Methodology
How every number is priced.
One deterministic cost function prices every scenario — engineer-led, AI-heavy, hybrid, intern, contractor — over the same buckets, so comparisons are apples-to-apples by construction. No language model computes a number.
One cost function, identical for every scenario.
Every cost and duration is an O–E–P range: optimistic – expected – pessimistic. The constants below are the engine defaults; every analysis lists the assumptions it actually used, with their provenance, next to its results.
| Bucket | Formula | Applies to |
|---|---|---|
| Labor | Consumed effort hours × the resource's fully loaded hourly rate range (interns carry a 1.5× ramp factor). | Every human assignment: employees, contractors, interns. |
| AI usage (tokens) | Agent operating hours × the tokens an agent consumes per hour × the model's price per million tokens — counting cached context reads, cache writes, tool calls, and retries separately. | Every AI-agent assignment. |
| AI orchestration (labor) | 0.15 human hours per AI effort-hour at the plan's average human rate — specifying, steering, integrating, and accepting agent work, priced into engineering labor. | Every AI-agent assignment. |
| Reviewer / QA / security review | Review-chain hours × the reviewing resource's rate, floored at the AI-review policy minimum. | Every work packet's review chain — the same chain shape whether a human or an agent executed the work. |
| Mentoring | Mentorship hours per week × plan duration × the mentor's rate. | Plans that staff interns. |
| Coordination | k · n(n−1)/2 × the average human rate × weeks, with k = 0.05 and n = the number of people on the plan. | All humans on the plan. |
| Rework | Rework probability × 30% of the affected cost — with the AI-generated share repriced at the human fix rate, never at token rates. | Every work packet. |
The rules that keep comparisons honest.
Human time is never free
People are priced at fully loaded hourly rates whether or not they are already on payroll — an opportunity-cost basis. An engineer-led plan is not “free because we already pay them”: the hours it consumes are hours unavailable to everything else. The same rule prices the human hours inside AI-heavy plans.
Review is priced the same everywhere
Review chains attach per work packet, not per scenario, so every plan over the same backlog carries the same review obligations. That is also why review capacity — not agent speed — is usually the binding constraint.
Rework is human work at human rates
When AI output needs rework, the affected share is repriced at the reviewing engineer's rate. Without this rule, AI-heavy plans would show near-zero rework cost — reworking cheap tokens is arithmetically free — while human plans pay full price.
AI compute is token-metered
Retries and repeated context reads are counted in token spend. Regeneration is cheap; the human cost of unreliability is captured where it actually lands — in review and rework.
Documented, deliberate.
A model you can audit states what it leaves out.
- AI seat and subscription costs, internal-model compute, and negotiated enterprise discounts are modeled when you provide them, alongside metered tokens.
- Cached-context reads and cache writes are priced as their own token types; batch pricing is not modeled.
- The AI-orchestration fraction (0.15 human hours per AI effort-hour) is an asserted constant, reported as an assumption; it is not yet calibrated from usage data.
- Agent token consumption per operating hour is a benchmark default, reported as an assumption on every analysis; it is not yet calibrated from your usage data.
- The coordination constant k = 0.05 is an asserted, Brooks’s-law-shaped pairwise term, also reported as an assumption.
- Range arithmetic is component-wise (no correlation model), so ranges are conservative.
- Rework probability defaults to 0.15, is overridden by measured history when provided, and calibrates from logged outcomes — every analysis labels which of these applied.
See the numbers at work in the sample analysis walkthrough — a fixed, illustrative dataset with every scenario's ranges and rejection reasons.