AI
How to Train a Bank Team on AI and Actually Verify the Skill
The first session for a bank team runs on one exercise. An employee solves a task from the procedure manual. First by hand, then with AI. Then they compare both answers and check each one against the actual rule text.
If you do not give the AI the procedure, it can invent a process your bank does not have. Training teaches people to spot where the system fills gaps and to pull it back to the document.
Pick a task the employee can verify
A branch manager, a call-center agent, and an analyst each have different daily work. The exercises should differ accordingly. The manager checks whether the AI drafted a customer reply according to the procedure. The agent spots whether the AI missed the condition under which a case cannot be closed remotely. The analyst cross-references whether the AI swapped numbers in a report or dropped a step from a workflow.
Every role needs a source document the employee can use to check the answer.
The Central Bank of Uzbekistan reported on May 7, 2026 that two-day training courses for branch managers of Tashkent-region banks took place May 4-5. This was not a Majento course, but it shows the region is already thinking about team preparation.
AI Masters is an education project by Sergey Shima, co-founder of Majento. His team programs adapt to a company's specific tasks. Below is a fictional training example for a bank team.
More on the first team session: our walkthrough.
Prepare the training materials
Before the session, agree on materials with the process owner, the IT department, and the bank's security service. Use fully fictional documents for exercises unless real ones can be anonymized with approval.
Training data kept separate from production data. An empty screen in the illustration does not prove safe file deletion.
"Your own server" does not mean data goes nowhere. Do not promise that the AI runs only inside your internal network until you have checked what your systems actually allow. A training exercise does not connect to live accounts and does not send messages to real customers.
Check the answer against the procedure
Training example. The procedure manual of fictional Bank X reads: "A request to change a contact number is accepted by a branch employee after verifying an identity document."
Customer message: "I am on a business trip. Can I change my number in the chat?"
Task for the employee: draft a reply strictly per the procedure. Quote the basis. Mark missing conditions. Do not invent a remote process.
Correct answer: do not promise a number change in chat. Refer to verification at a branch. Request the approved procedure for business trips or exceptions.
Wrong answer (fictional training case): "Of course, we can change your number right here. Send a photo of your passport in the chat and we will handle everything." The AI invented a process that does not exist in the document.
Notes on the draft remind the reader about manual review of the response.
This is a fictional bank and a fictional procedure. The text is not advice to customers of a real bank and not a legal rule.
How AI responds to company documents, we covered separately.
Assess the employee's independence
Compare two ways of doing the same training task. First the employee works by hand, then with AI. Record citations from the procedure, missed conditions, and time spent, including time spent verifying.
Then, an independent final assignment. The employee writes their own instructions to the AI without the trainer's guidance. Then checks the answer and marks missing conditions.
One more step: a colleague repeats the saved instruction on a different training case. Their questions and mistakes reveal what the instruction needs clarifying. One identical answer does not yet prove the skill is learned.
This is a proposed method, not a completed measurement. The acceptable error threshold and the scoring method get agreed on for the chosen task.
Identify champions and projects
A champion is a staff member who helps colleagues figure out AI and collects tasks for further testing.
Majento trains teams in groups of up to 15. Offline is preferred, but online works. The program adapts to the specific business's tasks and processes. After sessions we identify champions and select projects for implementation.
See what a training program looks like here. If the team is ready to move from sessions to implementation, the implementation section describes the next step.
Frequently asked questions
Can training happen remotely?
Yes. Groups up to 15. Offline is preferred, but online works. The program adapts to the tasks.
What if the bank has no procedure for the needed task?
Pick a different task for the first exercise, one where an approved source exists. Discuss process gaps with the owner.
How many employees per group?
Up to 15. Larger groups lose feedback, and the trainer cannot keep up with everyone.
Must AI connect to the bank's live systems?
No. The training exercise works on fictional documents or approved anonymized ones. Connecting to real accounts or customers is not required at the training stage.
Gather a training procedure and one example customer case. That is the starting point for preparing the exercise. Before the session, agree on materials and the verification method. Write if you need help choosing tasks: Telegram or email.