01
Capability cluster
- Output quality
- Does AI produce work that's usable as-is?
- Oversight need
- Minutes of human review per unit of output.
- Latency
- Fast enough for the task's real-world cadence?
Methodology · Open, versioned, monthly
How every number is built.
Don't read it — see it. Start with the output, then open how each figure is derived. Every number ships with its method and a confidence band.
See it first
One role, fully costed.
What the methodology produces, before any theory: a real-shape Wagecard — the exposure number with its confidence band, the operational AI cost, and the four-class task mix.
Financial analyst · US Tier-2 · mid-level
Augmentation territoryEconomic substitution exposure
44/ 100±6 pts
038–50 interval · high confidence100
Operational AI cost
$3,100/mo
Oversight, retries, and error cost priced in.
Market rate (p50)
$98k/yr
Median pay × city × experience band.
Task distribution
Replaceable tasks16%
AI-augmented41%
Human-led + AI-assisted28%
Human-critical15%
/m/v8 · computed live · illustrative sample
The number
Economic substitution exposure on a 0–100 scale — always shown with its ± confidence band, never a bare score.
Operational AI cost
What running the role on AI actually costs once oversight, retries, and error cost are priced in. Not the license fee.
The task mix
Every task lands in one of four substitution classes. The blend is the read — not a single verdict.
- Confidence band
- Operational AI cost
- Market rate
- Four task classes
The nine axes
What moves the number.
A weighted product, not a sum: capability gates, reliability multiplies, error cost divides, human advantage dampens.
02
Reliability and risk cluster
- Reliability
- Escalation and retry rate in production-like conditions.
- Hallucination risk
- Frequency of confidently-wrong outputs.
- Error cost
- Financial, regulatory, reputational severity when AI is wrong.
03
Operational economics cluster
- Integration overhead
- Context-prep, monitoring, prompt-iteration amortized.
- Unit economics fit
- Does the per-task AI cost hold up at the role's wage benchmark?
04
Human advantage cluster
- Human advantage dampener
- Composite from the task's irreducible value across five axes: trust, ambiguity, accountability, persuasion, context.
Go deeper
The full method, when you want it.
The classes, the axis that gates production use, the CFO read, and the outside evidence — collapsed by default, so you open only what you need.
The four substitution classes — where every task lands
High substitution
Replaceable
AI runs the task end-to-end with minimal human oversight. Substitution viability high.
Augmentation territory
AI-augmented
AI does most of the work; the human owns decisions and context.
Hybrid optimal
Human-led, AI-assisted
Human leads the task; AI accelerates tooling (drafting, search, summarization).
Human-critical work
Human-critical
AI delivers no net value (or negative value) due to trust, regulation, accountability, or relational complexity.
Reliability — the axis that actually gates production use
Capability is what most public AI-substitution framings reduce to: can the model do this task at all? It's the easier axis to measure, so it dominates the conversation. The axis that actually gates production use is reliability — does the model do the task correctly often enough that a human can stop watching every output?
Capability has run ahead of reliability across the matrix: 31 task-model cells score capability ≥ 75, but only 5 score reliability ≥ 80. That gap is where most of today's unsupervised-AI rollouts get reversed.
Every Wagecard surfaces three reliability numbers: hours-weighted average capability across your tasks, hours-weighted average reliability, and the gap. We also report the share of your role-hours sitting in the “capable but unreliable” zone — capability ≥ 75 with reliability < 80.
Capability gap
- 31
- cells capability ≥ 75
- 5
- cells reliability ≥ 80
Klarna reversed their 700-role AI deployment when CSAT dropped on complex tickets. Capability was sufficient; reliability wasn't.
How a CFO reads this — NPV, IRR, and payback
Enterprise AI deployment decisions go through three standard finance gates. We compute all three for any Wagecard with a salary input — treating AI substitution like any other capital project.
NPV
5-year
Sum of discounted annual savings minus Year 0 transition cost. Positive NPV means the deployment creates value at the given discount rate.
IRR
Internal rate of return
The annual return the project earns on its capital, compared against the firm's hurdle rate (WACC). A 35% IRR with cheap capital means 'do it now.'
Payback
Period
Years before cumulative savings recover the transition cost. A sanity check against NPV/IRR — positive NPV with a 6-year payback may still be rejected.
A full worked example — every figure with its method — lives in the Investment view on each Wagecard. Discount rate defaults to 10% (typical mid-market WACC) and is adjustable there on Pro accounts. We deliberately do not model option value, strategic redeployment value, or terminal value beyond Year 5 — the model is calibrated to be conservative for the individual and mid-market read.
External validation — the thesis, independently confirmed
Capability is not economic viability — and that the real cost of AI substitution runs to oversight, retries, error cost, and integration overhead is not our claim alone. The people deploying it say the same.
Nvidia VP of Applied Deep Learning
April 2026“For my team, the cost of compute is far beyond the costs of the employees.”
Fortune ↗MIT CSAIL
2024 study“AI automation economically viable in only 23% of vision-primary roles at current cost structures.”
Study ↗BCG
2025“Only 5% of companies are capturing AI value at scale; ~60% report no material value despite investment.”
BCG ↗Klarna + Uber
2025–2026“Klarna reversed 700-role AI deployment when CSAT dropped. Uber burned its full 2026 AI coding budget in four months.”
This is the gap Wagecore prices. Capability is rising. Economic viability is not — yet, and not uniformly. Our four-class taxonomy is calibrated to where AI is operationally cheaper today, not where it could be in 2030.
Evaluator & versioning — how the matrix is scored and refreshed
Matrix v1 (live) is scored by a single calibrated evaluator against a transparent rubric. An expanded evaluator methodology ships with v1.5 in Q3 2026. The methodology is open and versioned monthly; paid plans add depth and role-level detail.
Open methodology
The rubric and formula are public, so any number can be checked against the method that produced it.
Versioned monthly
The capability matrix refreshes on a published cadence.
Confidence band on every number
No bare scores — each figure ships with its ± interval.
Single calibrated evaluator
v1 scored against one transparent rubric; v1.5 expands it.
Run the math on your own role.
Pick a role, see the operational AI cost, the substitution mix, and where the human-advantage layer kicks in. Two minutes, anonymous preview.