# Macro WAG Session — Employment, Automation Exposure, and the Whole Economy

*Full-text archive of one thread from a longer Claude Code session, 2026-09-10 to 2026-09-11. 8 user turns, 8 assistant turns. Tool calls, tool output, and model reasoning are omitted; every human message and every assistant response is reproduced verbatim.*


---


**Ted:**

ok, this is all fascinating.  i know we need a lot more data, but after this first look, what would be your WAG about how many positions could be significantly affected by AI?  with an attempt a full automation (not copilot type tool approaches).  it looks like not a lot?

---


**Claude:**

Let me ground the WAG in the actual roster rather than pull a number from the air — a crude keyword pass over all 344 classes bucketed by full-automation exposure, weighted by FTE. This isn't the real scoring (that needs the duty statements and a lot more), but it anchors the guess.

Your read is right: for a genuine full-automation lens (remove the FTE, not speed it up), it's **modest — a WAG of roughly 10–15% of the workforce as the technically-plausible ceiling, and more like 5–8% as the realistically-removable floor.** In headcount that's somewhere around **60–170 of the ~1,120 positions**, and the honest center of gravity is the low end of that.

Here's what the crude first-pass bucketing (FTE-weighted, keyword-level, *not* the real scoring) shows:

| Bucket | FTE | Share | Character |
|---|---|---|---|
| **High** full-automation candidate | ~82 | **7%** | back-office transaction processing: accounting techs, court clerks, permit techs, records/disclosure, office assistants, utility billing/service reps |
| **Partial** (workload, not the whole job) | ~214 | 19% | dispatch, planners, inspectors, analysts, paralegals, HR generalists, service desk, librarians |
| **Low** (structurally insulated) | ~754 | 67% | sworn police/fire, utility & plant operators, maintenance/fleet/parks trades, engineers, managers, elected |
| Temp / unclassified | ~71 | 6% | seasonal labor + a few odd titles |

**The ~7% "High" bucket is the real full-automation target** — it's clerical/transactional work whose core *is* moving text and numbers through rules, which is exactly what AI-plus-workflow can do end-to-end. Add the automatable slice of the "Partial" roles and you reach a technical ceiling around 10–15%. But two things pull the *realistic* number down toward 5–8%: much of the Partial work has a statutory human-in-the-loop or public-trust core that makes full FTE removal illegal or unwise (911 dispatch is the clearest — you can automate call-classification but not remove the dispatcher), and many High-bucket roles are single-incumbent, so you can't cleanly delete 0.6 of a person.

**Why "not a lot" is the correct — and interesting — answer:** two-thirds of this organization is water pipes, fire trucks, patrol cars, and parks. A city is *structurally* automation-resistant in a way a bank back-office or an insurer isn't, precisely because its work is physical, in-person, emergency, and statutorily mandated. That's itself a finding: it says the choice of *which* organization you point AI at dominates everything. A benefits-administration agency or a permitting-heavy department would score two or three times higher on the same method; a general-purpose municipality sits near the resistant end.

Three caveats so "not a lot" doesn't get over-read:

1. **You set aside copilot deliberately, which is fair — but that's where most of the value actually is here.** The ~19% Partial bucket is where assistive tools plausibly remove 20–40% of the *workload* without removing a single FTE. Full automation understates AI's footprint on this org considerably.
2. **The automatable work clusters in exactly the "seams" the original thesis predicted.** Remember Bellingham is a net regional service *seller* — ~$30M/yr of Medic One billing, dispatch cost-allocation across member agencies, grant reconciliation, PFD accounting. That cross-entity reconciliation/reporting overhead is embedded across the High and Partial finance/admin roles and is disproportionately automatable — a fractional-FTE-across-many-positions pattern, not whole-class elimination.
3. **Automatable ≠ removable**, which is the one thing the org chart can't show you. A court clerk processes filings under statutory procedure; a permit tech issues under code. The legal-removability floor sits below the technical ceiling, and pinning it down needs the statutory/procedural layer we haven't built yet.

So as a first-look WAG I'd say: **~7% of positions are strong full-automation candidates, maybe 10–15% at the optimistic technical ceiling, realistically 5–8% removable** — and the more important number for actual impact is the ~19% whose workload (not headcount) AI could meaningfully cut. The next thing that would sharpen this a lot is running the 352 duty statements we already have through a real task-level exposure score instead of my keyword guess — that alone would move this from WAG to defensible.

---


**Ted:**

I don't think I realized that most of these employees were under some bargaining unit.  those negotiated contracts are probably going to be pretty careful about controlling ai replacement.  and then some of the others are under statutory control.  Let's keep WAGging for a minute.  at approx 1000 employees were at about 1% of Bellingham's population.  Do you think this is a reasonable rule of thumb for city employment?  And thinking bigger, what about county and state?  And what about including all the other local govt. entities, schools, ports, etc.

---


**Claude:**

First, your instinct on the contracts is right and it compounds: you've got three overlapping brakes on AI replacement here, not one. The bargaining units (85% of staff) will negotiate job-security and technology-change language; civil service rules govern layoffs and bumping for the classified positions; and statute dictates that certain work be done by a designated human. Even where something is *technically* automatable, replacement runs at the speed of contract cycles and legislation, not software. That's a big part of why the realistic-removable number sits at the low end.

Now to the per-capita WAG — this is a good one, and the answer has a genuine surprise in it.

**City layer: ~1% is a reasonable rule of thumb, with big caveats.** Bellingham is actually ~1.2% (1,120 / ~93k), and it's *inflated* because it staffs regional services — Medic One county-wide, dispatch for member agencies — so some of those bodies serve 230k county residents, not 93k city ones. A full-service city typically lands **0.8–1.5%** (8–15 employees per 1,000 residents). But "full-service" is doing a lot of work in that sentence: a contract city like Spokane Valley runs ~0.1% (118 for 107k) because it buys police, courts, and fire from others. So ~1% is a fair central guess for a city that runs its own police/fire/utilities, and near-zero for one that contracts them out. The rule of thumb is really "1% × (share of services delivered in-house)."

**The bigger reveal is that the city is a minority of the government workforce standing on that same ground.** Stacking the layers a Bellingham resident actually sits inside (these are round, memory-based US-typical figures — Census of Governments / BLS — so treat them as WAGs, not citations):

| Layer | ~% of the population it serves | Note |
|---|---|---|
| City (full-service) | ~1% | Bellingham 1.2%, inflated by regional roles |
| County | ~1% | courts, jail, sheriff, health, social services, elections |
| **K-12 school district** | **~2–2.5%** | **the biggest single layer — usually out-employs the city** |
| State (incl. public universities) | ~1.5% | universities + corrections + admin dominate |
| Federal civilian | ~0.9% | |
| Special districts (port, transit, fire, water, library) | ~0.5–1% combined | many small entities |

Add it up and the entire government apparatus touching a given patch of ground is roughly **6–7% of population** (about 22–23M government workers against ~335M Americans). For the project's universe — state + local, setting federal aside — it's ~**5.5–6%**. So the city you can see so clearly is maybe **one-sixth** of the total public headcount governing that geography.

**The K-12 surprise is worth sitting with.** Your own recon already hinted at it: Bellingham School District has ~12,000 students, and at the usual ratio of roughly one employee per 8–9 students that's ~1,400–1,600 people — *larger than the city itself*, for the same town. Schools are the single largest category of public employment in the country, bigger than all municipal government combined. If you're modeling "the organization that governs this place," leaving out the school district omits the biggest employer in the frame.

Two things this does for the automation thesis:

1. **The entity-graph point becomes quantitative.** Earlier we said a resident sits inside 8–10 overlapping governments; now we can say those governments collectively employ ~6–7% of the population, and the city is a sixth of it. Modeling only the city captures maybe 15–20% of the public workforce and a *different* 15–20% than where the automatable work concentrates.

2. **The automatable fraction shifts as you climb the stack, mostly upward from the city's low base.** The city was automation-resistant because it's so physical (pipes, trucks, parks). But the layers you'd add are more exposed: county government is heavy on *eligibility determination and benefits casework* (social services) and court processing — far more transactional than city work; state government is heavy on administrative and licensing bureaucracy; even schools, though half teachers (very low exposure), carry a large clerical/administrative/food/transport tail. So a whole-geography model would likely show a *higher* blended full-automation exposure than the ~7% we WAGged for the city alone — my rough guess is the county and state administrative layers could run 15–25% on the same crude method, pulling the combined figure up toward maybe 10–15%. Schools drag it back down because teaching is nearly untouchable.

So: **~1% is a good rule of thumb for the city, but the full public workforce is ~6–7× that, schools are the hidden giant, and the automatable share probably rises as you add the county and state layers where the work turns transactional.** Which, usefully, argues that if this project wants to find the automation, the *county social-services and state administrative agencies* are the higher-yield targets than the city you started with — the city was the right place to *learn the method* because it's the most legible, not because it's where the exposure is.

---


**Ted:**

great stuff.  how about those 22M gov jobs with a different denominator, all jobs, rather than per capita

---


**Claude:**

About **14.5% — roughly 1 in 7 nonfarm payroll jobs in the US is government.** That's the ~23M against ~159M total nonfarm payroll employment. (If you measure it the other way, through the household survey's "class of worker," you get ~20M government workers out of ~161M employed, or ~12.5% — about 1 in 8. The gap is definitional, not a real disagreement; establishment-based ~1-in-7 is the cleaner figure.)

Here's the same stack from last turn, now denominated in *all jobs* instead of population:

| Level | Share of all US jobs | ~Millions |
|---|---|---|
| Local government | **~9%** | 14.6 |
| — of which K-12 + community college | ~5% | 7.8 |
| — local non-education (cities, counties, districts) | ~4% | 6.8 |
| State government | ~3.4% | 5.4 |
| — of which public universities | ~1.6% | 2.6 |
| Federal civilian | ~1.9% | 3.0 |
| **All government** | **~14.5%** | **~23** |

Two things jump out when you switch denominators:

**Local government alone is ~1 in 11 jobs**, and it dwarfs the other two levels combined — which matches the entity-graph intuition. The federal government, the layer that dominates the political conversation, is the *smallest* slice at under 2% of employment.

**Public education is the real giant: ~10.4M jobs (K-12 + higher ed, state and local together) is roughly 1 in 15 of all jobs in the country, and nearly half of all government employment.** As an employer, the single biggest thing American government does is teach. That reframes the whole exercise — if you're modeling government as an organization, you're mostly modeling schools and universities, with everything else (police, fire, roads, courts, admin) as the other half.

For scale: government (~23M) is either the largest single employment supersector in the country or in a dead heat with health care & social assistance (~22M and climbing) — bigger than retail (~15M), far bigger than manufacturing (~13M).

And here's where it loops back to the automation thesis, with a genuinely macro implication: **this ~1-in-7 chunk of employment is disproportionately concentrated in the automation-*resistant* categories** — teaching, emergency response, physical infrastructure, in-person casework, statutorily-mandated process, all wrapped in bargaining units and civil service. So the largest, most stable employment sector in the economy is also one of the hardest to automate away. If AI hits private white-collar work hard over the next decade, government's *share* of employment could passively rise even with zero growth, simply because the denominator shrinks around it. The automatable slice within that 23M is real but concentrated in the administrative/eligibility/clerical tail — county social services, state licensing and admin, back-office finance — which is maybe a fifth of the total. That's the part worth modeling precisely, and it's a small, specific target inside a very large, very sticky whole.

---


**Ted:**

ok, last denominator would be by $, say gdp or more interesting total payroll (although maybe skewed by pass thrus?)

---


**Claude:**

By dollars you actually get **three very different answers depending on what you measure — and the spread between them is the whole story**, so let me give the hierarchy rather than one number. (These are round, memory-based macro figures — BEA/BLS/OMB, ~2024–25 — against GDP ~$29T and total US employee compensation ~$13.5T.)

| Measure | Government share | What it means |
|---|---|---|
| **Spending / GDP** | **~36–38%** | money *flowing through* government |
| **Value added / GDP** | **~11–12%** | what government actually *produces* |
| Jobs (from last turn) | ~14.5% | 1 in 7 payroll jobs |
| **Compensation / all compensation** | **~18–20%** | government's share of the national wage+benefit bill |

**Your pass-through instinct is exactly the key, and it's huge.** The famous "government is ~37% of the economy" figure is spending, and roughly two-thirds of that never touches a government payroll — it's transfers (Social Security ~$1.4T, Medicare ~$0.9T, Medicaid, unemployment), interest on the debt (~$0.9T), and checks to private contractors. Strip those out and government *as a producer* — the BEA "value added" measure, which is essentially government employees' compensation plus depreciation on public capital — is only **~11–12% of GDP**. So the productive footprint is about a third of the spending headline. That gap is precisely the pass-throughs you were worried about; measuring government by outlays triple-counts its real economic size.

**Compensation is the most interesting denominator for this project, and it lands higher than the job share: ~18–20% of all US employee compensation, about $2.6–2.7T/year.** Government is 14.5% of jobs but ~19% of the payroll bill because government workers skew higher-skill (teachers, nurses, professionals, engineers) and — critically for your automation question — carry **richer benefit loads**: defined-benefit pensions, retiree health, the kind of obligations private employers shed decades ago. Remember Bellingham's ~37% benefit load on top of wages; that's a public-sector signature, and it's part of why government's compensation share outruns its headcount share.

Two things this does for the thesis:

**The pass-through skew is level-dependent, which validates why the city was so payroll-heavy.** Federal spending is mostly checks — its own employees are a small fraction of its outlays (high pass-through). Local government is the opposite: it mostly *produces services directly* — schools, police, fire, water — so its spending is mostly payroll (low pass-through). That's exactly what you saw in Bellingham: personnel was 44% of operating spend and 68% of the General Fund. So the layer you started modeling is the most labor-intensive, least pass-through layer of the whole apparatus — the one where "the budget" and "the workforce" are nearly the same object. As you climb to county and state, more of the budget becomes Medicaid and grants (pass-through), and payroll's share of the budget falls even though the automatable *administrative* share of that payroll rises.

**The dollars at stake in automation are real but concentrated.** If the removable-by-full-automation slice is ~5–8% of government FTEs, and it's the higher-paid clerical/administrative tail rather than the low-paid physical work, call it very roughly 4–6% of that $2.6T compensation base — order of **$100–150B/year** of labor cost that's technically in play nationally, before the bargaining/civil-service/statutory brakes slow the actual capture to a crawl. That's simultaneously a large absolute number and a small fraction of a $2.6T, ~19%-of-all-compensation, benefit-heavy, contract-protected whole. Which is the same shape of answer we keep landing on at every denominator: government is a big, sticky, labor-intensive sector, and the automatable part is a specific, modest, well-defended slice inside it.

---


**Ted:**

yeah, that's not a large number ($100-150B).  ok, we're up here at 40,000 feet, lets keep going.  If we try to group the whole pie of employment as you started to do, government, health, manufacturing, what is our breakdown/percentage of total employment and what percentage of each would be considered "knowledge work" (is that a useful category or just a vague one we use?). and then we have the percentage under some kind of bargaining agreement or regulatory apparatus. feel free to run with this and see what you can find

---


**Claude:**

Let me lay the whole pie out, then put your two overlays on it. One honest caveat up front: the **sector sizes are real** (BLS, ~2025), but the "knowledge-work %" and "AI-exposed %" columns are *my* estimates — there's no official statistic for them, which is itself part of your question about whether the category means anything.

| Sector | % of all jobs | Cognitive / info-core | AI-exposed (routine-cognitive) | Union + license + civil-service friction |
|---|---|---|---|---|
| Trade, transport & utilities | ~18% | ~15% | ~10% | ~15% |
| Professional & business svcs | ~14.5% | ~60% | ~30% | ~15% |
| Government | ~14.5% | ~35% | ~10% | **~55%** |
| Health care & social assist. | ~14% | ~25% | ~12% | **~55%** |
| Leisure & hospitality | ~11% | ~10% | ~6% | ~5% |
| Manufacturing | ~8% | ~25% | ~12% | ~12% |
| Financial activities | ~6% | ~75% | **~40%** | ~35% |
| Construction | ~5% | ~10% | ~5% | ~20% |
| Other services | ~4% | ~20% | ~12% | ~15% |
| Private education | ~2.5% | ~30% | ~8% | ~40% |
| Information | ~2% | ~80% | **~35%** | ~10% |
| **Weighted total** | 100% | **~30%** | **~15%** | **~30%** |

Three things fall out of this, and they sharpen everything we've been circling.

**1. Most of the economy isn't desk knowledge work — it's hands and faces.** Weight the cognitive column and only about **30% of all employment is primarily information-at-a-desk work.** The three biggest sectors — trade/transport, government, health — are dominated by moving goods, responding to emergencies, teaching, and hands-on care. The popular mental image of "the economy = people at laptops" describes maybe a third of it.

**2. "Knowledge work" is a real category but the wrong one for predicting AI impact — and this is the important methodological point.** It fails because it lumps three things AI treats completely differently:
- **Routine cognitive** — claims processing, bookkeeping, permit intake, basic drafting, first-line support. LLMs eat this. *This* is the exposed slice.
- **Non-routine analytic/creative** — strategy, research, design, litigation. AI augments; rarely replaces the person.
- **Non-routine interpersonal** — teaching, nursing, management, sales, negotiation, care. AI barely touches it, because the value *is* the human accountability and relationship.

A management consultant, a medical coder, and a kindergarten teacher are all "knowledge workers," and their automation exposure runs from "augmented" to "replaced" to "essentially immune." So the predictive cut isn't knowledge-vs-manual (Drucker); it's Autor's task framework — **routine, codifiable, symbolic work with tolerant error bars and no legally-required human accountable in the loop.** By that sharper definition, the AI-exposed share is roughly **half** the "knowledge work" share — call it **~15% of all employment** where full automation is genuinely in play. Which, notice, is almost exactly the ceiling we WAGged for Bellingham (~10–15%), arrived at from a totally different direction. The macro and the micro agree.

**3. The exposure and the friction sit in almost opposite places — and that's the punchline.** Look at where the AI-exposed column peaks: **Information (~35%), Financial activities (~40%), Professional & business services (~30%).** Those are the sectors full of routine-symbolic work — and they're relatively *small* (finance 6%, information 2%) and relatively *unconstrained* (low unionization, though finance carries heavy compliance licensing). Now look at where the friction column peaks: **government and health at ~55%, education ~40%** — and those are *huge* and simultaneously *low-exposure* (physical care, teaching, emergency response, statutory process). 

So the automation pressure and the institutional protection are **negatively correlated across the economy.** The work AI is best at is concentrated in a minority of employment that's mostly unshielded; the majority of jobs are protected by some overlapping combination of physicality, human-contact, occupational licensing (~21% of all US workers need a license — bigger than union density, ~10%), civil service, and bargaining. About **a third of the workforce has a formal institutional layer between them and reorganization**, and it clusters precisely on the low-exposure sectors.

The macro implication is the one we hit last time from the dollar side: AI's full-automation force lands hard on a **small, specific, under-defended fraction — roughly the finance/professional/information back-office, maybe 15% of jobs and concentrated in the ~30% cognitive slice** — while the large, sticky sectors (health, government, education = ~30% of all employment between them) are shielded on *both* counts. That's why the aggregate labor-displacement numbers keep coming out smaller than the hype: the economy's employment mass is mostly in the wrong place for a symbol-processing technology to reach, and where it *can* reach, licensing and bargaining slow the capture.

Two things I'd flag as the softest guesses here, if you want to pressure-test: the per-sector cognitive/exposed percentages are my judgment, not measured (O*NET task data could firm them up considerably — that's the real version of this), and "financial activities" hides a big split between exposed back-office (very high) and licensed advisory/relationship roles (protected), so its 40% is a blend of two very different halves.

Want me to turn this into a proper visualization — a marimekko-style chart where each sector's *width* is its employment share and it's split by exposed-vs-shielded, with the constraint overlay? It'd make the "exposure and friction sit in opposite places" point land visually. It'd be built on these WAG estimates, clearly labeled as such, but it's exactly the kind of 40,000-foot picture that reads better as one image than a table.

---


**Ted:**

no thanks, good idea, but i can already see it.  let me drill a couople places: financial seemed surprising.  isn't a lot of that pretty heavily regulated?  even the back-office stuff?  can you break that down a little more.  and then i need a little more color on the "professional" bucket as that seems to be a pretty large catch all

---


**Claude:**

Both drill-downs are fair, and both expose the same weakness in sector-level labels: they average over bimodal distributions. Let me pull each apart.

## Financial Activities (~9.3M, ~6%) — regulated ≠ protected

Your instinct is right that it's heavily regulated, but the *kind* of regulation matters enormously, and it cuts the opposite way from what you'd expect. The sector splits roughly:

| Sub-piece | ~Emp | Exposure | Why |
|---|---|---|---|
| Insurance carriers & support | ~2.9M | **high (~50%)** | claims processing, policy admin, underwriting support — codified rules, huge back office |
| Banking / credit intermediation | ~2.6M | high (~40%) | tellers (already being automated), loan processing, ops, reconciliation |
| Securities / investment | ~1.1M | moderate (~30%) | settlement/ops exposed; licensed advisors & analysts protected |
| Real estate + leasing | ~2.5M | low (~20%) | agents/brokers licensed *and* relationship-based |

The key distinction is **process regulation vs. worker licensing**, and they're almost opposites for automation:

- **Process/institutional regulation** — SOX, Dodd-Frank, BSA/AML, fair-lending, GLBA privacy, state insurance codes — governs *how the transaction must be done* and demands audit trails and controls. This is codified, explicit, auditable rules. That is **automation-friendly substrate, not a brake.** A regime of "every step must be documented and consistent" is exactly the condition under which you can build a compliant automated pipeline with a cleaner audit trail than error-prone humans — which is why "RegTech" exists and why regulators increasingly *prefer* systematic automated controls over manual ones. The back office being heavily regulated is a reason it's *more* automatable, not less.

- **Individual licensing + liability retention** — this is the real friction, and it sits at the *front and the sign-off*, not the processing layer. Securities brokers/advisors (Series 7/63/65), insurance agents, mortgage loan originators (NMLS/SAFE Act), CPAs, actuaries: personal credentials gating specific roles. Plus liability that forces a named accountable human at decision points — the adverse-action notice on a denied loan, the suitability determination, the claims-denial sign-off, the compliance principal. Regulation *keeps a human at the accountable decision*, while freely allowing the processing underneath to automate.

So the picture is: **the back office is both heavily regulated and highly automatable — those aren't in tension** — while the protected part is the licensed advisor, the relationship, and the one accountable signatory the rules require. My earlier blended "~40%" holds, but it's really ~50–55% on the unlicensed processing half and ~15–20% on the licensed/relational half. If anything I was *understating* back-office exposure by lumping it with the protected front. Insurance in particular is the softest underbelly of the whole sector — it's mostly a giant claims-and-underwriting processing operation wearing a regulated coat.

## Professional & Business Services (~23M, ~14.5%) — two economies stapled together

This is the worst offender in the whole table — BLS files it as one supersector because it's all "services sold to businesses," but it bundles the highest-cognitive work in the economy with some of the lowest. It's really three things:

| Sub-sector | ~Emp | What's in it | Cognitive | Exposed |
|---|---|---|---|---|
| Professional, scientific & technical | ~10.5M | law firms, accounting/tax, architecture & engineering, software/IT consulting, management consulting, advertising, R&D, design | ~80% | ~30–35% |
| Management of companies | ~2.5M | corporate HQ, holding cos — execs, corp finance, HR, strategy | ~70% | ~20% |
| **Administrative, support & waste** | ~9.5M | **temp/staffing agencies (~3M), janitorial (~2M), landscaping (~1.4M), security guards (~0.7M), call centers, office admin, waste hauling** | ~25% | ~15% |

That third row is the surprise that drags the sector's real profile down: **roughly 40% of "Professional & Business Services" is low-wage physical or basic-service labor** — janitors, landscapers, security guards, temp manual workers, waste haulers — with essentially zero AI exposure. When people cite "professional services" as the AI-vulnerable frontier, they're picturing row one and unconsciously excluding rows that are 40% of the bucket.

And even row one is bimodal:
- **Exposed:** the junior/routine layer — paralegals, bookkeepers and routine tax prep, junior analysts, first-draft legal document review, routine architectural drafting, tier-1 IT. This is genuinely in the crosshairs and it's a big chunk of entry-level professional employment.
- **Protected:** the senior advisory/judgment/relationship layer — partners, principal engineers (PE-licensed), architects (licensed), senior consultants, rainmakers. Protected by licensing (bar, CPA, PE), by liability, and by the client relationship being the actual product. Plus the well-known "the AI does the junior work" dynamic, which hollows the *bottom* of the pyramid, not the top.

So the honest restatement: the exposed routine-cognitive work in this giant bucket is concentrated in **the junior tier of professional/technical firms + call centers + corporate back-office** — maybe 5–7M of the 23M — while ~9.5M is physical services and ~4M is senior/licensed/relational work that AI augments rather than replaces.

**The meta-point both drill-downs make:** the BLS sector label is an accounting convenience, not an automation-relevant category — same lesson as the city budget. "Financial activities" hides a codified-and-automatable back office behind a licensed-and-protected front; "professional services" hides a janitorial workforce behind a law-firm nameplate. You cannot reason about AI exposure from the sector name; you have to go down to the role and the task, which is exactly why the whole project is built on duty statements rather than job titles. The reassuring corollary for the "not a large number" thread: when you actually decompose these two big buckets, the genuinely exposed cores are smaller and more specific than the headline sector sizes imply — junior professional-services roles, insurance/bank back-offices, corporate admin — a real but bounded target sitting inside 20% of employment, most of it unshielded by licensing but also, notably, mostly *not* unionized, which is where it differs sharply from the government and health sectors we started with.

---


**Ted:**

Holy Bad Schemas Batman!  That's some vestigial stuff.  They have to pull that Admin category out of there.    Ok, after this whole look, you want to summarize what you found here.  It seems like a pretty limited total number of possible automation relevant jobs.  obviously software, insurance, back office...

---


**Claude:**

Your read is right, and the striking thing is how the *same answer kept falling out no matter which denominator we used*. Here's the whole arc distilled.

**The number is bounded and it converges around ~15% as a ceiling — realistically less.** We came at it four ways and they agreed:
- Bottom-up from Bellingham's actual roster: ~7% strong full-automation candidates, ~10–15% technical ceiling, ~5–8% realistically removable.
- Top-down from the whole employment pie: ~30% of jobs are desk/cognitive at all, of which maybe half is *routine* enough for AI to take end-to-end — again ~15% of employment.
- By dollars: the removable labor cost nets to ~$100–150B/yr — a big-sounding number that's ~1% of the $13.5T national payroll.

The micro and the macro landing on the same ~10–15% from opposite directions is the most reassuring thing in the whole exercise. It's not a large pool.

**Where it concentrates — your instinct nailed it, with two additions.** The genuinely full-automatable work is routine, codifiable, symbolic work with tolerant error bars and *no legally-required accountable human*: insurance claims and underwriting back-offices, banking/financial operations, **corporate administrative and finance functions**, the **junior tier of professional services** (bookkeeping, routine tax and legal document work, junior analysts), call centers, the routine tiers of software, and a thin clerical/eligibility layer inside government and health. Software, insurance, back office — yes — plus corporate admin and the bottom rung of the professional-services pyramid. Call it 5–7M jobs, sitting inside maybe 20% of employment, most of it *not* unionized.

**Why it stays bounded — three walls, and every big sector hides behind at least one:**
1. **The employment mass is physical or interpersonal.** Two-thirds of it is trades, transport, construction, manufacturing production, hands-on care, teaching, emergency response, enforcement, management, sales. A symbol-processing technology simply doesn't reach it.
2. **Institutional friction** — unions (~10%), occupational licensing (~21%, the bigger one), civil service, and liability-driven human-in-the-loop requirements — covers roughly a third of workers.
3. **The accountable-human requirement** keeps a person at the decision/sign-off even where the processing automates (the loan denial, the claims adjudication, the permit issuance, the compliance principal).

**The deep structural finding is that exposure and protection sit in opposite places.** The work AI is best at concentrates in *small, lightly-shielded* sectors (finance and professional back-office, information). The *large* sectors — government, health, education, ~30% of all employment between them — are low-exposure *and* high-friction, shielded twice. Automation pressure and institutional defense are negatively correlated across the economy. That's why the aggregate displacement numbers keep coming out smaller than the hype: the reachable work is a minority, and it's the specific minority that also happens to be transaction-heavy, benefit-rich, and — except in finance's licensed front — not especially protected.

**And the methodological punchline, which is really what this project is about:** you cannot see any of this from the schemas. The BLS "Professional & Business Services" bucket stapling janitors to law partners, "Financial Activities" hiding an automatable back-office behind a licensed front, a city budget that renders its own position tables as scrambled character-runs — the categories we inherit are vestigial accounting conveniences that actively *mislead* on automation. The only way to get a real answer is to go down to the task and the duty statement, which is exactly why Bellingham had to be built role-by-role. The city was the right place to *learn the method* precisely because it's the most legible org in existence — even if, ironically, it turned out to be near the *least* exposed, since it's mostly water pipes and fire trucks.

Three honest caveats so the "limited" conclusion doesn't get over-read:
- **Full automation was the deliberate lens; the copilot footprint is far larger.** Assistive AI touches the whole ~30% cognitive slice, not just the ~15% replaceable core — it raises output-per-worker rather than cutting headcount, and that's where near-term value actually accrues.
- **These are WAGs stacked on WAGs.** The soft layer is the per-sector exposure estimates; running real O*NET task data (or Bellingham's 352 duty statements) through a proper scorer would firm the whole thing considerably.
- **Two things could move the frontier:** robotics is the wildcard that could eventually reach the physical two-thirds AI-text can't, on a different and slower timeline; and the "accountable-human" wall is partly a *policy* variable, not a fixed technical one — law could move it either direction.

Net: **the full-automation target is real but modest and specific — roughly one in seven jobs at the outer edge, concentrated in back-office finance, insurance, corporate admin, junior professional services, and routine software, mostly outside the big protected sectors — and it's invisible from the top-line categories, which is the entire case for building it up from the tasks.** For the project itself, the implication is that if you ever want to point this method at where the *exposure* is rather than where the *legibility* is, the high-yield targets are transaction-heavy administrative bodies — benefits and eligibility agencies, insurance operations, financial back-offices — not general-purpose cities.

Want me to drop this synthesis into the project as a short `macro-context.md` memo? It's the strategic framing for the whole effort and worth having next to the Bellingham reports.

---


**Ted:**

sure

---


**Claude:**

Saved to [macro-context.md](macro-context.md) at the project root, alongside the Bellingham reports. I also folded a pointer into the project memory so a future session picks it up as strategic framing.

Done. The memo captures the whole arc — the four-denominator convergence on a ~10–15% ceiling, where the exposure concentrates, the three walls that bound it, the exposure⊥friction structural finding, the two schema drill-downs (finance's "regulated ≠ protected," professional services' bimodal grab-bag), the caveats, and the strategic takeaway that cities are the most legible but least exposed orgs. It's clearly marked as WAG-level synthesis to distinguish it from the data-grounded phase reports.

The project folder now holds three tiers of document: the **data-grounded** Bellingham phase reports (built from parsed source files), the **recon** for candidate selection, and this **strategic framing** at the top. Whenever you're ready to move from validating the method to hunting the exposure, the memo's closing point is the compass — point it at benefits/eligibility agencies, insurance operations, and financial back-offices rather than another general-purpose city.

---