Exclusive: Google Cloud reshaping finance with agentic AI

Google Cloud is urging financial institutions to scale their use of agentic AI, arguing that fragmented deployments are limiting returns. A governance-led platform approach, it says, is essential to unlock broader value.

Platform shift

"Over the last few years, we've seen a lot of implementation. People have implemented AI for very specific use cases. What we're seeing is a value gap, because these individual implementations are not getting broad enterprise value," said Sid Nadella, Director, Financial Services Market Leader, Capital Markets, Google Cloud.

"It's not going to be one agent or two agents or three agents. It's going to be tens of agents doing different things across the enterprise. You need that governance layer to implement agentic applications across the enterprise," added Nadella.

The focus is shifting towards a platform layer that connects multiple agents while maintaining oversight and compliance. This reflects the operational complexity of financial institutions, where systems must integrate across trading, risk, compliance and customer functions.

Deterministic balance

"I wouldn't call it deterministic. The nature of LLMs is still probabilistic, but we've used techniques like grounding and fine-tuning to make them more explainable and grounded in enterprise data," said Nadella.

"But an agent can have deterministic skills. For example, computing portfolio risk is a deterministic calculation. You give the agent that skill so it uses an approved model rather than generating its own code," added Nadella.

This hybrid approach combines probabilistic reasoning with deterministic processes, aligning with existing financial controls.

"Every bank has approved models. You're not allowed to use an unapproved model to compute risk. Agents are going to act in the same way, using the same skills a human would," said Nadella.

Regulation pace

"I think there are different regulations and requirements for every geography. It can be data sovereignty or how data needs to be set up," said Nadella.

"The pace of development in AI is so fast that we've never seen that before. Digital transformation in financial services took 20 years. We've seen exponential growth in AI capabilities over the last four years," added Nadella.

This pace is creating tension between innovation and oversight. In fraud detection, for example, machine learning reduces false positives, challenging established regulatory expectations.

"When you implement machine learning, the number of false positives goes very low. Regulators are used to seeing false positives, so they think something is wrong. But the system has just become better at detecting patterns," said Nadella.

Addressing this requires close collaboration with regulators. Work with HSBC on anti-money laundering involved extensive engagement to secure approval.

"We worked with HSBC on anti-money laundering, and a big part of that was engaging regulators and taking the model to regulatory approval. It took time because we had to educate them on why this approach is different," said Nadella.

A similar process is under way with the Chicago Mercantile Exchange as it moves core operations to the cloud.

"We work closely with regulators like the CFTC. They analyse the infrastructure and certify it, so we work together to make sure it's acceptable," added Nadella.

Security focus

"Security has always been foundationally important across all our products," said Nadella.

"As AI becomes more powerful, it can be used for malicious reasons, so you need AI for defence as well. We're bringing agentic defence that can be more proactive than reactive," added Nadella.

Security remains the primary concern for financial institutions adopting cloud services, with most engagements starting at the infrastructure level.

"It's not one feature. It's a pattern built over time across our infrastructure," said Nadella.

Investor access

"There's an increased appetite to invest using sophisticated financial instruments that used to be considered complex," said Nadella.

"What is sometimes missing is the educational part. It's hard to understand which instrument to use, what the risks are, and what is right for your situation," added Nadella.

AI is expected to play a growing role in improving financial literacy, particularly among younger investors.

"Even if regulators say you can't use it for advice, it helps people get educated. It can explain products in simple language and personalise it based on your knowledge level," said Nadella.

Use cases

"I think compliance and fraud detection have a lot of potential, and we're already seeing that," said Nadella.

"Settlement is about operational efficiency. Agents can detect operational breaks and fix them quickly. Risk management is still very deterministic, so it will be a combination of agentic AI and deterministic skills," added Nadella.

Adoption is likely to advance first in compliance, fraud detection and operational processes, while trading and risk functions remain closely tied to established models.

"I think the misconception is about what the current capabilities are and how you architect an agent. You need foundational platform capabilities and the right governance and observability," said Nadella.