What a Kerala Bank Should Ask Before Letting an Agent Touch Customer Records

Before an AI agent reads a single customer record at your branch, you need answers to a short list of questions, and none of them are technical. The list below is written for a branch manager rather than a CTO. If your vendor cannot answer these in plain language, that is your answer.

The one line version: you stay responsible for the customer's data no matter whose software is processing it, so ask where it goes, who can see it, what is recorded, and how you would stop it.

Where does the data physically sit

Ask where records are stored and processed, country by country. Ask whether any copy leaves India, including for support, backups, or model training. Ask it about every part of the system, because the main database and the logs often live in different places. It helps to know in advance every place a copy appears when an agent reads a record.

Indian data protection duties and the expectations banks work under have views on storage and localisation, and those expectations change over time and differ by the type of data. Take the answer to your compliance team and check it against current guidance rather than assuming that a vendor's summary is the whole picture. Banks with Gulf operations have a second set to satisfy, since UAE rules reach AI in HR workflows.

A related question that catches people out: if the vendor uses a third party for any part of the processing, where does that third party store things. The chain is only as short as its longest link.

Is the customer's data being used to train anything

Ask directly whether records processed by the agent are used to improve any model, and get the answer in the contract rather than in an email. Ask the same about support access: can vendor staff read customer records while helping you, and is that logged.

The safest answers are the boring ones. Some systems keep processing entirely within your own environment. SicherOne, for example, allows private models to be self hosted, which changes this question from a matter of trust to a matter of where your servers are. Whether that is right for your branch depends on what your bank can actually run, so read up on what self hosting changes for a compliance officer before the meeting.

What did the customer agree to

Consent is not a single tick box. Work out, for each use, what the customer was told when they gave their information and whether an agent reading it falls inside that.

Questions to put to your compliance team, not the vendor:

• Does our existing notice cover automated processing of this data
• Do we need to tell customers that an agent is involved, and in what terms
• What happens if a customer asks us not to process their data this way
• How would we honour a request to delete, when the agent has produced summaries built from that record

That last one is the one people miss. Deleting a record does not automatically remove what was written about it.

Who sees what inside the bank

An agent with access to customer records is a new door into them. Ask who can prompt it, and whether the agent respects the same permissions your staff have. A teller who cannot open a loan file should not be able to ask an agent to summarise it.

Ask whether access is by role, whether it can be restricted by branch, and whether you can see a list of who has it. Then ask who at the vendor can grant themselves access, and how you would find out if they did.

What gets recorded

You need a record that answers, months later, what the agent did, on which records, at whose request, and what a person approved.

Ask for these specifically:

• A log of every record the agent read, not only the ones it changed
• The prompt or request that triggered each action
• The name of the person who approved any output that left the bank
• How long logs are kept and whether you can export them
• Whether the logs themselves are protected from editing

If an inspection or a customer complaint arrives, this log is your position. Without it you are relying on memory.

What can the agent actually do

Draw the line before switching anything on. An agent preparing a draft for a person to approve is a different risk from an agent that can act on an account.

The safe default for customer records is that the agent reads and prepares, and a named person approves anything that reaches the customer or changes a balance, a limit, or a status. Anything irreversible should need two people or no agent at all.

Who carries the blame

Ask the vendor, in writing, what happens if their system leaks or misuses customer data. Ask what they will tell you, how quickly, and what they will pay for. Then assume the regulator will still come to you, because in most arrangements the bank remains answerable for its customers' data regardless of who built the software.

The last question

Ask how you would turn it off. Not the marketing answer. The specific steps, how long they take, who can authorise them, and what happens to the records the agent has already produced.

A vendor who has thought carefully about your side of the risk will have a clear answer ready. That answer tells you more than any demo.