← Back to the savings calculator
Every default in the calculator is tied to a published study. The combined figure is our model – not a measured customer result yet.
| # | What the study found | Used for | Source |
|---|---|---|---|
| 1 | Rework consumes ~40 % of project effort; 70–85 % of that rework is caused by requirement errors. | Rework share, requirement share | Leffingwell, Calculating the Return on Investment from More Effective Requirements Management, American Programmer 10(4), 1997 · Boehm & Basili, IEEE Computer 2001 |
| 2 | A requirement error costs 21–78× more to fix in integration & test and 29–1500× more in operation than when found in the requirements phase. | Value of finding contradictions early | Stecklein et al., NASA JSC, Error Cost Escalation Through the Project Life Cycle, 2004 |
| 3 | Inadequate testing infrastructure costs the US economy $59.5 bn per year; finding and fixing defects accounts for ~80 % of development cost. | Rework share | Tassey, NIST Planning Report 02-3, 2002 |
| 4 | Verification is labour-intensive and may reach ~60 % of overall software development cost for dependable (DO-178C) systems. | V&V share (default 40 %) | Sun, Brain, Kroening et al., Functional Requirements-Based Automated Testing for Avionics, 2017 |
| 5 | Airbus replaced unit testing by formal "unit proof" on A380/A400M/A350 software; Dassault gained ~1 person-month per flight-software release by replacing robustness testing with Frama-C. | Verification effort avoided | Moy, Ledinot, Delseny, Wiels, Monate, Testing or Formal Verification: DO-178C Alternatives and Industrial Experience, IEEE Software 2013 |
| 6 | Semi-automatic generation of test models from structured requirements saved 86 % of the time compared with manual creation, at equal quality. | Verification effort avoided | Fischbach et al., Automated Generation of Test Models from Semi-Structured Requirements, 2019 |
| 7 | EARS patterns, developed at Rolls-Royce for jet-engine control, gave qualitative and quantitative improvements over conventional text: less ambiguity, vagueness and omission. MBSE case studies with cost metrics attribute success mainly to defect prevention in requirements. | Requirement defects avoided | Mavin et al., EARS, IEEE RE'09 · Carroll & Malins, Sandia SAND2016-2607 |
| 8 | Generative AI: 56 % faster task completion in a controlled coding experiment; 14–15 % more output in a field study; but −19 % for experienced developers on their own large codebases. LLMs can generate Capella architectures and SysML v2 artefacts from requirements. | AI authoring & modelling saving (default 35 %, deliberately below the coding figures) | Peng et al. 2023 · Brynjolfsson, Li & Raymond, QJE 2025 · METR 2025 · LLM-generated Capella architectures, J. Eng. Design 2025 |
| 9 | Our model: 40 % × 35 % (verification) + 30 % × 70 % × 60 % (requirement rework) + 35 % × 35 % (AI authoring/modelling) ≈ 39 % of the engineering budget; conservative case ≈ 23 %. | Headline figure 30–40 % | EASY.SAFE – see calculator above; to be validated with pilot customers. |