicine, Orthopedic Surgery, Anesthesiology — are not there because AI cannot reason about their clinical problems. It can. They are in Tier 3 because the physician's physical presence is the treatment. You cannot automate a knee replacement. You cannot automate airway rescue. The specialties in Tier 1 — Radiology, Internal Medicine, Dermatology, Family Medicine, Endocrinology — are there because their workflows are dominated by cognition, synthesis, and documentation, with physical intervention consuming a smaller share of total effort. The implication is straightforward. AI's near-term value is not about replacing any specialty. It is about absorbing the cognitive and administrative burden that consumes 40-60% of every physician's workday across every specialty. The procedural work stays human. The paperwork does not have to. Health systems investing in AI as a documentation, intake, and decision-support engine will see returns now. Health systems waiting for AI to replace proceduralists will be waiting a long time. Post 3 of a series. Post 1: consensus ranking. Post 2: adversarial reconciliation methodology.

GLP1-agonists for the win. Again.

NEWS: Waymo has just announced that their autonomous fleet has now driven a combined 170 million rider-only miles without a human driver as of Dec 2025, up from 127 million in Sept 2025. That's 467,000 miles per day on average. Waymo also released updated safety stats. Locations: • LA: 37.8 million • SF: 53.5 million • PHX: 68.6 million • Austin: 10.7 million

AI in 2026 cannot palpate an abdomen, intubate a patient, feel a thyroid nodule, test a patellar reflex, reduce a dislocated shoulder, perform a colonoscopy, or deliver a baby. That is not a temporary limitation. It is structural. When we scored AI capability across seven clinical dimensions for 240 visit reasons in 20 specialties, the physical/procedural dimension averaged 1.5 out of 5. The cognitive dimensions averaged 3.0 to 4.1. No specialty broke 2.0 on procedure. Not one. History-taking averaged 4.1 — approaching specialist level. Patient communication 3.6. Follow-up management 3.5. Documentation, which runs through every workflow component, is arguably where AI already outperforms most physicians in speed and completeness. The 2.6-point gap between the cognitive ceiling and the procedural wall is not closing with larger language models. Language models do not have hands. Closing that gap requires robotics, haptic sensing, and physical infrastructure at clinical scale — none of which exists beyond narrow research applications. This matters for how we think about workforce planning. The specialties in Tier 3 of our ranking — Ophthalmology, General Surgery, ENT, Emergency Medicine, Orthopedic Surgery, Anesthesiology — are not there because AI cannot reason about their clinical problems. It can. They are in Tier 3 because the physician's physical presence is the treatment. You cannot automate a knee replacement. You cannot automate airway rescue. The specialties in Tier 1 — Radiology, Internal Medicine, Dermatology, Family Medicine, Endocrinology — are there because their workflows are dominated by cognition, synthesis, and documentation, with physical intervention consuming a smaller share of total effort. The implication is straightforward. AI's near-term value is not about replacing any specialty. It is about absorbing the cognitive and administrative burden that consumes 40-60% of every physician's workday across every specialty. The procedural work stays human. The paperwork does not have to. Health systems investing in AI as a documentation, intake, and decision-support engine will see returns now. Health systems waiting for AI to replace proceduralists will be waiting a long time. Post 3 of a series. Post 1: consensus ranking. Post 2: adversarial reconciliation methodology.

AI in 2026 cannot palpate an abdomen, intubate a patient, feel a thyroid nodule, test a patellar reflex, reduce a dislocated shoulder, perform a colonoscopy, or deliver a baby. That is not a temporary limitation. It is structural. When we scored AI capability across seven clinical dimensions for 240 visit reasons in 20 specialties, the physical/procedural dimension averaged 1.5 out of 5. The cognitive dimensions averaged 3.0 to 4.1. No specialty broke 2.0 on procedure. Not one. History-taking averaged 4.1 — approaching specialist level. Patient communication 3.6. Follow-up management 3.5. Documentation, which runs through every workflow component, is arguably where AI already outperforms most physicians in speed and completeness. The 2.6-point gap between the cognitive ceiling and the procedural wall is not closing with larger language models. Language models do not have hands. Closing that gap requires robotics, haptic sensing, and physical infrastructure at clinical scale — none of which exists beyond narrow research applications. This matters for how we think about workforce planning. The specialties in Tier 3 of our ranking — Ophthalmology, General Surgery, ENT, Emergency Medicine, Orthopedic Surgery, Anesthesiology — are not there because AI cannot reason about their clinical problems. It can. They are in Tier 3 because the physician's physical presence is the treatment. You cannot automate a knee replacement. You cannot automate airway rescue. The specialties in Tier 1 — Radiology, Internal Medicine, Dermatology, Family Medicine, Endocrinology — are there because their workflows are dominated by cognition, synthesis, and documentation, with physical intervention consuming a smaller share of total effort. The implication is straightforward. AI's near-term value is not about replacing any specialty. It is about absorbing the cognitive and administrative burden that consumes 40-60% of every physician's workday across every specialty. The procedural work stays human. The paperwork does not have to. Health systems investing in AI as a documentation, intake, and decision-support engine will see returns now. Health systems waiting for AI to replace proceduralists will be waiting a long time. Post 3 of a series. Post 1: consensus ranking. Post 2: adversarial reconciliation methodology.

AI in 2026 cannot palpate an abdomen, intubate a patient, feel a thyroid nodule, test a patellar reflex, reduce a dislocated shoulder, perform a colonoscopy, or deliver a baby. That is not a temporary limitation. It is structural. When we scored AI capability across seven clinical dimensions for 240 visit reasons in 20 specialties, the physical/procedural dimension averaged 1.5 out of 5. The cognitive dimensions averaged 3.0 to 4.1. No specialty broke 2.0 on procedure. Not one. History-taking averaged 4.1 — approaching specialist level. Patient communication 3.6. Follow-up management 3.5. Documentation, which runs through every workflow component, is arguably where AI already outperforms most physicians in speed and completeness. The 2.6-point gap between the cognitive ceiling and the procedural wall is not closing with larger language models. Language models do not have hands. Closing that gap requires robotics, haptic sensing, and physical infrastructure at clinical scale — none of which exists beyond narrow research applications. This matters for how we think about workforce planning. The specialties in Tier 3 of our ranking — Ophthalmology, General Surgery, ENT, Emergency Medicine, Orthopedic Surgery, Anesthesiology — are not there because AI cannot reason about their clinical problems. It can. They are in Tier 3 because the physician's physical presence is the treatment. You cannot automate a knee replacement. You cannot automate airway rescue. The specialties in Tier 1 — Radiology, Internal Medicine, Dermatology, Family Medicine, Endocrinology — are there because their workflows are dominated by cognition, synthesis, and documentation, with physical intervention consuming a smaller share of total effort. The implication is straightforward. AI's near-term value is not about replacing any specialty. It is about absorbing the cognitive and administrative burden that consumes 40-60% of every physician's workday across every specialty. The procedural work stays human. The paperwork does not have to. Health systems investing in AI as a documentation, intake, and decision-support engine will see returns now. Health systems waiting for AI to replace proceduralists will be waiting a long time. Post 3 of a series. Post 1: consensus ranking. Post 2: adversarial reconciliation methodology.

AI in 2026 cannot palpate an abdomen, intubate a patient, feel a thyroid nodule, test a patellar reflex, reduce a dislocated shoulder, perform a colonoscopy, or deliver a baby. That is not a temporary limitation. It is structural. When we scored AI capability across seven clinical dimensions for 240 visit reasons in 20 specialties, the physical/procedural dimension averaged 1.5 out of 5. The cognitive dimensions averaged 3.0 to 4.1. No specialty broke 2.0 on procedure. Not one. History-taking averaged 4.1 — approaching specialist level. Patient communication 3.6. Follow-up management 3.5. Documentation, which runs through every workflow component, is arguably where AI already outperforms most physicians in speed and completeness. The 2.6-point gap between the cognitive ceiling and the procedural wall is not closing with larger language models. Language models do not have hands. Closing that gap requires robotics, haptic sensing, and physical infrastructure at clinical scale — none of which exists beyond narrow research applications. This matters for how we think about workforce planning. The specialties in Tier 3 of our ranking — Ophthalmology, General Surgery, ENT, Emergency Medicine, Orthopedic Surgery, Anesthesiology — are not there because AI cannot reason about their clinical problems. It can. They are in Tier 3 because the physician's physical presence is the treatment. You cannot automate a knee replacement. You cannot automate airway rescue. The specialties in Tier 1 — Radiology, Internal Medicine, Dermatology, Family Medicine, Endocrinology — are there because their workflows are dominated by cognition, synthesis, and documentation, with physical intervention consuming a smaller share of total effort. The implication is straightforward. AI's near-term value is not about replacing any specialty. It is about absorbing the cognitive and administrative burden that consumes 40-60% of every physician's workday across every specialty. The procedural work stays human. The paperwork does not have to. Health systems investing in AI as a documentation, intake, and decision-support engine will see returns now. Health systems waiting for AI to replace proceduralists will be waiting a long time. Post 3 of a series. Post 1: consensus ranking. Post 2: adversarial reconciliation methodology.

Three frontier AI models independently estimated that current AI can assist with 45-77% of clinical workflow across 20 medical specialties — with physician oversight. That is not a claim of autonomous practice. It is a structured capability estimate built from 240 visit reasons, seven scoring dimensions, and three rounds of adversarial reconciliation between GPT 5.4-Pro, Gemini Deep Think, and Grok Heavy. The ceiling is interesting. The floor is the real story. Radiology ranked highest at 77%. Anesthesiology ranked lowest at 45%, with Emergency Medicine and Orthopedic Surgery close behind at 46%. Even in the most procedurally intense, real-time specialties, nearly half the total clinical workflow is cognitively assistable right now. Here is the consensus ranking: Tier 1 — 62-77% AI-assistable: Radiology 77% | Internal Medicine 64% | Dermatology 63% | Family Medicine 62% | Endocrinology 62% Tier 2 — 50-59%: Cardiology 59% | Psychiatry 58% | Gastroenterology 57% | Pediatrics 56% | Pulmonology 54% | OB-GYN 54% | Neurology 53% | Urology 51% | Oncology 50% Tier 3 — 45-49%: Ophthalmology 49% | General Surgery 47% | ENT 46% | Emergency Medicine 46% | Orthopedic Surgery 46% | Anesthesiology 45% The universal bottleneck is physical. Across all 20 specialties, history-taking scored 4.1 out of 5. Physical/procedural work scored 1.5. The near-term role of AI in medicine is not physical replacement. It is cognitive leverage — intake, synthesis, decision support, documentation, patient communication, care coordination. What surprised me most: the three models agreed more than they disagreed. After reconciliation, 9 of 20 specialties had a spread of 3 points or less. The areas of genuine uncertainty — Urology, Orthopedics, Oncology — are exactly where the boundary between AI-assistable and human-essential is most contested. Most health systems are not organized to capture this cognitive value. The technology is here. The workflow redesign is not. Full methodology and dataset available on request. This is the first post in a series.

The first version of this analysis was wrong. Not wrong in direction. Wrong in calibration. And the way we found that is the most important part. When three frontier AI models independently scored AI capability across 20 medical specialties, the initial disagreements were enormous. Oncology had a 35.5-point spread. Radiology 16.6. Emergency Medicine 17.6. Gemini scored systematically high. GPT scored systematically low. The raw outputs were not publishable. So we made them argue. Each outlier model received the other two models' scores and rationales, then had to defend or revise with explicit justification. Three rounds. Gemini admitted its original calibration was wrong. It had scored what AI could theoretically do, not what it could realistically do under regulatory, liability, and deployment constraints. It dropped Oncology from 71.5% to 48%. Cardiology from 76.5% to 58%. Endocrinology from 78.2% to 62%. GPT admitted its scoring templates were too blunt. It had compressed heterogeneous visit mixes into worst-case archetypes, anchoring on the hardest 20% of each specialty's workflow. It moved Oncology from 36% up to 48%. Emergency Medicine from 35% to 45%. They converged on Oncology at 48%. Neither model's original score was right. The reconciled score was better than either one alone. After three rounds, average inter-model spread dropped from 13.8 to 4.0 points — a 71% reduction. Nine specialties landed within 3 points across all three models. This is adversarial multi-model reconciliation. Independent estimation. Structured disagreement. Iterative convergence. Transparent audit trail. Three things this reveals that single-model prompting cannot: Calibration bias is real and measurable. Gemini's theoretical framing and GPT's deployment framing are both internally consistent but produce materially different numbers. If you rely on one model without cross-validation, you are getting that model's bias, not ground truth. Forced justification beats forced scoring. When a model has to explain why its number differs from two independent estimates, it either mounts a compelling defense or it revises. Both outcomes generate information the original score did not contain. Persistent disagreement is signal. Urology and Orthopedic Surgery had the widest spreads after three rounds. That tells you something real about the specialty — the cognitive-physical boundary is genuinely contested — not something wrong with the method. Full methodology, reconciliation prompts, and complete dataset available on request. The master prompt is available to anyone who wants to replicate with different models. This analysis cost less than a single-site clinical study and produced a testable framework that updates as models improve. Post 2 of a series. Post 1 has the full consensus ranking.