In AI, sovereignty, trust and context cannot be solved separately.
By Professor Ashley Braganza
Sovereignty, trust and context are usually discussed as separate concerns. They are not. A bank cannot earn a customer’s trust without understanding that customer’s context, and it cannot own that context if the data and the intelligence layer beneath it are rented from someone else. Pull the three apart and each one weakens. Hold them together and they reinforce, which is the harder and more valuable discipline. This is the shift AI forces on financial services, and on any organisation that handles something a customer cannot afford to see mishandled.
I reached this view in conversation with Peeyush Aggarwal, senior partner at Deloitte, whose thirty-six years in enterprise technology run from writing FORTRAN for air cargo systems to advising bank boards on adopting AI safely. He kept returning to two words, context and trust, and beneath them a third, sovereignty. Peeyush refuses to let any of them collapse into a slogan, and the result is a practical account of what the three demand when they are treated as one.
Context is the asset, and it is harder to own than it looks
Context sounds soft until you try to own it. For a bank it means understanding what a customer actually needs, absorbing the full weight of regulation, and reconciling demands that genuinely conflict, all at once. Peeyush is precise about where the difficulty sits. The book of record captures what happened, one account paid another, but the intelligence lives in the why, the how and the who, and much of that has never been written down. It sits in procedures and in people’s heads. An enterprise that builds AI on transaction data alone is constructing half the picture and then wondering why the outcomes do not make sense. There is a commercial edge to this. A large language model predicts the next word in a sentence, so if two institutions send the same prompt to the same model, they receive the same answer. Anything a competitor can rent cannot be the source of your advantage. The context around your brand, your regulatory environment and your customer is the one thing a rival cannot buy off the shelf, precisely because it is specific to you and largely undocumented. Banks have spent years mastering the structured data problem. The unstructured layer, the reasoning and the intent behind each decision, is the work that remains, and it is the work usually deferred because it is genuinely hard.
Trust begins with context, and it cannot be deployed
Context is also where trust starts, and this is the link organisations miss. A customer will not trust an institution that visibly fails to understand their situation, so the quality of your context is, in effect, a signal of your trustworthiness. Peeyush draws the sharp distinction: trust, unlike technology, cannot be deployed. It has to be earned, transaction by transaction, and a single agent taking a wrong turn can undo it in one. He uses Waymo to make the point concrete. A self-driving car that turns the wrong way and causes harm does not lose a feature, it loses the public’s confidence. The same exposure now sits inside financial services. His warning about multi-agent systems belongs here too, because it is a trust problem before it is a technical one. The risk is not that one agent gets a decision wrong. It is the interaction between agents, where intent drifts as they talk to each other, and by the time a decision lands nobody can trace who decided what, when and why. Intelligence, he says, is in the interaction, and so is the loss of accountability. A bank that cannot explain how an automated decision was reached cannot defend it to a regulator or to the customer it affected, and it forfeits the very trust it was trying to automate. Traceability is not a compliance afterthought. It is the mechanism by which trust survives contact with autonomy.
Sovereignty is what keeps the trust defensible
Sovereignty is the third side, and it is what makes the other two hold. If the intelligence layer is rented, the context that earns trust is not truly yours, because it runs on infrastructure and models you do not control and a competitor can hire. Peeyush is blunt about the principle: you do not outsource your intelligence layer. He extends sovereignty down the stack, from energy and compute through the model to the data and applications, and up to the enterprise level as a design decision rather than a national slogan. The national picture gives the argument its urgency. A high proportion of UK public sector data already sits on infrastructure controlled by American interests, a position that accumulated quietly while everyone celebrated open networks. He points to DeepMind as British-born capability that moved under foreign ownership, and to a voice-model company he characterises as British that in his account drifted the same way, as talent and capital followed the money offshore. The lesson is not economic nationalism. A firm that does not hold its own data and intelligence ends up renting the very things that were meant to set it apart, and with them the trust of the people it serves. Sovereignty, context and trust turn out to be one commitment described from three angles. Own the context, earn the trust, and hold the sovereignty that protects both.
Name the metric before you approve the agent
None of this survives contact with a poorly framed strategy, so Peeyush offers four pillars, and they are more demanding than they first appear. Give people a genuine choice of tools, because they will solve their own problems once they understand the do’s, the don’ts and the maturity of what they are handed. Tie every deployment to a business metric a leader will own, whether growth, risk, margin or profit, and move the conversation away from the technology and towards the problem it solves. Instrument trust directly into the platform, because a governance decision that once took six months now has to resolve in a split second, closer to a nuclear reactor’s kill switch than to a committee. Put data and context at the heart, since the intelligence layer is only as good as what feeds it. I would add one caution. A single metric, chosen carelessly, is dangerous, and cost reduction is the usual culprit. You can cut the cost and still kill the business. The discipline is to name, in advance, the full set of metrics a deployment must move, and the context it draws on, and to hold it to both. Organisations that skip this will not notice the damage until the saving has already hollowed out the thing that made money.
One last insight from this stimulating conversation. If consumer AI moves onto the devices we already trust, and our preferences live in our own hands rather than in any single institution, the customer relationship may drift away from the bank. Or the bank becomes the ecosystem consolidator, the trusted layer between people and an automated world. Trust decides it, and trust rests on context, and context rests on sovereignty. Treat the three as one and an institution has a chance of keeping its place. Treat them separately and it will lose all three, one transaction at a time.
Listen to the full conversation with Peeyush on The AI Adoption Podcast.
Comments