On August 3 Anthropic announced that Claude will soon run inside India. Through Amazon Bedrock, requests sent to an India endpoint will be processed on servers in the country. The company was clear about who it had in mind: banks, financial institutions and government bodies that have kept frontier AI out of production because their data could not leave India.
"When that data can stay in India, AI moves from pilots into the systems that matter most," said Irina Ghose, who leads Anthropic in India.
Hosting Claude in India is an important step. It is also only the first one.
Residency answers one question
Indian banks work under strict rules about where data lives. Since 2018 the RBI has required payment system data to be stored only in India. The Digital Personal Data Protection Act adds its own duties for how personal data is handled. For many compliance teams, a model that sends customer information to servers abroad has been an easy no.
In-country inference removes that objection. It does not settle the harder questions a bank's risk committee will ask.
- Who controls the model? Claude is a closed model. A bank cannot inspect its weights or decide when a version is retired. When the provider updates the model, its behaviour can change.
- Who controls the data path? The data stays in India but still passes through infrastructure the bank does not run. The bank needs to know what is logged, for how long and who can see it.
- Who explains the decision? If a model helps decline a loan or flags a payment as fraud, the bank has to explain why to the customer, to its auditors and to the RBI.
- Can every action be audited? Regulators expect a clear record of what a system did and why. That applies to an AI system as much as to a person.
The RBI is already moving in this direction. Its FREE-AI framework, published in August 2025, recommends board-approved AI policies, audit mechanisms for AI systems and specific contract safeguards when banks rely on outside AI vendors. None of that changes because the servers are in Mumbai.
Two kinds of work
The likely answer is not one model for everything. Banks will split their AI work by risk.
For general enterprise work, closed frontier models make a lot of sense. Customer support, document summaries, internal search, coding help and report drafting all benefit from the best model available. The data is sensitive, but the decisions can be reversed and a person usually reviews the output. Local hosting makes this kind of use much easier to approve. The early names fit the pattern. Axis Bank and IndusInd Bank are using Claude for engineering and internal knowledge work. NPCI is building its agentic AI platform on it.
Core banking is different. Payments, fraud detection, credit decisions and customer financial records are a bank's crown jewels. Here banks want full control over the model as well as the data. We expect many to choose hybrid setups, with open-weight models running in their own data centres or private clouds for the most sensitive work and closed models handling the rest.
Whether banks trust Claude with those crown jewels is a separate question from where it is hosted. The answer will come slowly, one use case at a time.
Trust comes from visibility
Whatever mix a bank chooses, it will end up running several models from several sources. It might use a closed model through Bedrock, an open-weight model in its own data centre and AI agents that connect both to internal systems. Each comes with its own logs, its own access rules and its own ways of failing.
Trust in that setup does not come from the location of the servers. It comes from seeing what every model and agent is doing, catching risky behaviour early and showing auditors exactly what happened. That is the problem SaaviAIDR is built to solve. It gives security teams one view across closed, open-source and open-weight models and the agents built on them. Because SaaviAIDR is self-hosted, the record of that AI activity stays inside the bank's own environment too.
Data residency gets AI through the door. Governance, control and auditability decide how far inside it goes.
First shared on LinkedIn, August 4, 2026. Read the original
