Updated November 2026.
Most banks, insurers and NBFCs we speak to don’t have an ideas problem. They have a list of good AI pilots that never made it past InfoSec. This post is about that gap: what regulators in India and the UAE now expect, which workflows reach production first, and what an architecture looks like when the risk team signs it off rather than sends it back.
Why AI pilots in BFSI stall before production
We hear the same four stalls in nearly every first call with a financial institution.
- 1. The pilot never clears security. It ran on a public API with broad access, and the review asks questions the design can’t answer.
- 2. The back office runs on email and Excel. Three teams share one spreadsheet, so there’s no clean data for an agent to act on.
- 3. The bot answers FAQs, then stops. It can explain a nominee change but can’t make one, so customers still end up with a person.
- 4. The data can’t leave the network. Data residency rules out the cloud model the pilot was built on.
None of these is a model problem. Each one is an architecture or process decision made too late. We wrote about the general pattern in Why AI Pilots Stall Before Production; BFSI simply has the strictest version of it.
What regulators in India and the UAE now expect from AI
Both markets have moved in the last 18 months, and in the same direction: AI is a board responsibility, third-party AI is still your responsibility, and every system needs to be inventoried, risk-rated and explainable.
| India: RBI FREE-AI framework | UAE: CBUAE AI Guidance Note | |
|---|---|---|
| Issued | 13 August 2025, by a committee constituted by the RBI (RBI) | 11 February 2026, announced 23 February (WAM) |
| Who it covers | Banks, NBFCs, fintechs and other regulated entities | All licensed financial institutions, including insurers |
| Structure | 7 guiding principles (“sutras”), 26 recommendations across 6 pillars | Principles-based guidance on consumer protection and responsible AI and ML use |
| What it asks of institutions | Board-approved AI policy, AI risk management, explainability, incident reporting, safeguards in contracts with third-party AI providers | Board and senior-management accountability, an inventory of all AI systems including third-party ones, a risk rating for each system |
| What it means for your AI vendor | They must work within your outsourcing and IT-risk rules and give you logs, explanations and contractual safeguards | Their system has to fit your AI inventory and risk rating, and their failures are treated as yours |
Neither document is a ban or a licence. Both describe the evidence a supervisor will expect to see. Practically, that means any AI system in a bank or insurer should arrive with five things: a clear owner, a written risk assessment, an audit trail, a way to explain decisions that affect customers, and a human sign-off on anything involving money or policy changes.
In India the Digital Personal Data Protection Act, 2023 sits alongside FREE-AI. In the UAE the guidance is read with the federal data protection law, and institutions in DIFC and ADGM have their own regimes.
Where AI works first in BFSI: five workflows in production
Over a third of what KnackLabs builds is for banks, insurers and fintechs. These are the workflows we’ve seen reach production, with numbers where they’ve been measured.
| Workflow | What the AI does | Example and result |
|---|---|---|
| Tender and RFP response | Finds tenders across portals, ranks them by fit, writes a cited report and answers the bid team’s questions in Microsoft Teams | Tender CoPilot at The Manipal Group: 280+ portals watched, live across 17 business units, a cited tender report in about 9 minutes instead of a day, and 95% of 3,000+ questions answered in under 50 seconds |
| Policy servicing | Lets policyholders change nominees, bank details and PAN, start a loan or upload documents in one WhatsApp or web conversation | IndiaFirst Life, connected to policy admin, CRM, document management, payment gateways, OTP, bank account validation and loan systems |
| Back-office request tracking | Puts every client request, owner and SLA on one platform instead of spreadsheets, so agents have clean data to act on | Verity Knowledge Solutions, for a bank back office |
| Merchant onboarding by voice | Calls new merchants to activate a device, run a test payment and explain the plan, with a person on hand | Paytm SoundBox onboarding, on Ukti, our voice platform |
| Collections and payment reminders | Routes each call: reminders and order status to an AI voice agent, disputed charges to a person | Operon, the contact-centre model we built with Flamingo Infinite |
The pattern across all five: the agent does the full job inside existing systems, and people approve the decisions that matter. That’s the difference between an AI agent and a chatbot, which we explain in Agentic AI in Production.
An AI architecture your risk team can sign off
The fastest way through a BFSI security review is to design for it from week one. These are the controls we build into every AI agent for financial services.
| Control | What it means in practice |
|---|---|
| Your infrastructure | Public cloud, private cloud or on-premise, on infrastructure the institution controls |
| Models you approve | Frontier models, or open-source models such as Llama, Mistral or Sarvam running on your own hardware when data can’t leave the network |
| Permission-checked actions | Every action is checked at the boundary; the agent can’t grant itself new powers |
| Secrets outside the model | Credentials are held separately and never shown to the model |
| Append-only audit trail | Every read, write and decision is recorded with who approved it, and can’t be edited afterwards |
| Human sign-off | A named person approves money movement and policy changes |
| Evaluation harness | Each agent is tested against known answers; below the bar, it hands over to a person |
KnackLabs is certified to ISO/IEC 27001 (certificate GCAI-ISMS-FP64, issued 22 September 2026), and our SOC 2 audit is in progress. We don’t hold or claim any approval from the RBI, IRDAI or CBUAE; no vendor can. What we can do is give your InfoSec and risk teams the architecture, access model and logs to review before anything goes live.
India and the UAE: what changes between the two markets
The controls are the same. The details aren’t.
- Language. In India, servicing and collections often happen in Hindi, English and Hinglish, and voice matters as much as chat. In the UAE, workflows mix Arabic and English.
- Where the data lives. Indian institutions weigh FREE-AI alongside the DPDP Act and their own outsourcing rules. UAE institutions often need in-country hosting, and free-zone entities follow DIFC or ADGM rules.
- Who’s buying. In India we see the most demand from insurers, NBFCs and fintechs moving fast on servicing and onboarding. In the UAE, the CBUAE guidance has made AI governance a board topic, so the first conversation is often with risk, not IT.
We covered the wider differences between the two markets in our guide to AI-native services for mid-market companies in India and the UAE.
A six-week path to your first BFSI agent in production
- 1. Pick one workflow. One process, one owner and one number to move, such as hours per tender or cost per servicing request.
- 2. Build inside your walls. A forward deployed engineer joins your team and builds in your cloud or data centre. A working prototype usually exists in the first week.
- 3. Clear your review. InfoSec and risk see the architecture, access model and logs before launch, not after.
- 4. Go live, then extend. Production quality takes about six weeks. The next workflow reuses the same platform, controls and audit trail.
If you’re still deciding which workflow to start with, our AI strategy and roadmap work ranks them across lending, servicing and operations.
Eight questions your risk team should ask any AI vendor
- 1. Where will the model, the data and the logs physically run?
- 2. Can we use an open-source model on our own hardware if policy requires it?
- 3. How is every agent action permission-checked and logged?
- 4. Can the audit trail be edited after the fact?
- 5. Which actions need a named person’s approval?
- 6. How is accuracy measured before and after launch?
- 7. What security certifications do you hold today, and which are in progress?
- 8. Who owns the data, the configuration and the code when the contract ends?
See our work for banks, insurers and fintechs on the AI for BFSI page.
Frequently asked questions
What is AI in BFSI?
AI in BFSI is the use of artificial intelligence across banking, financial services and insurance: servicing customers, onboarding, collections, underwriting support, fraud detection and back-office operations. In 2026 the focus has shifted from chatbots to AI agents that complete tasks inside core systems under human approval.
Is the RBI FREE-AI framework mandatory?
FREE-AI is a framework of 7 principles and 26 recommendations from a committee constituted by the RBI, published in August 2025. It sets the direction regulated entities are expected to follow, including board-approved AI policies, AI risk management and safeguards in contracts with AI providers.
What does the CBUAE AI guidance note require?
The Central Bank of the UAE’s guidance note, issued in February 2026, applies to all licensed financial institutions including insurers. It expects board and senior-management accountability for AI, an inventory of every AI system including third-party ones, and a risk rating for each system.
Can AI run on-premise for banks?
Yes. AI agents can run on a bank’s private cloud or own data centre, using open-source models such as Llama, Mistral or Sarvam, so customer data never reaches an outside AI provider. Every action can be logged in an append-only audit trail.
What are the most common AI use cases in banking and insurance in India?
Common production use cases include policy and account servicing on WhatsApp, merchant and customer onboarding by voice, collections and payment reminders, tender and RFP responses, and back-office request tracking.
How long does it take to deploy an AI agent in a bank?
A working prototype can be ready in the first week and a production agent in about six weeks, provided the security review runs in parallel with the build rather than after it.

Book a 30-minute call and we’ll show you how it could run inside your own cloud.
Sources
- Reserve Bank of India, FREE-AI Committee Report (13 Aug 2025)
- KPMG India, RBI’s FREE-AI committee report in the financial sector
- Law.asia, RBI issues FREE-AI framework
- WAM, CBUAE issues guidance note on responsible use of AI in the financial sector (23 Feb 2026)
- Modulos, CBUAE Guidance Note FAQ
- KnackLabs, AI for BFSI and Tender CoPilot case study
Note: This article is for general information only and is not legal, regulatory or compliance advice. Summaries of RBI and CBUAE publications are in our own words and reflect the documents as of November 2026; read the originals and take advice from your compliance team or counsel. KnackLabs does not hold, and does not claim, approval from any financial regulator. Results from KnackLabs projects are specific to those deployments and do not guarantee similar results elsewhere. KnackLabs accepts no liability for any loss arising from reliance on this article. Third-party names and trademarks belong to their respective owners.
- AI Transformation
- BFSI
- Banking
- Insurance
- NBFC
- India
- UAE
- AI Governance



