MUFG's rollout put training, information boundaries and local support in place before broad access. A smaller financial services firm can use the same sequence with one team, one approved tool, a clear review boundary and someone close to the work who can help.
Mitsubishi UFJ Bank has been rolling ChatGPT Enterprise out to approximately 35,000 employees.
Before employees could access it, they had to complete mandatory e-learning. MUFG also established information-management processes and approval routes, set clear rules for use and appointed AI champions in each department.
Four months after custom GPT training began, employees had created more than 1,800 custom GPTs around their own work.
The scale is unusual. The operating sequence is still relevant to a much smaller financial services firm.
The boundary came before the login
Giving someone access to an AI tool is easy. The harder questions appear as soon as they try to use it with real work.
Can a client document be uploaded? Which internal records are approved? Can the tool draft client-facing material? Who reviews a summary that contains an unusual figure? What should someone do when the answer looks plausible but the source is incomplete?
MUFG dealt with those questions as part of the rollout. Its reported approach included e-learning before access, information-management processes, approval routes and clear operating rules.
That gives employees a starting boundary. They know where the tool may help, where they should pause and where another person needs to remain involved.
The training does not need to turn every employee into an AI specialist. It needs to help them make sensible decisions when the tool meets the information, responsibilities and exceptions inside their role.
Put support close to the work
MUFG appointed AI champions in each department.
That choice matters because useful questions usually come from the work itself. Someone preparing a client update may want to know whether a particular document can be used. An operations analyst may find that two records disagree. A manager may need to decide whether a draft is ready for review or should be stopped.
A central technology team can set standards and maintain the approved environment. It may still be too far from the daily workflow to answer every operating question quickly.
A local champion can help translate the boundary into the work people are trying to complete. They can collect recurring questions, see where teams are struggling and bring patterns back to the people responsible for the tool, the data and the control environment.
This also creates a more useful feedback loop. Leadership gets a clearer view of where employees are finding value, which requests keep appearing and which proposed uses require stronger foundations.
Let employees reveal the useful workflows
MUFG reports that employees created more than 1,800 custom GPTs within four months of custom GPT training.
That figure shows activity, although it does not tell us how many GPTs moved into recurring or production use. The distinction matters. A prototype can still be useful because it reveals the task someone wants to improve, the information they need and the rules that are currently carried in their head.
An employee who prepares the same report every month may know which spreadsheet arrives late, which client uses a different format and which number deserves another check. Once they start building or testing a workflow, that knowledge becomes easier to see.
The firm can then assess the idea properly. Which sources should the workflow use? How are permissions handled? Which exceptions should stop the process? Who reviews the output? What result would make the work worth maintaining?
Employee experimentation can surface useful opportunities. The operating review determines which ones deserve to become dependable capabilities.
What can a smaller firm borrow from MUFG?
A smaller firm does not need a 35,000-person programme to use the same logic.
One team and one approved tool may be enough for the first stage. The firm can give the group a short practical training session, explain which information is permitted, define the tasks that still require review and name the person who can answer questions.
The next step is to watch what people actually do with the access.
Which use cases appear without being assigned? Where do employees stop because the rule is unclear? Which ideas depend on information the tool cannot access? Which early workflows are useful enough to test against real examples?
After a few weeks, the firm can review the evidence with the people doing the work. Some ideas may remain personal productivity aids. Some may deserve a controlled pilot. Others may expose weak data, unclear ownership or a process that should be repaired first.
Each answer helps the firm decide where to invest.
A practical rollout sequence
Choose the approved environment
Define the tool, user group, information boundary and administrative owner.
Train people before access
Use examples from their work and explain what is allowed, what needs review and what should be avoided.
Name someone close to the team
Give users a practical place to take questions, exceptions and early ideas.
Collect the workflows people try
Look beyond usage counts and record the task, information, review requirement and intended outcome.
Review the evidence
Decide which uses can continue, which need a controlled implementation and which should wait.
Access should follow a decision
AI adoption becomes easier to manage when the firm decides what access is for before it distributes accounts.
The boundary gives people enough clarity to begin. Local support helps them handle the questions that appear inside real work. Their early experiments reveal where the useful opportunities may be.
Halyard helps financial services firms examine those workflows, define the information and review boundary, and turn credible ideas into narrow implementations that can be tested properly.
The first rollout can be small. The important part is giving people a clear place to start and a responsible way to continue.
Further reading
Founder of Halyard, an AI implementation company helping financial services firms find and implement practical AI safely.