From Code to Requirements: Agentic Reverse Engineering of Business Rules at Enterprise Scale
Garima Agrawal, Prasun Das, Priyanka L, Minisha N, Akshay Sarvade, Hemath Manivanan, Sravani Joshna, Prasad Kalyansundaram
cs.SE · cs.AI
Quick verdict
Seven-agent LLM system reverse-engineers BRDs from Java in ~9 min at $0.07; real deployment evidence, but the >98% saving is a modeled, asymmetric comparison.
Summary
This industrial case study from Cognizant presents a seven-agent LLM framework that reverse-engineers Business Requirements Documents (BRDs) from undocumented enterprise software, demonstrated on an anonymized Java programme of 18 services across 39 repositories. LangGraph-orchestrated agents scope the component, map user journeys, extract business rules (static analysis first, with LLMs reserved for translation and conflict resolution), run gap and risk analysis, and lock a BRD after human-supervised escalation points. Across the 14 instrumented services the pipeline discovered 280 journeys and 1,612 rules and mapped 327 tests, averaging 8.78 minutes and about $0.07 of LLM inference per service. Against a manual baseline modeled with IFPUG-derived complexity weights (34,981 hours, roughly $2.03 million), the authors report over 98% cost and 99.4% time reduction, while explicitly acknowledging the accounting is asymmetric and the figure indicative only. Quality checks map generated journeys to regression tests (traceability, not independent validation) and corroborate 115 extracted rules against production defect triage notes via an LLM-as-judge across five sampled services. Main limitations: a single Java codebase and one legacy pilot, no non-agentic baseline or ablations, unmeasured extraction precision/recall, static-analysis blind spots, no code or data release, and an unvalidated judge.
Review complete
Assessment at a glance
6
risk flags identified
● high 2
● medium 3
● low 1
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A seven-agent framework generates comprehensive, SDLC-ready Business Requirements Documents from undocumented enterprise software in an average of 8.78 minutes per service at roughly $0.07 of LLM inference cost.
- Against a manually modeled baseline built from IFPUG-derived complexity weights (34,981 hours, $2,028,869 for 39 repositories), the approach reduces cost by more than 98% and time by 99.4% - presented by the authors as an order-of-magnitude indication, not a controlled measurement.
- 115 extracted business rules across five sampled services were independently corroborated by production defect triage records via an LLM-as-judge, indicating the pipeline recovers rules genuinely present in the system.
- Static-analysis-first extraction with LLMs reserved for synthesis, plus embedded expert 'cognitive models' as agent-level capabilities, bounds token consumption and hallucination risk relative to prompt-heavy agentic designs.
Strengths
Genuine production deployment with measured traces, not a benchmark toy
The framework ran on a real anonymized enterprise Java programme (18 services, 39 repositories); Tables 2 and 4 report per-run LangSmith telemetry: input tokens 5,500-45,230, output 3,300-28,150, avg 47,694 tokens, avg 8.78 min and $0.066 Haiku cost across 14 instrumented services.
Unusually candid cost accounting in the body
Section 4.2b explicitly states the two sides 'are not accounted symmetrically' and calls the reduction 'indicative of order of magnitude rather than a controlled measurement'; Section 1 separates generation time from human review, which Section 4 reports separately (one hour per service).
Clever independent validation channel
Deterministic, auditable artifact assembly
Weaknesses
No baseline or ablation anywhere
Quality evaluation lacks a denominator and a validated judge
Baseline model is an unsourced-in-composition parameter stack
Narrow generalizability and unverifiable anecdotes
Internal reporting inconsistency
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Preliminary Review: From Code to Requirements (arXiv 2609.22719)
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
How much does this framework cost per service?
The paper reports average LLM inference of about $0.07 per service at Claude Haiku benchmark pricing (average 47,694 tokens), plus a stated one hour of human review per service and a one-time setup of roughly three months that is excluded from the recurring per-service figure.
Is the claimed 98% cost reduction independently verified?
No. It compares measured agentic cost against a manually modeled baseline built from IFPUG-derived complexity weights, ISBSG phase ratios, and a $58/hr blended rate. The authors themselves state the two sides are not accounted symmetrically and call the result an order-of-magnitude indication only.
What data was the system evaluated on?
A single anonymized enterprise Java programme (18 microservices spanning 39 repositories) plus one legacy pilot feature. There is no public benchmark, and no code, prompts, or generated artifacts are released.
How is the quality of the generated requirements checked?
Two ways: mapping generated journeys to existing regression tests (traceability only, since tests were also a pipeline input) and matching extracted rules against production defect triage notes with an LLM-as-judge - 115 rules were corroborated across five sampled services.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure 1 (pipeline diagram) cannot be analyzed from extracted text; pipeline understanding relies on Section 3 prose alone.
- The cost-model equations render as inline boxes in the extraction; parameter values were cross-checked against surrounding prose, but a minor misreading is possible.
- All enterprise identifiers are anonymized and example artifacts masked in the paper itself, so deployment claims cannot be externally verified from the text.
- Per-service defect counts and identifiers are withheld by the authors; the corroboration analysis cannot be independently checked.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.