Early in my career, working in a leading global investment bank, I regularly worked 80-hour weeks. Much of that time involved assimilating information into a cohesive and compelling narrative for customers. We pulled numbers and insights off multiple systems, crunched them, and wrote slides through the night. We then sent our deck to a design team who tidied it up and returned it two hours later so we could revise it again. All that work went into a single pitchbook that made our case for why a company should choose our firm to underwrite its debt issuance or advise on its merger.
The manual assimilation and sifting was phenomenal training for my analytical brain, which is still wired that way, but it was enormously time intensive. Well-trained people were spending untold hours assembling information before we could do anything with it. For those working across different parts of the financial services ecosystem, this likely sounds familiar.
How we got here
None of this happened by accident. Banks grew by acquisition, by product line, and by regulatory mandate, and each wave brought its own systems. Fragmented customer data is a well-trodden story, but it remains the reality for most incumbent institutions. I've worked with companies who ran over 20 separate data centres before they consolidated onto the cloud and could finally form an integrated picture of each customer.
The chatbots we deployed a decade ago exposed the problem rather than solving it. A customer would ask a reasonable question, and the bot would reach across disconnected sources. Sometimes it assembled a coherent answer, but more often it couldn't, so it handed the customer to a human who then toggled between systems and constructed an answer by hand. We added a conversational layer that improved the interface but not the underlying fragmentation.
Then generative AI arrived. Banks started using it in the middle and back office, which was completely logical. In a highly regulated industry, you don't run your first experiments in front of customers. We looked at loan processing and asked how AI might extract details from a file that a person had been reading line by line. I think of that work as crushing paper out of processes, and it was genuinely useful.
As confidence grew, institutions became braver and moved closer to the customer and to higher regulatory scrutiny. While some of those programs produced real results, very few produced enterprise-level gains. Banks are now running dozens of successful pilots that haven't yet changed the shape of their business, and finance teams are asking a fair question about the cost of running multiple models against every problem when a cheaper tool could have derived the answer.
The gap between individual wins and enterprise value is the defining frustration of this period. We got what we designed for when we pointed AI at discrete tasks. It made those discrete tasks faster, but the multiple systems remained separate.
What has to be true before AI agents can run workflows
Something has shifted in the last six to nine months. Banking executives no longer ask me whether agents can draft useful documents. They want to know if agents can run a whole workflow in a way that's auditable and explainable and works for all three lines of defense, something easy to say that's difficult to do.
It's difficult because large language models are probabilistic by design, while much of banking is deterministic by regulation. Any credit decision, capital calculation, or exposure limit must run on a model the bank has documented, validated, and defended to a supervisor. Ask the same question twice and you'd better get the same answer, with documentation on how you arrived at it. No responsible institution will let a model invent its own method for computing a number the bank has to defend during an examination.
To run defensible workflows, agents need deterministic skills. An agent can gather filings, read disclosures, and assemble an analysis, but when the workflow arrives at a step that calculates a number, the agent runs the bank's approved model instead of working out a method of its own. In other words, the agent's reasoning is probabilistic while its execution is deterministic. Every output carries its provenance in cited sources, visible methodology, and a line traceable back to the original data.
This combination of probabilistic reasoning, deterministic execution, and provenance on every output is what Gemini Enterprise for Financial Services delivers.
Introducing Gemini Enterprise for Financial Services
At the center of Gemini Enterprise for Financial Services is the Financial Research agent, skills and connectors that run research end to end, pulling together public, licensed, and private data to produce documents, decks, and models. It reaches that data through connectors configured inside the bank's own environment, which let the agent query a source directly, in that source's own format, whether it's a licensed market data provider, a regulatory filings database, or the bank's own internal systems. On top of that sit more than 50 role-based skills for specific workflows, such as a relationship manager preparing for a client meeting or an investigator working a financial crime case.
Consider how this helps the relationship manager. Let's say her client is a multinational with subsidiaries in a dozen jurisdictions, a complex cross-border structure, and exposure to interest rate volatility and climate risk. Proper analysis is precisely the work an agent can do in a fraction of the time it usually takes an entire team. She can bring to the meeting not only a better grasp of the client's structure, but also a view of what their exposure means and which of the bank's products actually address it.
Deutsche Bank served as one of our design partners for the Financial Research Agent. They provided deep banking domain expertise to bring the requirements of a major European regulated bank to bear on security, governance, and data residency from the beginning.
"Starting in the Corporate Bank, we see significant potential to reduce manual research effort, improve the consistency and auditability of outputs, and give our teams more time for client conversations," said Marie-Jeanne Deverdun, Chief Technology, Data and Innovation Officer, and Member of the Deutsche Bank Management Board. "This is an important step in applying AI where it can make a practical difference: safely, responsibly, and at scale."
By using the Financial Research Agent in its Corporate Bank, Deutsche Bank aims to help relationship managers prepare more effectively, identify relevant client needs and opportunities earlier, and turn insights into more targeted conversations.
Where should the time go?
Highly skilled people currently spend the bulk of their workweek assembling information, when their real value lies in judgment, reflection, and hours spent with customers. Handing them assembled research, rather than raw data scattered across a dozen tools, turbocharges people who are already very good at their jobs.
The institutions I see getting real value from agents are the ones that start with foundational capabilities, including governance, compliance, observability, and secure access to their own data. With this foundation, they can deploy the second hundred agents as safely as the first.
Finance is so deeply woven into the fabric of society that, honestly, it's hard to predict exactly where all this will go. Much will come down to how people actually choose to use these tools. But the direction is unmistakable, from banks as balance sheets to banks as advisers. This requires bringing the intelligence of all your data to every customer, so you can see them clearly enough to anticipate what they'll need next and to serve them in new ways.
Explore how Gemini Enterprise supports financial services firms
