Vibe coding reaches enterprise SaaS: an SMBC-group lessor uses it for per-customer customisation (エンタープライズSaaSにもバイブコーディングの波 三井住友系が顧客ごとのカスタマイズに活用)
B · Useful, limited ITmedia AI+ report from the launch press event of Sumitomo Mitsui Finance and Leasing (SMFL) and Replit; first-hand quotes on the control design, but the results are company-reported and partly forecast.
Summary based on the full text·Original: about 5 minutes to read
In short
ITmedia AI+ (6 November 2026) covers SMFL, a leasing company jointly owned by SMBC Group and Sumitomo Corporation, which on 4 November announced "enterprise vibe coding" built on Replit and a "personalised SaaS" for corporate clients. Clients can have its asset-management service assetforce customised through natural-language app building. SMFL put a gateway between Replit and assetforce, split context per task to save tokens, and checks generated code with a quality-control agent running deterministic tests, with an SMFL engineer approving every change.
Key numbers
- 6 people x ~6 months to 1 person x 1 weekassetforce modification estimated under the old method versus done with enterprise vibe coding
- 1/3expected development and test time after requirements for a 200+-screen depreciation system, versus the old method
- 200+companies with paid contracts for SMFL's asset-management cloud service assetforce
- 10x in 5 yearsSMFL's revenue target for its digital services
As reported by the source.
Japanese·Web page·opens itmedia.co.jp
What it reports
- SMFL holds about 9 trillion yen of operating assets. It gained an in-house development team when it bought GE Japan's leasing arm in 2016; assetforce went on sale in 2021 and has paid contracts with more than 200 companies.
- Customising assetforce for each client's requests was costly; Replit was chosen to cut that effort.
- A development platform (ADP) connects assetforce and Replit, checks authentication and permissions on both sides and passes only data and operations the client has allowed.
- Context is split by work unit so each developer's AI session receives only the design information needed, to stop token use ballooning in multi-person projects.
- A quality-control agent independently verifies AI-written code with deterministic tests; an SMFL engineer always reviews before release, and final decisions are not delegated to AI.
- Claimed results: an assetforce change estimated at 6 people for about half a year was done by 1 person in 1 week; a 200+-screen depreciation system is expected to need one third of the usual development and test time after requirements.
- Replit CEO Amjad Masad argues vibe coding reduces dependence on individual developers because AI writes in a more standard style. SMFL concedes a risk of over-personalising systems and is training forward deployed engineers.
How they got it
A reporter's account of the press event, with quotes from SMFL president Tetsuro Imaeda and Replit's CEO and slides from SMFL. No measurement method is given for the time savings.
Limits
- All results are self-reported at a product launch; the depreciation-system figure is a forecast.
- The 6-person, half-year baseline is an estimate, not a measured project.
- No data on defects, running cost or maintenance after release.
Why it matters for companies
A regulated financial-group company is putting natural-language app building in front of its own clients, with explicit controls: a permission gateway, deterministic tests by a separate agent, and a human sign-off. That control pattern is reusable for any firm letting non-developers build on top of core systems. SMFL's own point that requirements matter more once coding is cheap fits the wider evidence. Treat the 6-people-to-1 figure as a launch claim.
Key passages in translation
Japanese original
金融や会計を扱うシステムでは、計算結果が正しいこと、その計算根拠を説明できることが欠かせない
In systems that handle finance and accounting, it is essential that calculation results are correct and that the basis for the calculation can be explained.
Japanese original
assetforceの改修では、従来手法では6人で約半年かかると見込んでいた開発を、1人・1週間で完了した。
In the assetforce modification, development that was expected to take six people about half a year with conventional methods was completed by one person in one week.
Japanese original
最後は必ず同社のエンジニアが確認してからassetforceに実装するものとし、AIに最終判断を委ねないことを開発のポリシーとしているという。
The final step is always a check by one of the company's engineers before release to assetforce; not leaving the final decision to AI is its development policy.
Summary based on the full text. Read the original for the full argument and data.Read the original on itmedia.co.jp