feral-analyses · four-model panel · 2026-07-20

Automation Moats

Which industries, professions, and specialties are least likely to be automated — under the lens that robots can own all future work once the installed base is rebuilt for them, but rebuilding the installed base takes ~50 years.

GPT-5.6 Sol · 20 min reasoning Grok Heavy · 75 sources Gemini · Ultra Claude Fable 5 · Max

The consensus

All four models converged, independently, on the same structural claim: the moat is not skill — it is the installed base.

A task falls to automation when its substrate is standardized and the work can be brought to a fixed station (phlebotomy: a patient, a chair, an arm). It resists when the machine must travel into bespoke, undocumented, legacy environments (a wet 1920s crawlspace). Moat duration ≈ replacement cycle of the substrate — and for the US built environment, that is 50–100 years.

Top protected headcount: home & direct care → bedside nursing → brownfield trades → childcare/K-12 custody → emergency response.
Top protected dollars: hands-on healthcare (~$5T sector), the repair/retrofit annuity (~$600B of a ~$2.2T built-environment sector), education's custody core (~$1.6T sector).

The sharpest caveat (Sol, echoed by Claude): occupations persist longer than headcounts. Remote experts and robot supervisors shrink the workforce long before the job disappears — "regulation preserves a human signature more effectively than it preserves human headcount."

The shared framework

Each model articulated the same three-factor moat in different vocabulary — plus one test that explains why phlebotomy fooled everyone.

Moat factorSolGrokGeminiClaude
Substrate is bespoke / legacy"brownfield penalty""legacy infrastructure lock-in""unstructured legacy environments""bespoke substrate"
Machine must travel into chaosfour automation routes"robots choke in 1970s crawlspaces"Moravec's paradox"machine-must-travel"
Human accountability required"legal legitimacy""licensing, safety calls""high-stakes moral judgment""liability fuse"

Sol's decisive test: a profession is only safe if all four automation routes stay hard — build a general robot, move the task into a workcell, rebuild the environment for machines, or eliminate the service. Phlebotomy fell via the workcell route, not the general-robot route — which is why it looked unassailable right up until it wasn't.

Claude's betting formula: bespoke substrate × machine-must-travel × liability fuse × demographic tailwind. "Everything with a chair, a highway, or a stainless-steel counter is on the clock."

Ranked by humans employed

Ranked by strength of the resistant core, not sector size. Ranges span the four models' figures; they chose different denominators (occupation vs sector), so both are shown.

Protected workforce — resistant core, US
Millions of workers · midpoint of the panel's figures · sorted by size
Childcare / K-12 custody core
7.9M
Occupied-space cleaning
5.7M
Bedside nursing
5.0M
Home & personal care aides
4.3M
Field social & crisis work
2.9M
Brownfield trades (core)
2.8M
Emergency response
2.4M
Personal / body services
1.2M
Veterinary & animal work
0.75M
Size ≠ rank: the consensus ranks home care #1 on moat depth (triple moat, unanimous), while the custody core is bigger but rests on social legitimacy rather than physical lock-in.
#Protected groupHeadcount (US)AgreeWhy the moat holds
1Home health & personal care aides4–4.3M core
sector to 16–18M
4/4The one unanimous #1. Bespoke homes × nonstandard bodies × wages too low to amortize a robot. Largest single US occupation, growing on demographics.
2Bedside nursing (RN/LPN/CNA)~5M3/4Charting, triage, monitoring automate; the hands and the liability fuse don't. Sol impedance 91/100.
3Brownfield trades — plumbing, electrical, HVAC, elevators, line work2.5–3M core
sector 11.7–13.6M
4/4The prompt's thesis, confirmed. Sol's highest impedance (94). Grok's detail: 757K electricians, 541K HVAC, 466K plumbers, 131K line installers, 24K elevator techs.
4Childcare, preschool, special ed, K-12 custody core~7.9M core
sector 8.4–10.6M
3/4"Instruction automates, custody doesn't. Schools are load-bearing childcare." Grok omits it — his lens was purely physical.
5Emergency response (fire, EMS, SAR, policing)1.4–3.4M4/4Every scene unrepeatable, adversarial, liability-soaked. Ordinary security guards are not protected.
6Occupied-space cleaning & remediation~5.7M3/4Occupied homes, biohazard, hoarding, disaster restoration. Standardized commercial janitorial is exposed.
7Personal / body services~1.15M core2/4Hair, massage, tattoo — intimate work on moving bodies.
8Field social & crisis work~2.85M sector2/4The involuntary, court-facing, carry-the-license slice survives; commodity talk therapy is already being eaten by AI companions.
9Veterinary & animal work0.7–0.8M1/4Patients who can't be told to sit still (Sol only).

Ranked by market spend

Sector footprints, US annual. The panel warns these are not comparable accounting measures (expenditure vs investment vs revenue) — and total spend overstates protected employment.

Sector footprint vs the resistant slice
US annual, $ trillions · panel figures (Sol's where models diverge)
Healthcare
$5.3T
Built environment
$2.2T
Education
$1.6T
Food away from home
$1.4T
Public safety
$0.34T
Facilities services
$0.28T
Personal & laundry
$0.23T
Pet industry
$0.16T
The resistant slice is smaller than each bar: drug spending doesn't protect nurses, greenfield construction doesn't protect renovation plumbers, and only ~$600B of the built-environment total is the repair/remodel annuity all four models agree is durable. Food away from home is the weakest moat on the board (Sol impedance 63) — kitchens are rebuildable workcells.

The plumbing math

Only Sol quantified the premise. The US has 148.3M housing units and issued 1.43M permits in 2025 — about 1% of stock, much of it additive rather than replacement. At 1% annual conversion:

converted after 50 years = 1 − 0.99⁵⁰ ≈ 39%

Even 2%/yr converts only 64% in 50 years. So the question's "50 years" is optimistic for full replacement: majority robot-compatible plumbing in 40–70 years is plausible, all legacy plumbing gone in 50 is very unlikely, and a premium human market for old, undocumented, historic, or illegally modified systems is potentially permanent.

The phlebotomy fact-check

The premise is only partially true — and the lesson survives anyway.

Sol audited the study behind "China's robotic phlebotomists are better than humans": 94.3% success on patients actually punctured (5,628/5,966); better fill volumes and lower reported pain in a separate 154-person comparison; but ~205 seconds per draw vs 52 manually, with excluded patients and no simultaneous first-stick human control. Superior on selected dimensions — not blanket superiority.

The deeper lesson all four models drew: fine dexterity is not a moat once the body is brought to a stationary, instrumented, high-volume workcell. Claude's autopsy: standardized substrate + work-comes-to-you + no liability fuse = robot food, regardless of how "skilled" the task feels.

Where the panel disagrees

  1. Care-work dollars vary 4×. Grok's $150–170B (home healthcare services only) vs Claude's ~$600B (all long-term care) vs Gemini folding it into $4.5T healthcare. Different denominators, all defensible.
  2. Education. Three models rank it top-4; Grok omits it entirely. The custody moat is behavioral, not infrastructural — it counts only if social legitimacy is a moat class.
  3. Food service. Only Sol includes it (8.29M workers) while scoring it his weakest moat. Kitchens are the phlebotomy pattern waiting to happen.
  4. Medicine's interior. Claude: procedural medicine's moat is "political… real but rented" (autonomous suturing already in labs). Gemini lists surgeons as least-automatable. Sol splits it: emergency and complex care durable, fixed-station tasks exposed.
  5. Feels safe, isn't (Claude's Tier 3). Long-haul trucking, radiology/pathology reads, pilots and ATC, fast-food kitchens, standardized-building janitorial — technically automatable now or soon; their moats are psychology and regulator pace. Rented, not owned.

What each model uniquely contributed

GPT-5.6 Sol

chatgpt.com · "Worked for 20m 0s"

The four-routes test, the housing-conversion math, the phlebotomy study audit, impedance scores per sector.

Thesis: the durable role is a "trusted, accountable exception-handler operating where automated systems meet legacy infrastructure, living beings, and rare high-consequence failures" — and it can persist while headcount still falls.

Claude Fable 5

claude.ai · Fable 5 Max

The three-question autopsy, moat-duration ≈ substrate replacement cycle (homes 50–100y, vehicles ~20y, bodies ∞ but transportable), the "feels safe, isn't" tier.

Thesis: bespoke substrate × machine-must-travel × liability fuse × demographic tailwind. Capability isn't adoption: certification → insurance → code approval → acceptance is 10–20 years per trade, per jurisdiction.

Grok Heavy

grok.com · 75 sources

Occupation-level BLS granularity: 757K electricians, 541K HVAC, 466K plumbers, 131K line installers, 24K elevator techs. US housing stock ~140M homes, average age 40+.

Thesis: demand tailwinds (aging population, infrastructure bills, energy transition) grow these roles faster than automation displaces them.

Gemini

gemini.google.com · Ultra

Widest sector coverage (adds grounds/janitorial and protective services with sizes) and the cleanest statement of the inversion.

Thesis: the historical assumption flipped — the data-entry clerk, junior copywriter, and entry-level software engineer now face higher displacement risk than the electrician or occupational therapist.