AI in Banking in Bangladesh: Where to Start in 2026
TL;DR: For AI in banking in Bangladesh, start with an internal, staff-facing use case such as circular and policy search or KYC document extraction, not a customer chatbot or a credit model. Internal tools keep customer data inside the bank, show value against a measurable baseline, and let your team learn to run AI in production before sector rules are final.
Most bank and MFS leaders in Bangladesh have stopped asking whether to use AI. The harder question is where to begin without creating a compliance problem or a pilot that never reaches production.
Where Bangladeshi banks stand today
AI in Bangladeshi banking is not new, but it is still shallow. Eastern Bank launched a Facebook Messenger chatbot, EBL DIA, at the end of 2017 and described it as the country’s first AI-based banking chatbot. A 2025 op-ed in the Daily Observer lists Dutch-Bangla Bank and bKash among the providers with chatbots in their apps and websites.
Readiness is the bigger issue. A Bangladesh Institute of Bank Management (BIBM) study in 2024 found that only 6% of banks were well equipped to integrate AI and that no bank had adopted it at large scale, although 65% had implemented it to some extent. Later BIBM research, reported by The Business Standard in November 2025, found that 60% of banks had no AI cybersecurity policy and 68% had not written AI into their operational policies.
So the typical starting point is a few vendor chatbots, some analytics and no shared way of governing models.
What the regulator has said, and what it hasn’t
Commentary often overstates this, so here is what is on the record.
- A sector AI policy is in preparation. In November 2025 Bangladesh Bank formed a seven-member team to draft an AI policy for the banking sector, and said it was considering an in-house large language model to reduce the risk of data going to foreign servers (TBS). As of October 2026 we have not found a published sector-wide AI circular, so check bb.org.bd before planning around one.
- The central bank restricts its own staff. A June 2026 office order bars Bangladesh Bank officials from entering confidential banking information into AI tools and requires prior approval for AI use in official work (TBS). It does not bind commercial banks, but it shows how the supervisor views public chatbots.
- Existing rules already apply. Bangladesh Bank’s March 2023 cloud computing guideline covers banks, NBFIs and MFS providers, requires due diligence and gives the regulator access to monitor cloud services. The ICT security guideline, outsourcing requirements and banking law on moving records abroad apply to an AI system as they do to any other.
- The national policy is still a draft. The draft National AI Policy 2026–2030 classes credit assessment as high-risk, requiring algorithmic impact assessments and strict human oversight (Daily Star, January 2026).
The practical reading: design as though customer data cannot leave your environment, log every model decision, and keep a named person accountable for anything that affects a customer’s money or credit. Confirm the specifics against current requirements with your compliance team.
Where AI pays off first, and why
AI pays off first where the work is high-volume and text-heavy, the data is already in the bank, and a person still checks the output. Our banking and MFS page describes the use cases themselves. The table below ranks them as first projects.
| Use case | Customer data leaves the bank? | Regulatory exposure | Data you need | Good first project? |
|---|---|---|---|---|
| Circular and policy assistant for staff | No, if deployed on-premises | Low: staff verify cited sources | Circulars, policies, product manuals | Yes |
| KYC and loan document extraction | No, if deployed on-premises | Medium: a reviewer approves exceptions | Scanned NIDs, trade licences, forms | Yes |
| Fraud and AML alert triage | No | Medium: analysts still decide | Labelled alert history | Second project |
| Bangla customer-service assistant | Depends on design | High: speaks to customers | Chat and call transcripts | After the first project |
| Credit scoring or decisioning | Depends on design | High: listed as high-risk in the draft national policy | Repayment history, bureau data | Later, with model risk governance |
Internal tools come first because their errors are cheap to catch. When a staff assistant gets an answer wrong, the employee checks the cited paragraph. When a customer chatbot gets one wrong, it becomes a complaint or a fraud vector. Get evaluation and monitoring habits in place before you take on that second kind of risk.
For MFS providers the order shifts. With around 144 million registered MFS accounts, according to Bangladesh Bank data cited in the Daily Star, contact volumes and social-engineering fraud dominate. Alert triage and Bangla customer service move up the list, and document-heavy work moves down.
What is different in Bangladesh
Bangla and mixed-language data
Customers write in Bangla, in English and in Bangla typed in Latin script, often in one message. Older internal documents are frequently typed in Bijoy encoding with SutonnyMJ-style fonts. They look like Bangla on screen but extract as meaningless Latin characters, so convert them to Unicode before indexing or your retrieval system will index garbage. Scanned NID cards and handwritten forms add OCR errors on top.
Model choice matters as well. The authors of TigerLLM, a Bangla model family published at ACL 2025, report that widely used open multilingual models still underperform in Bangla. Rather than trusting a vendor’s language list, score each candidate on an evaluation set built from your own material, and re-run it whenever a model, prompt or document changes:
# eval/circulars.yaml (one case from a set of a few hundred)
- id: fx-012
question: "রপ্তানি আয় প্রত্যাবাসনের সময়সীমা কত দিন?"
must_cite: "the current FE circular on export proceeds"
expected_language: bn
pass_if: cites_correct_circular AND answer_matches_reviewerData residency
For sensitive workloads, plan on open-weight models running on your own hardware or in an in-country data centre. A retrieval assistant on a mid-sized open model needs a GPU server, not a cluster. If a hosted model is useful for low-risk work such as drafting marketing copy, define which data classes may be sent to it and enforce that at the gateway rather than in a policy document.
Legacy core banking
Most cores were never designed for AI workloads, and a first project should not write to them. Read from a reporting replica, the data warehouse or the document management system, and put an API layer between the AI service and anything transactional. That keeps the core vendor and your change-control board off the critical path for the pilot.
Engineering talent
Few Bangladeshi banks have engineers who have run LLM systems in production, from retrieval tuning to GPU operations and prompt-injection defences. Hiring them is slow. That is the strongest argument for a delivery model in which your own engineers build alongside experienced ones and keep the system afterwards.
How to pick your first use case
Score each candidate against five questions, and pick the one with the most clear “yes” answers rather than the most exciting one.
- Is there a baseline? Can you measure today’s handling time, backlog or error rate?
- Is the data in reach? Is it digital, or one OCR step away, and can you legally use it?
- Does a person already check the output? Can the AI draft rather than decide?
- Can it run inside your environment? Is there a deployment that meets your data-location obligations?
- Does a business owner want it? Will a department head commit staff time to a pilot and own the result?
A circular and policy assistant for compliance and branch staff usually scores well on all five, which is why it is a common first choice.
Build versus buy
| Option | Strengths | Risks in the Bangladesh context |
|---|---|---|
| Buy a SaaS AI product | Fast to start | Data may leave the country; Bangla quality unproven on your data; hard to exit |
| Buy a vendor platform deployed on-premises | Data stays in; support contract | Licence cost; limited control over retrieval and evaluation |
| Build on open-weight models | Full control; data stays in; models can be swapped | Needs engineers you may not have yet |
| Hybrid: buy components, build the core layer | Speed with ownership | Needs someone to own the architecture |
We usually recommend the hybrid. Buy the commodity parts and build the layer that encodes your knowledge and risk decisions: retrieval over your documents, integration with core systems, evaluation sets and guardrails. Owning that layer means you can switch to a better Bangla model when one appears without starting again.
A 90-day starting plan
Days 1–15: decide. Run an Executive AI Workshop so leadership agrees on the first use case, the baseline metric and who owns risk. Classify the data involved and get compliance sign-off on where it may be processed.
Days 16–45: build a thin version. Convert and index the documents, deploy an open-weight model on-premises and build the evaluation set from real staff questions. Measure answer accuracy and citation accuracy separately.
Days 46–75: pilot. Put the system in front of one department and log every question, answer and correction. Review failures weekly and add them to the evaluation set. Bring internal audit in early, so they see the logging design before they have to audit it.
Days 76–90: decide again. Compare results with the baseline, then scale, fix or stop, and record why. Write the model-change process: who approves a new model or prompt, and which evaluation it must pass. Then choose the second use case.
Throughout, pair your engineers with the build team. Our approach is built around that handover, and the Applied AI Engineering programme teaches retrieval, evaluation and deployment to teams who will run these systems themselves.
FAQ
Has Bangladesh Bank issued rules on AI for banks?
Bangladesh Bank began drafting a banking-sector AI policy in late 2025, and in June 2026 it barred its own staff from putting confidential data into AI tools. We have not found a published sector-wide AI circular as of October 2026. Apply existing ICT security, cloud and outsourcing requirements to AI systems, and check bb.org.bd for new circulars.
What is the best first AI use case for a bank in Bangladesh?
Usually an internal, staff-facing tool: a Bangla and English assistant over circulars and internal policies, or KYC document extraction with a human reviewer. Both use data the bank already holds, can run on-premises and have a measurable baseline.
Can generative AI work well in Bangla?
It can, but quality varies by model and task, and research shows many open models still underperform in Bangla. Mixed-language chat and old documents in non-Unicode Bijoy encoding add further problems. Test candidate models on an evaluation set built from your own data.
Does customer data have to stay in Bangladesh for AI systems?
Treat it as if it does unless your legal and compliance teams confirm otherwise. Bangladesh Bank guidelines, banking law on moving records abroad and the data protection framework all bear on it. Open-weight models deployed on-premises or in-country avoid the question for sensitive workloads.
Should a bank build or buy its AI systems?
Usually both. Buy commodity components such as model weights, OCR engines and vector databases. Build the retrieval layer, core-banking integration, evaluation sets and guardrails, because owning those lets your team change models later.
How are AI use cases different for MFS providers?
MFS providers handle far higher transaction and contact volumes, a large agent network and customers who mostly use Bangla on mobile. Fraud and social-engineering detection, Bangla customer service and agent analytics come first; document-heavy work such as credit memos matters less.