Many enterprises run into the same obstacle when adopting generative AI: the initial team finds it useful, but when the time comes to scale it across the company, those responsible cannot produce results compelling enough to secure budget approval—and they are unsure how to proceed. The approach taken by Japanese credit card company Credit Saison offers a useful reference. It first gave 315 employees access to ChatGPT Enterprise for approximately six weeks. Less than a month after this small-scale rollout ended, the company approved a company-wide deployment. From the start of the initiative to the company-wide launch, the entire process took about three months.
This article examines how Credit Saison designed its rollout plan, why it was able to make a decision so quickly, and what additional work enterprises must undertake before scaling up—areas not covered in the case study.
How Credit Saison Designed the First Phase of Its Rollout
Credit Saison’s core business is credit cards. It has approximately 31.16 million cardholders and about 3,700 employees in Japan. According to a case study published by OpenAI, the company ran the first phase of its rollout from late June to early August 2025. A total of 315 employees participated—fewer than 10% of its workforce in Japan. In September of the same year, the company rolled out ChatGPT Enterprise to all employees.
Participants included employees who volunteered, employees nominated by the project team, and representatives from each department. Every senior executive, including the president, also took part. As a result, the people who would ultimately approve the budget were themselves among the first users.
Before the initiative began, the company had already established its targets: a 500% return on investment (ROI) and 7.5 hours saved per employee per month. Credit Saison reported that by the end of the initiative, ROI had reached 954%, while each employee saved 14 hours per month—both figures nearly double their respective targets. In the post-initiative survey, 75% of managers and 70% of non-managerial employees said the tool had become indispensable to their work. Approximately 80% of respondents in the legal department believed their capabilities had improved, while more than 80% of the IT department reported higher job satisfaction.
Credit Saison Made Three Smart Moves That Enabled Approval Within a Month
It defined success criteria in advance. The targets were documented before the initiative began. When it ended, the company could simply compare the results against those figures to determine whether it should scale up, without debating whether the outcome qualified as a success. When no targets are set at the outset, an initiative usually ends with little more than a collection of user impressions, which rarely persuades finance teams or senior management.
Senior executives used the tool themselves. Chief Digital Officer and Chief Technology Officer Kazutoshi Ono believed that simply encouraging employees to use the tool was not enough—managers also needed to use it regularly. Executive President and Chief Operating Officer Katsumi Mizuno therefore created a custom GPT for corporate research and incorporated it into his daily work. Senior executives were also particularly interested in another custom GPT that let them rehearse shareholder meeting Q&As by voice, and requests to share the prompt soon followed. Because executives had experienced the benefits firsthand, the decision to approve a company-wide rollout did not depend solely on presentations from their teams.
It enabled open knowledge sharing. The Center of Excellence (CoE), a dedicated internal team responsible for promoting AI adoption and providing guidance, created a dedicated Slack channel where participants could openly share both successful and unsuccessful experiences. OpenAI also provided executive training and hands-on workshops for beginners. As a result, techniques discovered by individual employees quickly became practices that others could adopt directly.
The use cases in the case study also built on existing work. The retail counter team continued holding its usual daily review meetings, while a custom GPT used the recordings to summarize each employee’s performance, meeting notes, and handover items for the following day. This practice is now in use across all locations. The corporate sales team used the tool to research clients’ business challenges before meetings. Inside sales teams had AI role-play as prospective customers, allowing new hires to practice sales conversations by voice without taking up the time of experienced colleagues. These use cases did not disrupt the teams’ existing work habits, making them accessible even to employees using AI for the first time.
What Enterprises Still Need to Put in Place Before Scaling Company-Wide
The Credit Saison case study focuses primarily on how the first phase was designed and implemented. It does not explain how the company managed processes, access permissions, and costs after scaling up. These are among the most common obstacles enterprises face when moving from a small-scale rollout to company-wide adoption, and they require deliberate planning.
The first step is to turn effective practices from the initial phase into standardized processes. During a small-scale rollout, employees who know the business understand where to find a missing data field and which content must not be sent directly to clients. When a new group of employees begins using the system, outputs will vary from person to person if this tacit knowledge exists only in conversation logs. A practical approach is to begin with one recurring task whose inputs and outputs are easy to verify. Clearly define the data sources, what the AI Agent must produce, who is responsible for reviewing the output, and which situations require the process to pause for a human decision. EgentWrX can configure recurring work as tasks and connect multiple tasks into a workflow. It can also require human approval before work is handed off to the next step.
The second step is to differentiate data access by role. The first group of users is small and may have similar permissions. Once access is extended to sales, procurement, finance, and customer service, however, each function requires different data. Enterprises can begin by defining which AI Agent each role may use and which data it may access. For example, procurement staff need access to supplier quotations for price comparisons, but that does not mean the same role should be able to read every customer communication record. EgentWrX provides role configuration based on organizational hierarchy, along with audit logs. Administrators can review the operational process and verify whether records have been altered.
The third step is to manage costs at multiple levels. As the number of accounts grows, usage will increase as well. EgentWrX can set budget caps and alerts by tenant, business unit, member, or AI Agent. Enterprises must still decide in advance which unit will handle alerts and what results must be reviewed before additional budget is approved.
Use Your Own Benchmarks for Your Own Rollout
Credit Saison’s 954% figure was published by the company itself. The case study does not explain how ROI was calculated, and the result reflects the specific conditions of a single enterprise. Other enterprises can learn from its approach, but they should not treat this figure as a promise of their own potential returns.
A more reliable approach is to follow Credit Saison’s example of defining targets in advance. Before the first phase begins, establish benchmarks for the existing process, such as how much labor each task requires, how long teams wait for responses from other departments, and why cases are returned for revision. Then select several metrics directly tied to the purpose of the work—for example, whether a customer service draft can be used after human review and revision, or whether the data in a procurement price comparison is complete. At the end of the first phase, management can compare the actual results against these benchmarks and make an informed decision about whether to scale up and which units should be included first.
FAQ
How many people should participate in the first phase of the rollout?
There is no fixed percentage. Credit Saison included fewer than 10% of its employees, but the participant group represented every department and included all senior executives. The composition of the group matters more than its size: it should include both the people who perform the work and the people who ultimately approve the budget.
How should targets be set before the rollout begins?
First, calculate how much time and labor the existing process requires. Then define figures that can be directly compared when the initiative ends, such as the number of hours saved per employee per month. Credit Saison set a target of 7.5 hours saved per employee per month, allowing it to determine clearly at the end whether the target had been achieved.
Do managers need to use the tool themselves?
Yes, this is recommended. Managers who have used the tool themselves are better positioned to determine which results are credible and which tasks are suitable for AI. They can also approve a broader rollout with greater confidence. Credit Saison’s president and all of its senior executives were among the first users.
If the first phase delivers strong results, can the company immediately open accounts to everyone?
A phased rollout is recommended. Begin by allowing new units to use processes that have already been tested, limit the scope of accessible data, and retain human review. After the first group has used the system for a period of time, examine exceptions and costs before deciding whether to proceed with the next group.
References
- Credit Saison’s ChatGPT Enterprise Pilot Achieved 954% ROI — OpenAI’s official customer case study describing the design and results of the first phase and the company-wide rollout process.
