How the verdict is calculated

Every formula behind the audit, published. Most automation calculators do not show their math, because there usually is not much. Here is all of it.

Most automation calculators don’t show their math, because there usually isn’t much. Here is all of it.

The readiness score

One number from 0 to 100, and the only thing that sets the verdict. It has two halves: the shape of the task, and the size of your numbers. Both move it.

shape = repeatability × risk factor × judgment factor
economics = 0.6 × (cost-per-unit improvement ÷ 30%) + 0.4 × (manual spend ÷ 6 × fixed cost)
score = shape × (0.55 + 0.45 × economics)

Both halves are capped at 1.0. Shape can carry a task no further than 55 points of its potential on its own — the rest has to be earned by the numbers you entered in step 2. A perfectly repeatable, zero-risk task running at a volume too small to cover its own platform fee will score in the fifties, not the nineties, and that is deliberate.

Input	Setting	Multiplier	Why
Risk	Bearable	1.00	Cheap and quick to fix, so cost can lead the decision.
Risk	Costly	0.85	Mistakes cost real time or money to undo.
Risk	Serious	0.55	Costs trust or a client. Heavily penalised.
Risk	Burns	0.00	Legal, financial or safety damage. Blocks the verdict outright.
Judgment	Low	1.00	Mostly the same decision every time.
Judgment	Medium	0.85	Regular exceptions that need a person’s read.
Judgment	High	0.60	Real professional judgment, not pattern matching.

A score of 65 or above returns Automate. 35 to 64 returns AI + human. Below 35 returns Human-led. Any Burns-tier risk returns Don’t automate yet regardless of the score.

The oversight contract

The verdict answers whether to automate. The oversight contract answers what you are accountable for once you have. It is not a second model — it is the same oversight premium the cost side already charges, written out as an instruction and then checked against itself.

sampling rate = base(risk) × judgment multiplier × (1.15 − 0.15 × repeatability)
units to review = monthly output × sampling rate
review hours = units to review × (1 ÷ rate) × review effort(judgment)

Base sampling runs 2% at bearable risk to 20% at serious risk. The judgment multiplier runs 1.0 to 1.8, because errors in judgment-heavy work are harder to see. Review effort assumes checking a unit costs 30% to 65% of the time it took to produce.

The self-check. Scenario C already budgets oversight at 20% to 60% of the retained salary. The contract compares the hours that budget buys against the hours its own sampling rate demands. When the sampling costs more than the premium, the page says the saving figure is optimistic and by roughly how much. A tool that quotes an oversight premium and then hands you a supervision schedule it cannot pay for is not being honest with you.

Sample adequacy. Seeing no faults in a sample of n only rules out error rates above roughly 3 ÷ n. So detecting a 1% error rate needs about 300 reviewed units a month. Where the sampling rate does not produce a sample that large, the page says so rather than implying a confidence the arithmetic does not support.

Liability-gated and capability-gated verdicts

Every task is marked as one or the other, and the distinction decides how long a verdict is worth trusting.

Capability-gated means the verdict rests on what models can currently do reliably. Document review, escalation triage and quote preparation sit here. These are re-scored after each frontier release and they are expected to move.

Liability-gated means the verdict rests on who is accountable when the work is wrong. Contract analysis, medical coding, claims assessment and resume screening sit here. A more capable model does not change who gets sued, so these verdicts are structural and expected to hold.

Conflating the two is the most common error in this category. A calculator that treats every refusal as permanent will be wrong within a year; one that treats every refusal as temporary was never worth reading. The framework version and last recalibration date are printed on every result, and every change is published in the changelog.

Output, and why hours cannot inflate the saving

You enter units finished per person per hour. Monthly output is people × rate × hours × 22, so raising the hours raises the cost and the units by the same factor. Cost per unit does not move, the readiness score does not move, and the saving scales against a baseline that scaled with it. There is no field in this tool you can turn up to manufacture a bigger number.

How AI changes output

One person supervising AI produces their own unaided output plus a leverage multiple of it, scaling with repeatability from 0.15× at the bottom to 1.60× at the top:

hybrid output = manual output × (staff kept ÷ staff today) × (1 + leverage)

Output therefore rises with staffing and with repeatability, and falls when either drops. Cutting people cuts capacity — this tool will not tell you that six people and one person produce the same thing.

The headline saving is like-for-like

Comparing a cheaper scenario that also makes fewer units is not a saving, it is a smaller business. The banner figure is the cost of producing exactly today’s output with AI helping, against what you pay for that output now:

staff needed to match today = people ÷ (1 + leverage)
saving = manual cost − cost at that staffing

The stepper in Scenario C lets you staff above or below that point. When you do, total cost and total output both move, so the honest comparison there is cost per unit, which is the line the column leads with.

The three cost scenarios

Scenario	Monthly cost
A — All manual	people × salary × (hours ÷ 8)
B — AI-led	supervision floor staff cost + platform fee + (output × per-unit API cost) + spot-check cost + (setup ÷ 12)
C — Hybrid	remaining staff cost × (1 + oversight premium) + platform fee + (output × per-unit API cost) + (setup ÷ 12)

Per-unit API cost is (input tokens ÷ 1,000,000 × input rate) + (output tokens ÷ 1,000,000 × output rate), using the published list rate for the model you selected.

An honest word about that line. On every task in this tool, the difference between the cheapest and the most expensive model is under one percent of the monthly figure, and usually far under. Token cost is not what decides whether automating this pays. Salary displaced, platform fee and setup are. The results page prints the exact spread so you can see it for yourself rather than take it on trust, and it is the reason this page does not ask you to treat model choice as a serious decision input. If list rates here have drifted since they were last checked, no verdict on this page changes.

The corrections most calculators skip

Correction	Value	What it’s for
Leverage multiplier	0.15 1.60	Extra output one supervised person gets from AI, as a multiple of their unaided rate. Scales with repeatability.
Oversight premium	+20% / +35% / +60%	Reviewing AI output and handling exceptions is different, better-paid work than the routine task. Scales with judgment.
Spot-check cost	2% / 6% / 12%	Even “full automation” gets checked. Scales with risk, charged per person’s output reviewed.
Output scales with staffing	—	Fewer people means fewer units, AI or not. The hybrid column can never claim that cutting the team leaves output untouched.
Economics in the score	55% / 45%	Task shape alone caps the score at 55% of its potential. The remaining 45% comes from cost-per-unit improvement and whether manual spend covers fixed cost.
Setup amortisation	÷ 12	One-time integration and testing cost, spread across the first year.
Economic gate	—	If matching today’s output with AI costs the same or more than doing it by hand, the recommendation is withdrawn no matter how good the score is.

What this model does not do

It doesn’t price redundancy, retraining, morale, or the management time a transition consumes. The leverage multiplier is a calibrated judgment, not a measured constant, and your own pilot will beat or miss it. It doesn’t know your specific vendor quotes. It assumes a task is roughly uniform, which most aren’t. It uses published list-price API rates, which you may beat at volume or exceed with retries and long context. And it is calibrated for teams of roughly one to fifty people on a single task — outside that range it will warn you and you should treat the output as directional only.

Run it against your own numbers

The formulas above are the whole model. Enter four figures and three judgements and it returns a verdict, an oversight schedule and the arithmetic behind both.

Start the audit