Dr. Ryan Ries here. In case you didn’t tune in last week (shame on you), I kicked off an “AI in your Industry” series. This week we’re covering financial services (most highly requested).
I’ll be running this series for a few weeks. Some of the next ones I’ll be covering are:
- Manufacturing
- Sports and Gaming
- Agriculture
- Education
Use Cases I’ve Personally Worked On
Building a Daily Market Intelligence Engine
I personally found this use case and customer’s business very interesting. They are a crypto brokerage that uses data science to execute trades with precision, eliminating slippage and hidden fees.
They came to us with a problem: they had a genAI system that was pulling in market intelligence from multiple sources — crypto news, CoinAPI price and volume data, derivatives data via Amber, macro trends from Seeking Alpha, and their own internal OTC desk.
The problem wasn't data (they had a ton of it), which is oftentimes a problem I see organizations have. The problem was turning that data into something a human could act on within 24 hours.
We built a solution to automate the transformation of their current AI system’s output into a dynamic visualization layer that could plot meaningful insights on demand as market conditions shifted. We used AWS Glue for data ingestion, S3 for storage across raw, cleaned, and consumption zones, Redshift for modeling, and Amazon Bedrock for the generative layer. QuickSight Q (now QuickSuite) handled the visualization.
Getting a foundation model to synthesize derivatives skew data, OTC buy/sell ratios, and macroeconomic equity trends into coherent, actionable market color in near real time required careful architectural thinking and not just clever prompting.
Teaching a Machine to Find the Next Great Investment
This next one is particularly interesting for private equity firms.
We had a customer that tracks thousands of companies simultaneously, trying to identify the ones worth investing in years before anyone else does.
Their challenge was institutional knowledge that lived entirely in people's heads.
When associates evaluate a company, they're drawing on a web of factors like their own experience, portfolio metrics, network synergy, market position, management quality, transaction dynamics. None of that is cleanly documented and new employees spend years absorbing it.
We developed a methodology for the customer that uses numeric data from their Top Prospects dataset. The system filters the universe of tracked companies down to a manageable candidate set
By layering a ton of numeric fields + a gamut of calculations that I won’t bore you with the details on, we could dramatically reduce the number of companies requiring active associate attention.
Then, and here’s the really cool part, we used LLMs to extract latent signals from unstructured data (news, analyst notes, PR releases) and relationship mapping to identify network connections among investors, directors, and executives across portfolio companies.
I like this use case a lot because it clearly shows that AI does not replace the analyst. It helps them to work smarter, faster, and armed with intelligence.
When the Document IS the Bottleneck
Last but not least, we have a customer who is a cloud-based mortgage loan origination system. Their whole value proposition is speed and precision – a platform built for originators who need to move fast.
The biggest problem was the paper and the unstructured data they were working with.
Underwriters were spending enormous amounts of time manually reading PDFs and images like pay stubs, bank statements, tax returns, debt schedules and re-entering that information into the loan system by hand.
This customer came to us for a document intelligence system. The goal was to figure out how to use AWS to scan, classify, and automatically populate loan data from applicant documents, and then layer LLMs on top to augment the analysis with additional context and insights.
The technical challenge here was OCR on documents that aren't always clean. Mortgage applicants don't submit pristine PDFs. They submit cell phone photos of W-2s taken at bad angles in bad lighting. Low-resolution images of bank statements printed, scanned, and re-emailed. That's where super-resolution algorithms came into the picture. Preprocessing the imagery to improve legibility before extraction even begins.
The roadmap we developed moved through five workstreams in sequence: data pipelines, image processing, image classification, loan application processing, and GenAI augmentation. Each one builds on the last.
Then, the human validation layer was a non-negotiable part of the design. For any document where extraction confidence fell below threshold, a human underwriter reviewed the source before the data moved forward.
A lot of teams want to jump straight to the GenAI layer: the insights, the recommendations, the augmentation. But if the underlying document extraction is unreliable, the LLM is just adding sophisticated analysis on top of bad data. Garbage in, garbage out, as we always say (except in this case it now has a confident-sounding explanation).
Getting the foundation right FIRST is the most important thing you can do.
What I'm Watching in the Industry
The broader trend I'm seeing in the FSI industry is hyper-personalization in consumer finance specifically.
Most fintech apps today are reactive. They categorize your spending, they send you an alert after you've already overspent, they surface a monthly summary you glance at and ignore.
The experience is in desperate need of some revamping.
What's coming, and in some places, what's already here, is something qualitatively different. Think of it as the difference between a weather app that shows today's forecast and one that says: "You have a vacation in five days, a pattern of impulse spending on weekends, and $200 less in discretionary income than last month. Here's what that means and what you should do."
Wealthfront's Path recalibrates investment and savings strategies as income changes, life events shift, and market conditions evolve. It adapts to your life with a “goal-based AI” model.
Cleo and NOVA Money take a behavioral coaching approach using gamification and personalized nudges to change financial habits. People don't change behavior from data alone, they change behavior from feedback loops.
Capital One Eno and Monarch Money are doing aggregated contextual intelligence that pulls all accounts into a single view, forecasting cash flow, and surfacing insights in the moment rather than the month-end report.
What Ties This Together
Here's the pattern I keep seeing across all of these use cases.
The data is rarely the problem. The interface between data and human decision-making is usually where the problem lies.
In every case, the real work was designing an AI system that could take complex, high-dimensional information and translate it into something actionable at the right moment for the right person. That requires three things most teams underinvest in: prompt architecture, memory management, and trust design.
I think I probably said something similar last week about the healthcare industry, but financial services is ripe for AI innovation. There are some huge opportunities here for companies to make big impacts for their customers.
If you’re curious about any of these use cases or have a cool one you’ve seen recently, reply and let me know!
Also, if you’re interested in building out a use case for your financial services company, reply directly to this email or reach out to our sales team here.
Until next time,
Ryan
Now, time for this week's AI-generated image and the prompt I used to create it.
Generate an image of me as the "Monopoly man". I need to have a monocle and look like a cartoon. I should be standing atop a game board. My cane should have the Amazon Bedrock symbol on it.
