On October 23, NTT announced that it has established technology
that organizes the relationships among the specifications, design documents,
code, and tests of existing systems
and converts them into a state that generative AI can easily understand and use.
The approach combines technology that detects and fills in inconsistencies
between documents and code with technology that automatically generates missing tests.
NTT will apply it to actual development and maintenance projects
and aims for practical use.
Background: Years of Modifications Cause Design Documents and Actual Behavior to Diverge
According to the announcement, about 60% of software development consists of
additional development such as modifying, maintaining, and redeveloping existing systems.
This figure is based on the “Software Development Data Analysis 2022”
(ソフトウェア開発分析データ集2022).
NTT says productivity improvements through generative AI are needed
not only in new development but also in additional development.
In systems that have been used for many years,
design documents and code can stop matching as modifications accumulate.
Developers need to investigate where to make changes
and how changes affect other functions,
and it also becomes difficult for generative AI
to accurately understand the system’s structure and specifications.
When tests are not sufficiently developed and automated,
verifying changes also takes extra effort.
How It Works: Organizing Links Between Artifacts and Unchecked Conditions
The first, “traceability analysis technology,”
organizes specifications, design documents, code, and other materials by function
and structures the relationships among them.
It also covers information on development intent, such as meeting logs,
and organizes it based on international standards and guidelines
for requirements engineering, including ISO.
It aligns differences in the amount of information and granularity across documents
to determine relationships,
then detects and fills in inconsistencies and omissions between development phases.
The second is technology for automatically generating tests.
It organizes the perspectives that should be checked
based on specifications, design, code, and ISO quality characteristics,
and visualizes which areas existing tests cover and which are missing.
It fills in gaps such as behavior under abnormal conditions,
boundary values, combinations of conditions,
and behavior when functions work together,
using test techniques such as decision tables,
and generates tests that can be run automatically.
Preliminary Experiments: 33 of 41 Embedded Errors Detected
The evaluation used “ATRS,” a sample application of the Macchinetta Framework,
with a code size of about 32,000 lines.
Of 41 contradictions and omissions intentionally embedded in the design documents,
the conventional method, RAG, which searches and references related documents,
detected 11, while the proposed method detected 33.
The detection rates were 27% and 80%, respectively.
In the evaluation of test generation,
the method detected 4 of 6 intentionally embedded bugs.
In addition to the embedded bugs, it also found 3 latent bugs
related to input validation and error messages.
NTT says that when tests were generated directly from specifications
and other documents using generative AI, none of these bugs were detected.
The figures are results of preliminary experiments using this sample application.
Next Steps: Testing from Modification to Verification in Actual Development and Maintenance
NTT will evaluate the technology on systems of different sizes and characteristics.
It will expand the documents and programming languages covered,
and verify the effectiveness of working with generative AI
through understanding the system, identifying the scope of changes,
making modifications, and generating and running tests.
The research results are scheduled to be exhibited
at “NTT R&D FORUM 2026” from November 18 to 25.
LC Signal’s Perspective: Sorting Out Discrepancies in Existing Assets Before Handing Them to AI
In these preliminary experiments, error detection results differed
between simply referencing related documents
and organizing the relationships among specifications, design, and code.
When using AI for modifications,
how documents are linked to each other and to code,
not just what the AI is given to read,
seems likely to matter.
Key Sources
-
NTT official announcement (Japanese), October 23, 2026:
“人が読む”既存ソフトウェア資産を、”AIが読む”開発資産へ
～ドキュメントとコードの不整合や不十分なテストを自動で補完し、AI-Native開発への転換を加速～
(Turning existing software assets “read by people” into development assets “read by AI”:
Automatically filling in inconsistencies between documents and code
and insufficient tests to accelerate the shift to AI-Native development)
This article was prepared with AI assistance using publicly available information
as of October 24, 2026.
Announcement dates follow the dates shown on the official pages.
Results confirmed in preliminary experiments are distinguished
from future plans toward practical use.
“LC Signal’s Perspective” reflects this publication’s interpretation
of the available information.