What Is AI-Driven Development?

What is AI-driven development?

As far as we could confirm, no public body has set out a unified definition. GitLab’s official Japanese blog (updated July 7, 2026) describes it as a development approach that makes use of AI technologies such as generative AI and machine learning.

On its page "Software Development Using AI," IPA (the Information-technology Promotion Agency, Japan) lists these efforts: requirements definition with conversational AI, program development with generative AI (code generation, pair programming, and so on), and testing with AI. The term "AI-driven development" does not appear on that page (last updated September 11, 2026).

GitLab also writes that this is not an approach in which AI replaces everything. It describes a collaborative style of development that rests on humans reviewing the code and suggestions the AI generates and taking on the final decisions and responsibility. Number X sees it as development that uses AI coding tools.

What AI coding tools can do

Here are the tools each company describes in its official documentation. The Gemini Code Assist page carries a note that new purchases have ended. The OpenAI Codex page shows no date.

Experiment results, and what the companies say

There are experiments showing development got faster, and experiments showing it got slower.

In an experiment GitHub itself published in 2022, the group that used GitHub Copilot finished the task 55% faster than the group that did not. The subjects were 95 professional developers, and the task was a single one: writing an HTTP server in JavaScript (updated May 21, 2024).

In contrast, a randomized controlled trial that METR published on July 10, 2025 found the opposite. In a trial using early-2025 tools, 16 experienced open-source developers worked on 246 tasks, and when they used AI, completion took 19% longer. The developers had expected beforehand that they would be 24% faster.

The two experiments differ in both subjects and tasks, so the numbers cannot simply be lined up and compared.

Companies have also made statements. All of these are the companies’ own claims, not independent verification. At Alphabet’s Q3 2024 earnings call (October 29, 2024), Sundar Pichai said that more than a quarter of Google’s new code is generated by AI and then reviewed and accepted by engineers.

Microsoft’s Satya Nadella was reported (by TechCrunch) to have said at Meta’s LlamaCon on April 29, 2025 that 20 to 30% of the code in the company’s repositories is written by AI. The wording was "a fifth to three-tenths." We have not been able to confirm the original remarks.

Points to watch out for

GitHub’s official best practices recommend carefully reviewing Copilot’s suggestions. They also list checking with tests and automated tools, and making sure the code is not similar to existing public code.

GitHub’s official documentation covers matches with public code. Copilot has a mechanism that detects suggestions matching public code, and a setting lets you block matching suggestions. The chance of a match is said to be low.

On security, the OWASP Top 10 for LLM Applications (2025) lists "Improper Output Handling." It says that passing LLM output downstream without enough validation and sanitization can lead to XSS, CSRF, SSRF, privilege escalation, and remote code execution. This is a definition for LLM-based applications in general, not a description limited to code generation.

IPA’s "Top 10 Information Security Threats 2026" (published January 29, 2026) newly selected "cyber risks surrounding the use of AI" as number 3 for organizations. This is a risk of AI use in general, not limited to development.

Number X’s support, and our own AI services

Under "Development" in our AI adoption support, we plan AI-powered business tools tailored to your operations and AI products to build into your own services, and handle everything from screen design to development. Using AI-driven development with AI coding tools, we aim to shorten the time from prototype to launch.

Below are introductions to services that have AI built in. This page does not say whether they were built with AI coding tools.

Number X also develops and runs AI-powered services of its own. GenAI Tool DB runs a pipeline from data collection through validation to publication, records every change made by AI agents, and keeps them reversible. Important items are applied only after a person approves them.

Realshop Checker is a service that judges public information with an LLM (Claude). It always records the reason for each judgment, and switches to rule-based judgment when the LLM is unavailable. In LEATHER CARE DB, the LLM is limited to extracting information from manufacturers’ public information, and compatibility is decided by rules. In STAY LOG, a collection agent finds and updates hotels, stay reports, and videos.

At Cognify Lab., in which our members take part, we develop and operate 16 AI products.

Members who go into the field as FDEs are involved from organizing the problem through adoption and development. Our member Shota Kamiyama (AI Engineer / AI Designer) has worked in web design and software engineering since 2007. For about the past two years, he has been intensively researching and practicing AI system development.

If you have an in-house development team, we can also support introducing AI-driven development to the team.

Frequently asked questions

Is there a unified definition of AI-driven development?

As far as we could confirm, no. GitLab describes it as a development approach that makes use of AI technologies such as generative AI and machine learning. IPA’s page lists efforts such as requirements definition with conversational AI and testing with AI.

Does using AI make development faster?

Experiment results are mixed. In GitHub’s 2022 experiment, one task was finished 55% faster. In METR’s 2025 trial, completion took 19% longer for 16 experienced developers. The two differ in both subjects and tasks and cannot be lined up and compared.

Can it be done without engineers?

You need a setup in which engineers review the work. GitLab explains this kind of development as resting on humans reviewing the code the AI generates and taking on the final decisions and responsibility. GitHub’s official guide also recommends reviewing AI suggestions and checking them with tests. We organize the setup to fit what you consult us about.

Can you also help our development team adopt it?

Yes. If you have an in-house development team, we can also support introducing AI-driven development.

Is consultation free?

Yes, consultations are free. After hearing from you, we propose what to work on and a quote.