Agentic coding has changed what software work costs. Token consumption, agent runtime, and human supervisory effort are not interchangeable with the engineer-hours that Scrum was calibrated for. Most organizations adopting agentic delivery are still pricing and budgeting it with the estimation tools they used in the prior era, and the resulting numbers do not survive scrutiny from sales, finance, or delivery leadership.
This paper proposes a three-tier maturity model for cost-benefit analysis of agentic versus traditional software delivery. Each tier corresponds to a state of project knowledge, and each delivers usable output at its own level of certainty. Together, they form a single estimation discipline that follows a project from first contact through retrospective review.
What the paper covers
The paper is organized around three tiers, each with methodology, an illustrative formula, a worked example, and an honest discussion of limitations.
Tier 1: Retrospective analysis
Performed after an agentic project completes. The realized agentic actuals (tokens, compute, supervisory hours, allocated overhead) are compared against a counterfactual Scrum team constructed at fully loaded rates. The worked example covers a customer-facing data ingestion service delivered in three weeks agentically, against a nine-week, $260K Scrum counterfactual. The realized delta was approximately $233K.
Tier 2: Project-inception analysis
Performed at the start of a project whose scope is well defined. The agentic estimate is built from a reference class of prior Tier 1 outcomes, decomposed by component type (CRUD endpoints, integration adapters, data pipelines, UI screens). The worked example projects an order-routing dashboard at $186K agentic vs. $498K traditional, with non-overlapping eighty-percent confidence intervals.
Tier 3: Consultative engagement analysis
Performed during sales or advisory conversations, before scope has been fixed. Both pathways are expressed as distributions over plausible scope realizations, propagated through the Tier 2 machinery using Monte Carlo sampling. The worked example, an analytics platform for a mid-market manufacturer, produces a 78% probability that the agentic pathway costs less than half of the traditional pathway, plus a defensible not-to-exceed at the ninetieth percentile.
How the tiers dovetail
The tiers are not alternatives. The same analytical artifact narrows from a Tier 3 distribution at discovery, into a Tier 2 point estimate at scope-lock, and into a Tier 1 retrospective at delivery. Tier 1 outcomes are then recycled into the reference class that calibrates future Tier 2 and Tier 3 analyses. The compounding effect is the strategic asset: each completed engagement sharpens the estimation accuracy of every future engagement that resembles it.
The dovetail also supports active cost control during execution. Variance against the Tier 2 estimate becomes an early warning indicator. Breaches of the Tier 3 envelope trigger commercial renegotiation rather than internal blame. Sales, delivery, and finance argue from the same numbers.
Who this is for
The paper is written for executives, sales engineers, and delivery leaders who need a defensible answer to three questions:
- Should we use agentic delivery on this project?
- What should we charge or budget for it?
- How do we govern it once it is in flight?
It is not a productivity study. It does not argue that agentic delivery is universally faster or cheaper. It argues that the comparison is answerable, and that the answer is more useful when produced through a tiered discipline than through intuition or vendor benchmarks.
What the paper does not claim
The framework is silent on questions of quality, residual risk, regulatory posture, and strategic fit. Cost and schedule are not the whole picture, and the paper is explicit that these considerations should accompany any tiered analysis as qualitative supplements rather than being silently absorbed into the cost figures. The paper also acknowledges the cost of maintaining the reference class and the risk of drift as agentic tooling continues to improve quarter over quarter.
The download below contains the full paper: methodology, formulas, three worked examples, the lifecycle dovetail diagram, and a discussion of tradeoffs and failure modes.