How to Prove ROI for a Financial Services AI Agent

In 2025, most of my conversations with leaders about agents focused on experimentation. Few had an agent supporting a business-critical process in production, and the ones that did often couldn't predict what it would cost in the first year, or how much value it would generate.

In 2026, teams are higher on the learning curve, and building a simple agent is now close to trivial. Models reliably call tools; agent SDKs now include basic architecture out of the box, and Model Context Protocol (MCP) makes it easier to connect an agent to a database, a CRM, or the web. And coding agents can tie all of it together in one afternoon.

With the build cost collapsing, leadership wants to see that it works reliably and economically in production. They want proven ROI.

Across our work with enterprise customers, many in the Fortune 500, I've found that what makes any implementation successful is very straightforward: clarity. You need to define what value the project should deliver up front, before designing the architecture or beginning development.

Clarity gets you to production

Earlier this month, I hosted our Tavily for Financial Services webinar with my colleagues Harsh and Yuval. We walked through what it takes to move an AI agent from pilot to a trusted production workflow, using KYC, investment research, and underwriting as examples.

We shared how, in financial services, the challenging part of getting an agent to production is everything around the model. The agent must find the right evidence (data), keep it current, and show exactly where each result came from.

Consider know your customer (KYC) and know your business (KYB) processes at a bank. For a corporate client, analysts must verify the company, trace who owns and controls it, screen it and its owners against sanctions, assign a risk rating, and document every source for audit. Because much of the process is gathering evidence from inside and outside the bank, it's an obvious place for an agent to help.

Banks increasingly recognize this opportunity. The Goldman Sachs 2025 Annual Report listed "client onboarding/KYC" first among six workstreams "ripe for disruption" under One Goldman Sachs 3.0, its new AI-driven operating model.

Now consider two banks building the same KYC agent with the same architecture and the same models. This is an illustrative scenario, but it's built from patterns I see with customers every week.

When beginning the project:

• Bank A records its goal as: "Use agentic AI to modernize KYC."
• Bank B records its goals as: "Cut corporate onboarding from 90 days to 30. Clear 80% of low-risk cases with no analyst touch. Miss no more sanctions hits than we do today."

There is a stark difference in the clarity of business goals the two banks set out to achieve.

Bank A described an action. Bank B described measurable outcomes, with numbers that shape every decision that follows. Here's what Bank B did at each stage to get there.

A clear goal becomes your business case

Bank B's goal contains everything leadership needs to approve, fund, and scale the project:

• A baseline: it takes 90 days to onboard a corporate client today
• A target: 30 days, with 80% of low-risk cases cleared with no analyst involvement
• A guardrail: no increase in missed sanctions hits

With those three in place, ROI becomes a comparison. Ninety days in, Bank B can show exactly how far onboarding time has moved, and how quality has been impacted. Bank A can only show that the agent is doing something, and that cases are completed.

We recommend our customers develop a template for their projects similar to the following. This applies to all use cases even beyond KYC or financial services.

What counts as "value" for your business case

Time saved is the number one metric we see Tavily customers focusing on, across all use cases and industries. It matters, but it's only one part of what belongs in your business case.

We recommend customers look for value in four other places:

• Capacity: The volume your team handles without adding headcount. In practice, analysts move through more reviews more quickly, instead of the process only scaling by adding people.
• Consistency: Whether results are repeatable; the same entity search returns the same evidence for every analyst.
• Coverage: How much of your universe you can now reach. In practice, a research team covers the long tail of companies it couldn't get to before.
• Cost per decision: Total cost, including tokens, per completed case. In practice, better retrieval sends less noise to the model, so you use fewer tokens per decision.

You don't need all four. Pick the one or two that matter most to the business workflow or decision your agent supports and add them to your business case with a baseline.

The best way to see how this works is in a real deployment.

Proof from the field: BMO

BMO has been in banking since 1817 and serves more than 10 million customers across North America. Its enterprise risk organization owns anti-money laundering (AML) compliance.

A clear problem, with a clear cost

Before BMO takes on a new counterparty, it needs to know what the public record says about them. BMO's analysts ran the time-consuming workflow by hand and manually reviewed the results. As the volume of cases grew, they needed either to increase headcount or refine their processes.

Fortunately, BMO had a sustainable answer. They built a retrieval layer through Tavily with secure, tunable search parameters and source-backed output. This gave investigators confidence, and a source of truth for audit purposes.

What changed

Now, when a BMO analyst searches an entity name, Tavily returns ranked URLs with the relevant data already extracted and can return full-page content when needed. BMO's own models score possible risk, and its investigators have the final sign-off.

The result maps directly to the value matrix above. Analysts move through more reviews, faster, and with confidence in the findings (impacting capacity). And every investigator works from the same evidence (creating consistency):

You've been clear on your goal, but results disappoint

If your pilot falls short on any of the metrics you've set, the model is usually the first place to look. But the model is just the brain. When we work with a customer to improve their agent, the first thing we do is separate retrieval from generation.

Look at what the model was given before it answered. More often than not, this is where the problem lies. The most common failure modes are the following:

• Stale knowledge (retrieval quality). Every model has a training cutoff. Ask a specific question about a company and you may get an answer that's months out of date. Ground your agent with live web data, so it has what it needs to gather exactly what you're looking for (sources, topics, timing, domains, etc.).
• Context pollution (useful context per token). Giving the model more data in the hopes that it will make the right decision is not always the best solution. Relevance scoring and filtering must occur before any data reaches the model.
• Lost sources (traceability). If your retrieval layer doesn't carry the source URL through to the output, you will not be able to verify any results once the case has been completed. Traceability should cover the agent too: which queries it ran, what came back, and when. Without that, you can't defend the result in an audit.

Each of these is a property of your retrieval and workflow, not the model. That's good news, because you can engineer all of them.

Production starts with a definition of success

The best advice we give our clients is also the simplest. Investing the time to find clarity on what you need to build, why, and the outcomes you expect it to produce will make every step of the project easier, and your agent more likely to succeed in production (which means ROI for your business).

In summary, our recommendations for developing a successful agent are as follows:

1. Before you build: Design backward from the decision. Connect every finding to evidence (data) and consider what categories of information your agent will need to make these decisions.
2. Before the pilot: Document the expected outcome metric, measure each layer against a baseline, and give every row in our template to an owner.
3. Before you scale: Check retrieval quality, context per token, traceability, and security.

If you're building in financial services, we offer a 30-minute review of your workflow. Bring your use case, and we'll tell you honestly where Tavily fits and where it doesn't. 

I'd love to hear how this framework works for you. Feel free to reach out to me on LinkedIn.