Banking has spent half a century digitalizing how money moves and how financial services operate. Now, with AI, it is beginning to digitalize something far more complex: the judgment behind banking decisions.

Ask an underwriter about a self-employed borrower’s uneven income, or an investigator why a string of ordinary payments feels wrong. Their answer will be based on their experience and the questions they know to ask as long-time experts in their field.

AI can help them answer those questions faster. In 2024, the Bank of England and the UK’s Financial Conduct Authority (FCA) found that 75% of firms responding to their survey were using AI, up from 58% in 2022. As the use of AI continues to increase, banks now face a critical question about the ownership of institutional knowledge: When an AI system learns how we make decisions, who controls what it has learned?

I believe every major bank will need sovereign AI. I do not mean a bank-owned frontier model or a warehouse full of GPUs. I mean the ability to use powerful AI while retaining control of the knowledge it learns, the actions it takes and the decisions the bank must defend.

The reverse information paradox

Satya Nadella frames this well in describing what the term “reverse information paradox” means. Nadella explains that “In the AI age, the buyer risks giving away knowledge, just in order to use what they bought.”

His point is simple — and uncomfortable: A company pays for AI, then reveals its own knowledge to make that AI useful.

In a bank, that knowledge is not just a document fed into a prompt. It is the credit officer’s correction, the fraud analyst’s rejected lead, the exception an experienced underwriter accepted, and the reasons these decisions were made.

All this information teaches an AI system how the bank thinks. I am not suggesting that enterprise AI providers automatically use the prompts, documents or other data submitted by a bank to train models shared with other customers; many provide protections against that. The more subtle risk is that a bank’s evaluations, feedback, memory and workflow logic become so embedded in one platform that it cannot take its hard-won expertise elsewhere.

In my view, this knowledge layer is the more valuable asset than the model itself. A bank can use external models and still own its private evaluations, approved knowledge, feedback and orchestration. If it changes models next year, that institutional memory should not be lost.

Data location is not the same as data control

It is tempting to make this a question of where the server sits. Data residency is important, but a model can run in a bank’s own building and still leave the bank unable to explain a decision or switch suppliers.

Sovereignty comes down to a few practical questions: Do we control our data? Can we explain the decisions the system makes? Do we retain what it learns? And can we keep operating when something fails?

An agent’s instructions alone will not provide this level of control. Its permissions, network access and spending limits need to sit outside the model. Critical actions need approval and an effective way to stop them from happening if that approval is not given. These controls can exist in a public cloud; they can also be missing from a private deployment.

The FCA expects firms to apply existing duties and governance to AI, including consumer outcomes and senior accountability. That means being able to answer simple questions: What evidence did the system use? Which policy applied? What did the agent do? Who challenged its recommendation? What explanation did the customer receive?

In a previous blog, I called the record behind a material outcome a decision passport. It should capture the relevant inputs, permissions, policy and model versions, tool actions, human interventions and final explanation. A bank must be able to recreate that decision record even after it changes models. For consequential decisions, the evidence must be more durable than the model itself.

A mortgage shows what the architecture must do

Take a mortgage application submitted by someone who has an irregular income. AI could reconcile payslips and bank statements, find the relevant policy, flag conflicting evidence and prepare an affordability recommendation.

For this decision, I would want the underwriter to see the conflicts and assumptions, not a confident paragraph that hides them. If the application is declined, the customer deserves a clear reason for the decision and a way to correct the record or seek review.

Then, the bank needs to learn from what happens next. An upheld appeal may expose a faulty interpretation. If a borrower later falls into arrears, the bank can assess whether the original affordability assumptions adequately reflected the risk. A complaint may reveal that the explanation failed the customer.

Each is useful feedback that can be tested before it changes a live system. I would keep a clear boundary between learning and production: learn in a sandbox, test against cases the system has not seen, then release the change under control.

There is a fairness test too. Historical lending data can reflect past exclusions, while declined applicants can create gaps in the outcome data. The bank must look at who is referred, who is rejected and which circumstances the system struggles to understand as the economy changes.

The same principle applies beyond mortgages, in areas such as fraud, customer service and operational resilience. What matters is the outcome, not how convincing the AI sounds. Are there fewer avoidable referrals? Lower false positives? Faster resolution? Sounder decisions? Better customer experiences?

That is the real test of sovereign AI: whether the bank remains in control of how it learns, decides and improves.

What a bank actually needs to own

Banks do not need to own every part of the AI stack. But they do need to control what makes their AI distinctly theirs: their data, knowledge and evaluations. They need the freedom to change models without losing context, alongside clear controls over what agents can do and decision records that survive each technology change. For critical services, they also need a tested fallback. Some tasks warrant private inference; others are better served by a smaller model or a clear deterministic rule.

So where should a bank start? I would choose one material process and establish what a good decision costs today, how long it takes, where it goes wrong and what customers experience. Then measure whether AI actually improves those outcomes. Cost per token tells me very little if the result is a faster, cheaper mistake.

I would also think carefully about what happens to human judgment as more of the work is automated. If AI builds every first-pass credit pack, how do junior underwriters learn to challenge it five years from now? Give them supervised exceptions, simulated cases and access to model traces. And make human oversight meaningful — a person clicking “Approve” without the opportunity to investigate and learn from the evidence is not going to add value to the decision or its outcome.

These are the choices I believe banks need to make now. Powerful AI will be available to everyone. The real difference will be whether banks can turn it into a governed extension of their institutional judgment without giving up control of what they know, how they learn or the decisions they remain responsible for.

A bank can rent a model. It cannot rent accountability.

This is why I expect sovereign AI will become part of core banking infrastructure.

What to do next

Download an IDC InfoBrief on sovereign AI for banking and financial services, sponsored by NTT SATA to learn more about the private and sovereign AI architecture gap in banking, the enterprise AI factory approach, strategic use cases and more.