Governing the AI your staff already use: lessons from our banking workshop.
Obrix Academy
29 September, 2026
00 Comments
In most banks, generative AI arrived the way it did everywhere: staff started using it on personal accounts, with no policy, no verification standard and no record of what went in or came out. The useful question is no longer whether people should use these tools. It is whether that use becomes governed.
That is the starting point for our 2-Day AI Adoption Workshop for Nepalese Banks, built for tellers, relationship officers, credit and trade finance teams, compliance and internal audit, finance, treasury, HR, planning and IT. Nobody needs to be technical, and nobody is assumed to have used AI before.
AI as an assistant, with a human owner
The workshop treats AI as a drafting and analysis assistant that always works under a human owner. It deliberately does not cover AI that decides anything: credit approvals, risk ratings, alert closures, or customer letters sent unread. Every exercise ends at a human gate, and accountability stays with the bank.
What the two days cover
- Day 1: fundamentals and vocabulary, what the tools do well and badly, what may be typed into a tool, prompting, context, spotting low-quality output, and the draft, critique, refine, verify loop.
- Day 2: choosing tools and understanding data flow, turning unstructured documents into clean tables, spreadsheets and variance commentary, and ethics and governance, including prompt injection, deepfakes and AI-enabled fraud.
- Optional sessions: a function lab on each department's own workflows, and a session on workflow design.
Nepal-specific traps
General AI training misses the details that trip up staff here. The workshop shows live how a tool can give a confident wrong answer, and works through local traps such as Bikram Sambat dates, lakh and crore, Devanagari output, and answers written for India when the question was about Nepal. NRB's draft AI guidelines are used as the governing position for the governance session.
Real workflows, fictional data
Everything is practised on workflows people actually have, using fictional data throughout. A few examples:
| Team | Workflow | How AI helps |
|---|---|---|
| Branch staff | Replying to a customer complaint | Drafts a polite reply in Nepali and English for staff to check and send |
| Credit | Writing a credit memo | Drafts the memo from key facts, ready for the credit officer to review |
| Trade finance | Checking trade documents | Compares a letter of credit with its invoice and shipping documents and flags anything that doesn't match |
| Compliance | Comparing policy versions | Shows what changed between the old and new version, in plain words |
| Finance | Explaining budget differences | Writes a short explanation for each line where the numbers are off |
A separate briefing for leadership
The CEO, executive committee and board do not attend the workshop. They get a separate 45 to 60 minute briefing on what the tools do and do not do, the bank's exposure from ungoverned use, and the accountability position. It ends with the four decisions a steering committee will be asked to make: approved tools, the data tier policy, the incident path, and ownership.
Cohorts are 20 to 25 people, with a second facilitator once a group goes above 20. If you would like to bring the workshop to your bank, send us an enquiry from the Workshops page.
Leave a comment