THE WORK ECONOMY, TASK BY TASK MODEL V2
A thousand tasks.
One fixed fleet.
Reconcile financial records. Review a contract. Resolve a support case.
Allocate the hours. Price the friction. Find what the fleet needs.
SIMULATION HORIZONSCENARIO · NOT A FORECAST
Oct 2027
202620282030
A fixed physical fleet. Serving efficiency keeps moving.
A REAL TASK REGISTER
Make the work specific.
1,000 distinct occupational task statements, organized into 20 functions. Each has an execution unit, an allocated hours budget, and its own editable economics. Select a row to inspect the work.
1,000distinct source tasks
20functional domains
130source occupations
—economic starting profiles
| Task / source occupation | Work-hours / yr | Units / yr | Acceptance Leader / laggard | Tokens / attempt | Gate |
|---|
Allocate the work & deploy a function20 independent function settings +
Hours shares are normalized from the weights below. Turning a task or function off leaves its allocated hours unused; it does not make other tasks larger. Integration is charged once per active function.
Task definitions: O*NET® 31.0, August 2026. Curation gives equal coverage to each function; this is not a global economic ranking. Source update dates vary by task.
THE DRIVER TREE
Drag the sliders to rerun the modelWork has to clear every gate.
FINITE OPPORTUNITYDEPLOYMENT & ACCEPTANCEPAID CAPTURE
01 / WORK UNIVERSE
—modeled execution units / month
Finite and explicitly assumed LEADERS
—of eligible work
Share of work, not firms LAGGARDS
—remaining work opportunity
Readiness and ROI still bind02 / ATTEMPTED WORK
—execution units attempted / month
03 / PAID TOKENS
—raw paid tokens / month
04 / VALUE CAPTURE
—net provider receipts / month
FIXED FLEETBusy is a physical measure.
Busy is a physical measure.
Paid is an economic one.
—paid inference utilization
Can work outrun efficiency?
PAID UTILIZATIONDemand for physical inference capacity, after serving efficiency.
Does the bridge last?
$B / MONTH↳
WHAT THIS SCENARIO SAYS
THE WORK MAP
Different work. Different friction.
Drag a task into a quadrant, or select it and use the classification controls. “Open world” is the taste, uncertainty, and real-world edge-case side of your “unsolvable” axis.
Counts include every matching task. Up to six representative tasks are shown in each quadrant; select one to inspect its assumptions.
SINGLE PLAYER PEOPLE, SYSTEMS & APPROVALS → MULTIPLAYER
01
Bounded & verifiable
One operator. An effective check.
1.0× coordination · 1.0× review02
Coordinated execution
Clear rules. Many dependencies.
1.5× coordination · 1.0× review03
Judgment & taste
One operator. A moving target.
1.0× coordination · 1.5× review04
The open world
Many actors. Uncertain outcomes.
1.5× coordination · 1.5× reviewTOP: BOUNDED / VERIFIABLE ↓ BOTTOM: JUDGMENT / OPEN WORLD
THE CUSTOMER'S ECONOMICS
ILLUSTRATIVE COMPANYA task is not a P&L.
Turn accepted work into actual cash savings. Review, rework, software, integration, and residual risk stay in the bill. A released hour only reduces spending if the organization can realize the saving.
$bn
THE LEGAL SPEND TEST
Cut the legal bill.
Legal budget derives from allocated employee hours × loaded cost. ACC’s 0.43% of revenue is a comparison anchor, not the model’s identity.
THE SALES & MARGIN TEST
Make the P&L move.
Savings improve profit dollar for dollar. Equivalent sales use the incremental contribution margin you set; they are not new sales forecasts.
REVERSE THE MODEL
ONE LEVER AT A TIMEWhat would you need to believe?
Hold the other settings fixed. Solve for the work, routing, or economics required to fill inference capacity and cover the cash bill. Infeasible targets stay infeasible.
THE FINITE-WORK CEILING
If every eligible task clears its gates…
Full adoption and perfect acceptance are an upper-bound experiment. Customer economics and finite opportunities still matter.
The identities behind the result
ACCEPTED WORK
Finite tasks × rollout × readiness × acceptance × economic gatePAID COMPUTE
Attempts × tokens per attempt × provider routing ÷ serving throughputCUSTOMER CASH SAVING
Realized avoided work − review − rework − AI − integration − expected residual lossPROVIDER CASH
Served tokens × realized price × channel retention − fixed costsKEEP THE EVIDENCE SEPARATE
AS OF 29 OCT 2026Measured facts. Editable beliefs.
Source statements and occupational ratings are measured source data. Execution units, work allocation, AI acceptance, adoption, prices and fleet finances are explicit scenario assumptions. They are not reported OpenAI or Anthropic results.
How the 1,000-task model works
One conserved work budget. Modeled workers × annual hours × catalogue coverage sets the finite hours pool. Function weights divide it once; task weights divide each function once. Opportunity units equal allocated hours × 60 ÷ minutes per unit. Disabling a task leaves its hours unused. Only explicit weight changes rebalance other shares.
The unit is part of the assumption. O*NET describes occupational duties, sometimes several activities at once. The model adds a scoped execution unit; ongoing duties use a 60-minute work block. That is not an observed completed job. Compare aggregate baseline work-hours, and inspect execution-unit counts within each task. Occupational importance and frequency are not global economic weights or annual volumes.
Profiles are editable starting points. Task semantics choose an economic profile for duration, tokens, acceptance, review, rework, coordination, cash realization and residual risk. These are analyst assumptions, not measured task-level AI results. Individual overrides take priority. Department controls govern deployment, readiness, integration and bulk changes.
Two cohorts, distinct gates. Leaders and laggards have different readiness, rollout and audited acceptance. Production tasks must clear their acceptance floor. Assisted tasks retain review and fallback costs. Buyer economics can block rollout even after quality clears. “Fund uneconomic pilots” overrides that cash gate, while quality and release timing still bind.
Demand is not guaranteed delivery. Accepted work-hours and company savings are potential demand-side outcomes before selected-fleet rationing. Provider token sales and receipts are capped at physical inference capacity. If demand exceeds that capacity, the customer outcomes require rationing, substitution or more supply that is not assumed here.
Company size also affects deployment. Global demand uses the representative company’s unit and function economics. Changing headcount or loaded cost can change world deployment by making fixed implementation affordable; these controls do more than resize a displayed P&L.
Company savings start with hours. Company employees × annual hours × catalogue coverage × loaded hourly cost defines the modeled cost pool. The same function and task shares allocate it. Review, rework and coordination consume human time. Cash realization converts net released work into avoidable spending. Model usage, application fees, integration and escaped-error losses are then deducted. Fixed integration costs occur once per active function. A diagnostic flags a modeled labor pool larger than company operating spending.
Failures still use compute. Base tokens × attempts per work unit × verification overhead gives tokens per attempted unit, before efficiency changes. Acceptance is measured after that retry budget. More retries never improve acceptance automatically: both assumptions must be edited if you believe they do.
The bridge is calibrated, not observed. Starting gross receipts of $5 billion per lab per month, $5 per million tokens and 30 paid compute units are stress-test assumptions. Bridge mode combines the frozen starting legacy residual with the current task catalogue. Catalogue-only mode removes that residual. A diagnostic exposes an incompatible catalogue and starting anchor; the model does not silently shrink tasks to fit.
A fixed fleet can get more efficient. Training, inference, reserve and free use are distinct physical allocations. More throughput and fewer tokens per task reduce the compute needed for the same work. Two identical fleets share one work universe. Free fill creates occupancy, not metered revenue. Price and routing are explicit assumptions, not a GPU-price or subscription-elasticity forecast.
Timing still matters. Physical work is gated by release month. Long-run robotics demand does not pay an earlier cash bill. Multimodal and physical tasks use editable token-equivalent serving demand; local edge control and training are not inferred from text tokens. No new financing is assumed in the cash bridge.
Coverage has limits. These are 1,000 curated tasks across 20 functions and 130 occupations, drawn from a U.S. occupational database. Equal task counts per function are a curation choice, not equal economic shares. Consumer entertainment, autonomous agents creating new demand, and nonlabor scientific compute need separate calibrated universes. Do not treat the catalogue as a measured global TAM.