Methodology · Positioning
A different question.
Nearly every AI-risk tool asks the same thing: can AI do your tasks? This model asks what happens to your income when AI reaches the people who pay you. Same risk, opposite side of the transaction.
Two sides of the same risk
The canonical AI-exposure literature (Frey & Osborne's automation probabilities, the Felten-Raj-Seamans AI Occupational Exposure index, Eloundou et al.'s "GPTs are GPTs", Goldman Sachs' labour-hours estimates) is supply-side: it decomposes an occupation into tasks and scores how much of that task content current AI can perform. That work is rigorous and essential, and this simulator consumes it directly: those scores drive the direct-displacement layer of every run.
What the task-based frame cannot see is the demand side. An occupation's income doesn't only depend on whether its work can be automated. It also depends on whether its customers can still pay. AI-Cascade models that second-order channel explicitly: each occupation's revenue is decomposed into customer industries, each customer industry's spending capacity contracts with its own AI exposure, and the effects propagate month by month.
| Task-based exposure frameworks | AI-Cascade | |
|---|---|---|
| Core question | Can AI do this occupation's tasks? | What happens to this occupation's income when AI reaches its customers? |
| Direction of risk | Supply side: the work itself | Demand side: the people who pay for the work |
| Unit of analysis | Tasks within one occupation | Revenue flows between an occupation and its customer industries |
| Typical output | A static exposure score | A month-by-month trajectory: headcount, revenue, income per remaining worker |
| Blind spot | An occupation can score "safe" while its customer base contracts | Customer-industry shares are judgement calls, documented and adjustable |
These are complements, not competitors. The simulator sits on top of the task-based literature: Frey-Osborne baselines, a documented modern overlay, AIOE percentiles, and observed usage from the Anthropic Economic Index all feed the direct layer. It then adds the demand-side cascade those frameworks don't attempt.
The immune-worker paradox
The practical consequence of a supply-only lens: a worker whose tasks AI can barely touch still reads as safe, even when their market is eroding. A real-estate agent negotiates face-to-face, reads local emotion, and walks physical property, so task-based models score them as resilient. But if the software developers, accountants, and financial analysts who buy and rent through them lose income in aggregate, the agent's transaction volume falls anyway. In this model that pressure arrives through two channels: the customer-industry pipes, and the Moretti local-multiplier drag that propagates tradeable-sector losses onto local services with a lag.
That is the case the simulator exists to make visible: zero task exposure is not zero income exposure.
Why isn't everyone modelling this?
It isn't laziness. It comes down to data and incentives.
- The data funnels research supply-side. O*NET gives every researcher a free, granular map of what each occupation does. No equivalent public dataset maps who each occupation sells to. Models follow the data that exists, so the literature looks inward at tasks.
- Demand cascades are genuinely hard to measure. Networked flows have feedback loops, and small parameter changes can move outcomes a lot. Institutions avoid publishing models that are hard to defend point-by-point. Our answer is not false precision. It's a deliberately simple, fully published mechanism (a few logistic curves and weighted sums), explicit uncertainty bands, and an audit trail separating derived values from judgement calls.
- The mainstream frame is productivity. Most institutional AI economics measures aggregate gains: how much more efficient firms become. A demand-side lens surfaces the less comfortable question of who loses purchasing power along the way, and that question has fewer natural sponsors.
What we don't claim
- Not derived from input-output tables yet. The customer-industry shares are expert judgement, bounded by trade data. Deriving them from BEA / ONS / ABS / StatCan input-output tables is the top item on the roadmap, and every current value is inspectable in the coefficient notes and adjustable via presets.
- Not a forecast. It's a structured way to inspect the consequences of a chosen adoption scenario. The uncertainty grows with every simulated month, and the simulator says so on screen.
- Not a career planner. It won't tell you how to reskill. It shows where the money that feeds an occupation comes from, and how a given AI scenario reshapes those flows.
Every input, formula, and assumption behind all of this is on the methodology page and the formulas page.