Your Data Met AI. Do You Know Where It Went?

Many banks we talk to have done the work. They have blocked public AI tools on bank devices, written acceptable-use policies, and trained staff on what never to share. They should be proud of that. It is a strong start.

But one bank technology leader we spoke with recently put the real challenge plainly. He went back to school to study AI and machine learning, and his view is that generative AI is easy to use in marketing, where a wrong answer has few consequences. In banking, it is much harder to put into production, because almost every answer touches regulated data. He had seen a promising AI project stall before launch once legal stepped in. The concern was never whether the AI worked. It was what would happen to the data.

In our conversations with community banks, this concern keeps coming up. Bankers want the productivity AI promises. They also want to know what happens to their customers’ data once it touches a model.

‍

Three ways customer data reaches AI

Blocking public chatbots closes one door. But customer data can still reach AI through three others, often through tools the bank already trusts:

1. Staff tools. Approved assistants and copilots built into email, office, and core-adjacent software help staff write, summarize, and analyze. AI workflows to automate your workspace. Where public tools are blocked, these sanctioned tools become the main door, and they still need clear rules about what data goes in.
2. Vendor features. The platforms you already use are adding AI: smart search, auto-generated summaries, chat assistants, recommendations. Sometimes these features turn on with an update, and the data flows to a model provider you never evaluated.
3. Document processing. Onboarding and lending increasingly use AI to read IDs, tax returns, bank statements, and business documents. These are some of the most sensitive files a bank handles.

None of these are bad on their own. The risk is assuming a guardrail on one door covers all of them.

‍

What bankers are worried about

When bankers raise AI concerns with us, they rarely ask whether AI works. They ask what happens to the data. The questions fall into four groups:

• Training. Will our customers’ information be used to train someone else’s model? Once data is inside a model, it cannot simply be deleted.
• Retention. How long are prompts, documents, and outputs stored, and where? A tool that keeps everything forever creates a new copy of sensitive data outside the bank’s control.
• Access. Who at the vendor, or the vendor’s AI provider, can see what was sent? Is it encrypted? Is personal information masked or tokenized before it ever reaches a model?
• Keys and control. Does the bank control the connection to the AI provider, or does the vendor? Can the bank turn a feature off, audit what was sent, and prove it to an examiner?

These are not technical questions. They are the same questions banks already ask about any third party that touches customer data. AI just adds a new party to the list.

‍

What should stay inside the bank

Safe AI use does not mean banning AI. It means drawing a clear line between what can leave and what cannot.

Should never go into an unapproved AI tool:

• Social Security numbers, account numbers, and card numbers
• Government IDs and identity documents
• Tax returns, financial statements, and credit information
• Anything that identifies a specific customer alongside their financial details

Can be used safely in approved, controlled tools:

• Customer data that is masked or tokenized before reaching the model
• Internal policies, procedures, and product information
• General drafting and analysis with no customer details included

The best AI in banking is designed so that sensitive data never has to leave in the first place. The model comes to the data, under the bank’s rules, not the other way around.

‍

Seven questions before you turn on any AI feature

1. Which AI model or provider processes our data, and where is it hosted?
2. Is our customers’ data ever used to train or improve a model? Is that in writing?
3. How long are prompts, documents, and outputs kept, and can we delete them?
4. Is personal information masked or tokenized before it reaches the model?
5. Who controls the connection to the AI provider, us or the vendor?
6. Can we turn the feature off, and will we be told before new AI features are switched on?
7. Can we see a log of what was sent and what came back, in a form we can show an examiner?

And one question to ask internally: do our employees know which AI tools they are allowed to use, and what they must never paste into one?

‍

AI your customers can trust

Community banks win on trust. Customers choose you because they believe their information is safe with people who know them. AI should strengthen that promise, not quietly put it at risk.

At Linker Finance, we hold ourselves to the same questions, and we built our AI, including Grandir, on a foundation designed to answer them:

• One governed home for the data. Our orchestrator brings core data together with KYC, external accounts, credit, and fraud signals into a single unified customer profile, with a data lake on top. Data lives in one structured place, not in copies scattered across vendors.
• Our own harness, model agnostic. Grandir runs on an AI harness we built and own. Customer data flows through our API gateway under the bank’s rules, and we are not locked into any single AI model provider.
• The model orchestrates. Our tools calculate. The language model decides which step comes next, but purpose-built tools do the math, like generating a personal financial statement or checking HELOC eligibility against the bank’s own credit policy. Results are consistent and repeatable, not guessed.
• Guardrails from the start. Entitlements, data segmentation by customer, Each bank gets its own data catalog, its own storage bucket and encryption key, and its own access roles to ensure data safety.
  Full data lineage and agent traceability enables you to answer the hardest questions regulators may send your way about what data your AI agent used when helping make a decision.
• Protected and auditable. Sensitive identifiers like Social Security numbers are tokenized, the platform is SOC 2 compliant, and every recommendation is explainable and auditable.

Knowing where your data goes is only the first step. The next is knowing who, or what, is acting on it. Read our piece, about Banks know how to verify people and businesses. The next customer at your door might be neither. Know your Agent to see why verifying AI agents is becoming as important as verifying customers.

If you are evaluating AI for your bank and want a second set of eyes on the questions above, let’s talk.

‍