How We Calculate How Long Until AI Replaces You (And Why You Can Audit Every Number)

1 · Why deterministic

A generative model can produce different month values for the same input on different runs. This is fatal for three reasons: auditability (the user cannot reconstruct why the number is what it is), reproducibility (we cannot tell whether a change between two readings is a real shift or model noise), and calibration (we cannot A/B the weighting without prompt surgery).

The Hedj architecture splits the work. The LLM grades five axis scores 0–100 and selects defense inputs from the user's quiz answers. A pure function — written once, version-tagged — converts those into months. The LLM is doing judgment; the function is doing arithmetic.

2 · The weights

The five axes — automation, half-life, wage pressure, AI investment, role demand — are not weighted equally. The weights reflect the empirical literature on task-level displacement: capability (what the model can do) and obsolescence (how fast skills decay) are upstream drivers; the others are downstream / lagging signals[1].

3 · The formula

risk     = 0.50 × automation
         + 0.25 × half_life
         + 0.15 × wage
         + 0.10 × ai_investment
         + 0.10 × (100 − role_demand)        // demand is inverted

defense  = ai_fluency_bonus
         + tenure_bonus(level)              // see "Tenure Trap" article

months   = clamp(
             18 − (risk − 40) × 0.45
                + defense × 0.6,
             2, 36
           )

The constants — baseline 18, risk slope −0.45, defense multiplier 0.6, clamp [2, 36] — are not arbitrary. The baseline was chosen so that a profile with risk = 40 (the midpoint of the 0–100 range) and zero defense lands at 18 months: a structurally neutral reading. The slope was calibrated so that the high-risk end (risk = 75) lands near 3 months pre-defense, and the low-risk end (risk = 25) lands near 25 months pre-defense — matching the empirical AEI exposure tails.

The clamp at [2, 36] is methodological honesty. Below 2 is alarmist theatre; above 36 the model curve moves faster than the horizon supports, and a confident 4-year-out reading would be pretending. Reading the manuscript carefully, no occupational forecaster in the literature publishes confident point estimates beyond 36 months for capability-driven shifts [2].

4 · The full pipeline

The user-facing months number is the final cell in a five-step pipeline. Three of those steps depend on the LLM (data extraction, question generation, axis grading); the rest are deterministic lookup and arithmetic.

5 · Sensitivity range

Alongside the point estimate, every reading publishes a sensitivity range computed by perturbing each axis by ±10 (clamped to [0, 100]) and re-running the formula. The reported range is the min and max across all four perturbations and the unperturbed center.

This isn't a true statistical confidence interval — it is a robustness band. It tells the reader: "if any one of your axis grades is off by 10 points in either direction, your reading would be somewhere in this range." The convention mirrors how macroeconomic forecasts publish bands around point forecasts of unemployment or GDP [3].

6 · Limitations

6.1 — Axis scoring depends on the LLM. The arithmetic is deterministic, but the inputs to the arithmetic are LLM judgments. The AEI floor on automation limits the damage on that axis. The other four axes (half-life, wage, investment, demand) remain LLM-graded with looser anchors. Future revisions should progressively anchor the remaining axes to public datasets — wage to BLS OEWS, demand to BLS Occupational Employment Projections, investment to public company AI capex disclosures.

6.2 — The formula is not the science. The choice of weights and constants is informed by literature but contains design decisions. We publish them transparently so they can be challenged, not because they are uniquely correct.

6.3 — Months units are aggregate, not personal. The output represents a population-conditional risk projection given your role + answers, not a guarantee about your personal trajectory. A reading of 12 months does not mean you have 12 months; it means a structurally similar profile, in aggregate, is at notable risk on that horizon.

References

1. [1]
   Goldman Sachs · How Will AI Affect the Global Workforce

   Cross-industry task-level displacement estimates that inform why automation + half-life carry the largest weights.
2. [2]
   OECD · Employment Outlook 2023, AI chapter

   Discussion of forecasting horizons in occupational displacement; rationale for short-horizon clamp.
3. [3]
   Federal Reserve · FOMC Summary of Economic Projections

   Macroeconomic point-estimate-plus-range convention. We mirror this format in the dashboard sensitivity strip.