Labor Displacement Index
A workload-level index tracking two distinct signals: the cost differential between human and AI execution, and the procurement-level evidence of actual substitution. These are not the same number.
There are two numbers on this page and they answer two different questions.
- Cost differential (structural) — what AI could displace at current pricing if every unit of work routed to AI tomorrow. A pressure gauge, not an outcome.
- Substitution rate (observable) — what procurement data shows is actually shifting. Far smaller, and intentionally so. PSC-level, FY-over-FY, government-wide.
The gap between them is the story. Don't average them. Don't compare across workloads as if the substitution number means the same thing in each context. Read the per-workload table for the real texture.
Trajectory
Cost ratio (left axis, log) vs. substitution rate (right axis, linear) since Feb 2025. Cost ratio tracks the structural pressure; substitution rate tracks observed procurement shift. The widening gap is the index's central observation.
Computed by least-squares slope on the historical series. Most input points are reconstructed from CPI snapshots with anchor-interpolated substitution rates — these numbers reflect the implied trajectory, not measurement-grade signal. The "samples_live" count grows with daily live runs.
Realized signals
Two derived numbers that take the structural cost gap and project it through the observed substitution rate. These are smaller, more conservative readings of the same underlying data — what's actually shifting versus what could shift.
(sub_rate/100) × volume × human_cost/unit.
Distinct from the structural cost-displacement number above.
Pilot Workloads
Nine federal workloads with legible SOC mappings, observable procurement records, and plausible AI substitution paths. Each row shows both signals — cost differential and estimated substitution — separately, plus the velocity (per-month change in substitution rate) derived from the historical series.
Human $/unit: BLS OEWS wages + ECEC multiplier (1.385×) × overhead (1.05×) ÷ 60 × O*NET min/unit. | AI $/unit: Compute CPI basket (classification, chat, or summarization tier). | Cost displacement ($): (human − AI) cost/unit × annual volume, if 100% routed to AI. Technical potential, not realized savings. | Sub rate: contractor spend YoY decline (60% weight) + AI procurement keyword-search YoY growth (40% weight), both from live USAspending.gov API, FY2023→FY2024. | Velocity: 90-day least-squares slope of sub rate, in pp/month. Mostly reconstructed today; live signal grows daily. | Absorption: OPM FedScope FTE delta (quantitative) + JOLTS/FPDS signals (qualitative).
Extension workloads
Same framework, different US contexts. Federal pilots use BLS + FPDS; these use the closest analogue — California state budget for state/local, BLS QCEW + private earnings disclosures for private sector. Static snapshots, not daily-refreshed; surfaced for cross-frame sanity-checking.
State/local: California chosen for budget-data transparency (CDT, DOF, LAO). Private-sector: BLS QCEW NAICS-level employment + S&P 500 productivity disclosures replace FPDS as the substitution signal. US-only by design — international comparators add jurisdictional noise that obscures the federal frame.
AI procurement by awarding agency
Where the federal AI procurement actually lives. Pulled from the same USAspending.gov keyword search that feeds the substitution-rate signal, deduped per award and rolled up by Awarding Agency. Sorted by FY2024 obligated amount.
Top 10 agencies by current-FY AI award value. — distinct agencies appear in the AI keyword search overall. The keyword set is intentionally specific (artificial intelligence, machine learning, LLM, OpenAI, Anthropic, Palantir AIP, ServiceNow AI, &c.) to suppress false positives — but it will under-count contracts that bury AI in technical descriptions or program names.
Measurement Pillars
Three data streams forced into a common unit — cost per workload. Each pillar has a distinct source, a distinct lag, and a distinct failure mode.
Core Formulas
Signal 1 — Cost Displacement (structural potential)
cost_ratio = human_cost_per_unit / ai_cost_per_unit
human_cost/unit = (annual_wage / 2080) × ECEC_multiplier × overhead / 60 × min_per_unit
ai_cost/unit = CPI_basket_cost[task_type] / 1000 (live from Compute CPI)
displacement = (human_cost_per_unit − ai_cost_per_unit) × annual_volume
This is the technical potential. It answers: if AI handled 100% of this workload, what would the cost difference be? It does not tell you how much has actually shifted.
Signal 2 — Substitution Rate (observable proxy)
sub_rate = contractor_decline_rate × 0.6 + ai_growth_rate × 0.4
contractor_decline_rate = max(0, −yoy_change_pct / 100)
ai_growth_rate = max(0, AI_award_total YoY growth, capped at 1.0)
Sources:
· USAspending.gov spending_by_category/psc/ — PSC contractor totals, FY23→FY24
· USAspending.gov spending_by_award/ — AI keyword search across description text
The contractor decline (60% weight) is from PSC-aggregated FPDS spend, year-over-year. The AI growth (40% weight) is from a keyword search on USAspending's spending_by_award endpoint — awards deduped per award_id, attributed per workload via description/recipient text matching, and rolled back into a 0–1 growth rate. Important caveat: PSC categories are government-wide and far broader than the specific workloads (R499 "Professional Support: Other" alone is $32B/yr across all of government). The substitution rate is a directional proxy, not a precise task-share figure.
Signal 3 — Absorption Classification
| Outcome | Signal Source |
|---|---|
| Eliminated | OPM separations, JOLTS layoffs, role abolished in budget docs |
| Reallocated | OPM reclassifications, position conversions, caseload shift to complex work |
| Upgraded | RFP language shifts, SOC reclassification to higher GS band |
| Frozen | Vacancy rate increase, position unfilled >6 months, no backfill in budget |
Methodology
Human Cost Computation
Step 1 — Annual wage (BLS OEWS). SOC-level mean annual wages from BLS Occupational Employment and Wage Statistics (May 2023). Five pilot SOCs range from $42,040 (Customer Service Reps) to $72,040 (Claims Adjusters).
Step 2 — Fully loaded compensation (BLS ECEC). BLS ECEC Q4 2024 shows wages are ~69.5% of total government compensation. Fully loaded hourly = (annual_wage / 2080) × 1.385 × 1.05 overhead. This is conservative — private-sector benchmarks run higher.
Step 3 — Cost per unit (O*NET). Minutes per workload unit from O*NET task-level data, cross-referenced with published program benchmarks. Cost per unit = (fully loaded hourly / 60) × minutes_per_unit.
Why These Five Workloads
Selection criteria: clear SOC anchor, legible FPDS procurement category, plausible current-generation AI substitution path, and observable annual volume from public data. These are not the largest federal workloads — they are the most legible.
- SNAP eligibility — Rule-based decision. 41M participants/yr. SOC 13-1041. High classification AI fit.
- UI claims adjudication — Statutory rule application. 21M claims/yr. SOC 13-1031. COVID surge exposed bottlenecks.
- Call center triage — First-tier routing and FAQ. 100M+ federal calls/yr. SOC 43-4051. GSA AI pilots since 2023.
- Document summarization — VA claims average 200+ pages. 750K hearings/yr. SOC 23-2093. 60-80% time reduction in pilots.
- IT help desk (Tier-1) — 40% of tickets are password resets. 18M tickets/yr. SOC 15-1232. AI ITSM widely deployed.
Known Limitations
Cost displacement is structural, not realized. The cost differential shows what AI could displace at current pricing — not what has been displaced. It is a structural argument for substitution pressure, not a measurement of actual substitution.
FPDS substitution rate is a rough proxy, not a census. PSC categories are government-wide and far broader than the specific workloads. R499 "Professional Support: Other" covers $32B/yr in spending across all federal agencies — it cannot be cleanly attributed to the 5 pilot workloads. The contractor spend decline signal is real (sourced from the live USAspending.gov API), but the inference from category-level decline to workload-specific substitution is indirect. The AI procurement component is also incomplete — isolating AI-specific awards requires keyword search on contract titles and descriptions, which is not yet implemented.
AI cost excludes integration and oversight. The AI cost per unit is inference cost only (from Compute CPI). Total cost of ownership including fine-tuning, human review, and error-correction is higher. The structural cost gap is real; the realized gap is narrower.
What we don't have. We do not have organizational-level data on how much volume is actually routing to AI vs. humans. That data does not exist in any public source. The gap between the cost differential (structural) and the substitution rate (observable) is the measurement problem this index is built to surface — not to paper over.
Data Source Citations
- BLS Occupational Employment and Wage Statistics (OEWS), May 2023 — bls.gov/oes
- BLS Employer Costs for Employee Compensation (ECEC), Q4 2024 — bls.gov/ect
- O*NET Online Task Database, v28.3 — onetonline.org
- USAspending.gov / FPDS Contract Data, FY2023-FY2024 — usaspending.gov
- OPM FedScope Employment Cube, Sep 2024 — fedscope.opm.gov
- USDA FNS SNAP Participation Data, FY2024 — fns.usda.gov
- DOL ETA Unemployment Insurance Weekly Claims — dol.gov/ui
- Occupant Compute CPI (AI cost per unit) — occupant.ee/cpi-data.html
Further reading
- After-action report — LDI, April 2026 — full methodology walk-through, what's working, what's broken, what's next.
- Deep dive: SNAP eligibility — workload-level numbers, sources, absorption signal.
- Compute your workload's LDI — drop in a wage, a minutes-per-unit, and a volume; get a cost differential at current Compute CPI.
Relationship to Compute CPI
The Compute CPI tracks AI cost deflation. The LDI uses that as the AI-side input to the displacement formula. As AI costs fall (CPI declines), the structural cost gap widens automatically. The CPI is currently — vs. the 100 baseline — meaning AI inference has fallen significantly since launch. That is why the cost ratios here are 1,000x+. That number will keep moving as the CPI moves.