- Fintech Brainfood
- Posts
- The rise of AI Agents in Financial Services
The rise of AI Agents in Financial Services
While 99% of companies are talking a good game, a small few are winning. Ignore the hyperbole; there's a platform shift happening, but not with co-pilots.
Welcome to Fintech Brainfood, the weekly deep dive into Fintech news, events, and analysis. You can subscribe by hitting the button below, and you can get in touch by hitting reply to the email (or subscribing then replying)
Hey Fintech Nerds đź‘‹
Elizabeth Warren says Fintech companies like Chime, Venmo, and Stripe should be banned from using the FDIC logo. They also request clear rules for managing third parties. Sounds bad. But this will be a good thing in time. Clarity is useful. (đź‘€)
Klarna got another press bump from using AI to “remove Salesforce and Workday.” Their pre-IPO, year-of-efficiency tour continues, but there are lessons here (👀)
Then Open AI released o1. It makes agents useful. It can error-correct, think step by step, and produce much more useful outputs, like solving crossword puzzles, math equations, or writing software.
The agent era is upon us. How will that impact financial services? That’s your Rant this week.
Here's this week's Brainfood in summary
đź“Ł Rant: The State of AI in finance & what it takes to win
đź’¸ 4 Fintech Companies:
đź‘€ Things to Know:
đź“š Good Read: The case for becoming a bank
If your email client clips some of this newsletter click below to see the rest
Weekly Rant đź“Ł
The rise of AI Agents in Financial Services
AI is now transforming finance. While 99% of companies are talking a good game, the very best are reducing costs, building new products, and changing how they work daily.
When Klarna announced they'd replaced half of their customer support with AI, the world of finance was captivated. Now, they've announced that Salesforce and Workday will also be replaced. Every incumbent CEO wants that, and every cynic wonders how inefficient they were before.
They're winning at AI because of the consistent, multi-year focus across the organization; they're using AI internally as a tool in their product and deploying agents.
Finance is an industry that relies on human judgment for risk, compliance, and “fairness.” Until now, software has not been able to replicate with high enough accuracy.
Now with Open AI’s o1, agents will become much more accessible. Every white collar job could hire 10 agents to help them.
We’re skipping past co-pilots straight to agents.
This piece explores how to think about GenAI in finance and how to win.
Why it's important to ignore the memes about Generative AI (good and bad)
Why the “agent” concept is imperfect but helpful marketing
Why Fintech is ideally placed to win at AI
Then, we can discuss the 4 ways to win at Generative AI in a vertical like finance.
Own the workflow of a growth customer and add AI or AI agents (e.g., Stripe or Ramp customers, for consumer Klarna customers)
AI Agents on top of incumbents as a wedge to displace legacy systems and own the workflow (e.g., Compliance automation like Casca* or Axle)
Own the workflow as an incumbent and add co-pilots (i.e., a system of record at, e.g., Salesforce, Microsoft, Oracle, etc.).
Innovate for the incumbents with AI Agents & growth companies and partner with or get acquired by them. (e.g. FIS, Fiserv, or the banks)
Ignore the memes.
Many people feel like AI is the emperor's new clothes. A large percentage of people had a few bad experiences with ChatGPT and decided the tech was a glorified party trick. The evidence disagrees.
I remain confused by the "GenAI is a dud" arguments.
We have controlled experiments showing 20%-40% gains in real work using GenAI. Adoption rates are the fastest in history. There is value.
I think what folks may mean is that "companies haven't captured the value created yet."
— Ethan Mollick (@emollick)
5:49 PM • Sep 13, 2024
Unquestionably, most legacy companies that are trying to invest in AI won’t get the benefit.
Yet some companies are meaningfully transforming themselves, adopting AI internally, in their products, or creating agents for their customers. The rest are quietly benefitting from these new services.
The mainstream discourse always involves "responsible AI" or ethics in AI. Bloomberg adverts for Workday feature aging rock stars talking about this stuff. Governments and regulators are getting together to discuss regulation and the existential crisis.
As I discussed a couple of weeks ago, regulation with AI can be managed; we know how to do this.
Another big meme is hallucination, AKA "creativity." This makes the models nondeterministic. You don't get a set answer like you would from a calculator. This isn't very clear for newbies but second nature for statisticians.
Complex problems are multivariate. Generative AI models like LLMs or image generators can help you imagine new futures, be creative, and problem-solve.
These memes don't communicate complex reality or help you figure out what to do with technology. Even if the memes and Thotboi Twitter threads are annoying, something genuinely paradigm-shifting is happening.
And Fintech is ideally placed to win at this new chapter in AI.
People seem to “get” the idea of an agent as a buyer and a GTM.
Three models to package AI
AI-as-product. Using AI and LLMs to make UI's more magical and conversational.
AI-as-copilot. Sitting outside a workflow and doing helpful things like adding data.
AI-as-agent. Displacing the previously human-only tasks and manual reviews.
It's this third one that I think the mass market buyer gets. AI as a product disappears into the background and never gets appreciated (like all great design). Co-pilots are awkward and a little janky. AI agents for software engineering, demand gen, compliance, and everything else? That's where we're headed.
The Agent-ification of AI is coming to Fintech. Here are some other Thotboi thought starters for your next brainstorming session.
Every white-collar job title could be an "AI agent." Examples include fraud analysts, compliance analysts, software engineers, designers, loan officers, and customer service agents.
Every back office process could be rebuilt to be AI-first instead of SaaS-first. Why use a workflow tool or software like a CRM when AI can help you build things just the way you want in your customer service team. Today, most start-ups are held up by Google Sheets, Airtable, and Zapier automation before implementing some SaaS tools. AI is like the Google Sheet that works just the way you need, but at scale.
Therefore: Every UI could be an AI Agent or Agentic. Wherever you see a UI, some organization staff member usually uses it. That UI could perform a better task than AI. For example, instead of using Zendesk or Salesforce for customer support, build AI agents to handle 70% of queries and have the AI engineer build a dashboard for exceptions.
What is an AI agent?
Software engineers often use the term "Marketichture" (marketing + architecture) to describe things that sound technical but are far from correct. As a practical reality, the term "agent" has no direct technical definition, but it's a helpful simplification. "Agent" is the perfect example of Marketichture.
Conceptually: The idea of an AI agent is software with the agency to make decisions, where humans can review but don't have to approve before an action is taken.
Fintech is ideally placed to win at AI.
Financial services have ridden every macro tech wave.
Most financial services now use the default cloud and mobile.
Most middle and back office tasks are increasingly on SaaS providers.
AI tooling is moving faster and becoming affordable.
Financial services are built on unstructured data, legacy systems, and complex contracts.
Finance relies on human judgment. Headcount is often the highest cost (just look at how much sponsor banks have had to hire in compliance)
Humans can only work at human speed. AI can work at machine speed.
AI that can problem-solve like a human is a breakthrough. AI can now solve the kind of human judgment problems financial services are filled with today.
The ability to perform complex tasks from a broad range of unstructured data opens the aperture of use cases. In a company every human job title could now have a meaningful percentage of its work completed by AI agents.
The internal use case is here. Whether it's risk assessing a business, a management team, or having a feel for compliance when building new products, a lot of assumed knowledge is hard won when you deliver financial services products to market.
Tiny models (like Apple Intelligence) that run on a mobile device can be private, operate on personal data, and begin to act as personal agents. It's still a little janky for consumers, but that is starting to change. The consumer use case is a matter of time.
And yet.
Most people still don't have a daily-driver LLM or set of AI tools.
The future is here its just not evenly distributed, which creates a perception gap.
Some Fintech growth companies live in the future, which are a great place to show who's winning at AI in Fintech and how to win.
1. Owning the workflow of growth companies is the best position to add or benefit from AI and Agents.
The biggest Fintech winner from AI is Stripe. Of the top 50 GenAI companies, 82% are Stripe payments or billing customers. Stripe has always indexed to the next growth customer, but they're also marketing masters.
Having the CEO tweet something like the below or having the Nvidia CEO show up at your customer conference is a unique and timely flex. Whatever the Silicon Valley wave is, Stripe is riding it.
a16z made this handy list of the top 50 AI Gen AI web products: a16z.com/100-gen-ai-app…
We checked, and turns out that 82% use @stripe.
We've been building a bunch of functionality that's useful for AI products, including usage-based billing (stripe.com/billing/usage-…) to handle… x.com/i/web/status/1…
— Patrick Collison (@patrickc)
11:26 PM • Sep 10, 2024
Stripe owns the payment and billing workflow and is naturally placed to use it behind the scenes (although they're being somewhat coy about the direct uses).
Another example is Ramp, whose CEO, Eric, has been on every VC and tech podcast lately, discussing how they use AI internally and for customers. Ramp is using AI as a product when it adds features like AI to understand what a user is trying to do and help them navigate to the right screen.
Whether it's their internal sales database bot Toby or the AI that can help you navigate their UI, they're becoming default-AI or AI-first. Not just in marketing but in tooling and going to market.
Ramp owns the spend management, CFO, and finance workflows. It's winning at the growth customer segment. Its AI story keeps it there, and it can build AI as a product. Users can use contextual search, and Ramp’s AI will drive the UI on their behalf. They’ve also been public about the use of AI agents like Cognition AI in their software engineering and internal use cases.
Klarna directly owns the shopping and checkout workflow. Their core business model of BNPL and shopping ecosystems has consistently driven growth. Using AI agents in customer support, they can continue to re-engage shoppers and manage issues like delivery or returns elegantly
2. Innovation on top of Incumbents with Agents
Loan agents sit on top of Loan Origination systems. Casca* offers a loan officer agent that can manage communications with customers, collect documents, and directly drive underlying legacy systems like a loan officer would.
Compliance agents deal with manual work. Greenlight and Parcha can investigate business customers flagged for manual review by a KYB process and approve or decline their onboarding. Axle can respond to transaction monitoring or sanctions alerts and approve/decline them. All of these are examples of AI-as-agents.
Often, they're sitting on top of a legacy system like NiCE Actimize or a loan origination system (LOS). Their natural buyers are incumbents who have legacy systems or very early-stage companies that don't have the staff breadth to handle the alerts or volume of manual work.
By leading with agents as the thing they sell, it's easy for the buyer to comprehend what the product does and how it fits their workflows today.
This is true for marketing, CRM, sales, and countless other parts of a company's value chain. Ramp and Klarna noted that they use GenAI for their marketing images. Every SaaS tool has an AI agent that can work with or displace that SaaS tool.
The interesting question for these companies is: can they can become the daily driver? We've seen Klarna displaced Salesforce and Workday. Will we see more stories like that? Will that become normal? Or was Klarna just not that big of a Salesforce user in the first place?
3. Incumbent workflows adding AI Agents
I haven't seen notable examples of this in the major payment companies, core banking providers, or legacy platforms that financial institutions use. Perhaps the AI agents from the SAP, Oracle, and Microsoft crowd have more opportunities.
Salesforce is actively marketing its AI. The market seems to think Salesforce is more threatened than it has opportunities. Their "Einstein GPT" creates personalized content for every employee. But it's ultimately ChatGPT embedded in Salesforce. The model here is AI-as-co-pilot.
Microsoft has been aggressive at embedding AI into its productivity suite, perhaps even more so than Google. Most back-office workflows at legacy businesses still exist on a Microsoft stack. Microsoft Excel and legacy Mainframes hold up the global financial system. Could they turn these into agents? This is also packaging AI-as-co-pilot.
The incumbents seem to have speed-run launching the co-pilot model to defend their market position, but ironically, they just wrapped ChatGPT and have not meaningfully innovated.
Currently, some Microsoft data integrations directly into Excel could be very interesting. This gives the potential for a Powerpoint to auto-update with live data, but this isn't what's being sold or what the incumbents seem to know how to buy.
They do know how to buy agents.
4. Incumbent workflows buying or partnering with AI agents.
Incumbents are no strangers to channel partners, integration partners, and the marketplace. Giants like Salesforce and SAP have businesses specializing in implementation and their own app stores.
An AI agent could sit on top of your existing provider. Today, they already do, but it's more of an early-Plaid-like agent, directly interfacing with back-end systems and driving them as a human would, handling human comms, emails, and documents. This could be done more formally with a simple re-seller or partnership arrangement.
An AI agent could be tightly coupled with incumbent core systems. What if agents to operate a system, come packaged with the system itself. It feels like the obvious "value added service" for these companies to sell directly. Legacy providers have a habit of acquiring companies that do this sort of thing eventually.
Winning the AI Era
Get the basics right, skip co-pilots, and package AI to solve tasks requiring human judgement
How to read this diagram: The fundamentals, compliance, and use of AI as a tool are prerequisites to capturing value from the sale of AI. I’m dubious about the long-term value of enterprise co-pilots; they’re less powerful than baking AI into the product and less obvious to users than agents.
Data Fundementals. You have to get the basics right in data. You need a team that understands how AI works from the top down and is using it daily.
What Shopify, Klarna, Ramp, and Stripe have in common is that despite being late-stage, they're founder-led. And I'm not saying this in some cringe "founder-mode" hyperbole. I'm saying it in the Tobi from Shopify tinkers with AI because it's fascinating. These founders and companies are using the tool to grow their businesses.
Compliance (and quality). Financial services compliance. If you communicate with customers, you must immediately worry about UDAAP and customer fairness, disclosures, etc.
Quality. There's nothing worse than an AI chatbot that doesn't work. The balance between low-cost open-source models and using RAG and fine-tuning is critical to creating and packaging any experience for users. Raw LLM models are an expert tool; productizing them requires work.
AI as a tool. You cannot build with AI if you don't use it every day. The number of CEOs on TV and in the media who don't use AI daily is so obvious. There's a massive difference between companies and cultures that use it to do their jobs and those that just buy it as a SaaS tool or use ChatGPT to occasionally write marketing copy.
I've included AI as a co-pilot for completeness. There's good and bad here. The good is using AI as your thinking buddy, your sparring partner—not always outsourcing whole tasks, but working with it, dancing with it. The bad version is wrapping ChatGPT into your product and saying "Einstein GPT." It looks odd.
AI as a product is the artful inclusion of AI in the UI. This can be much more powerful than a co-pilot because it doesn’t require a chat or prompt to complete. No new skill is required. Save agents for conversation, and hide AI in the UI where there’s a better shelling point for task interaction.
AI agents will be where the real unlock is. They can perform the tasks previously reserved for humans with creativity, problem-solving, and unstructured data. Building AI agents requires all of the fundamentals, quality, skills, and AI product experience.
They're the hardest to build, and today, they naturally get applied to existing processes. However, things could get spicy when they start getting creative as engineers, compliance officers, or loan officers. (Within legal and good product limits of course)
Summary
Ignore the memes, play with AI, and put the foundations in place to take advantage of it immediately. This is a generational shift, and like every hype cycle, the first to take advantage will be the small and growing companies. Your competitive advantage will ultimately come down to how far up the AI-era hierarchy you can climb.
Are you just wrapping up ChatGPT, or are you doing the hard data engineering work to rethink your entire value chain?
We're well past the age of shit chatbots.
If your experience of AI is not being able to get ChatGPT to do something the way you intended, you're doing it wrong.
The lesson here is that winning in the AI era is about shifting your default. Shifting your thought pattern. Shifting to every UI could be an AI "agent." Every white-collar job could be an agent.
Not because this is the perfect technical description.
Because, like all marketing, it's a helpful metaphor.
Play more.
Let go and dance with the AI.
Because it's only by engaging with it you'll figure out where the true value is.
Part of me groans every time I see a Klarna press release because, no doubt, it hides a lot of complexity, like being a great cover for headcount reduction and vendor consolidation. You also have to see why it's working.
They're aggressively engaging with the technology, and showing you how they do it.
Stop what you're doing right now, open ChatGPT o1, Claude or Replit, and try creating a prototype Trello board app or UI. Don't quit until you've got something working.
Or, keep trying to get to inbox zero.
Your call.
ST.
4 Fintech Companies đź’¸
1. Agree - Docusign + Stripe Billing + GenAI
Agree helps users create, send, and sign digital agreements and contracts. It bakes in invoicing and payments to the agreement platform. Users can create, collaborate, and negotiate an agreement between multiple parties. The AI can clarify documents in plain English and allows users to ask questions about the contract. Then, the platform handles invoicing, payments, and payment collection.
🧠This is a clever business model. Give the agreement signing away for free vs. charging like Docusign, but monetize the following payments, invoicing, and billing. In the process, you also reconcile any agreement against the billing. A contract is, by nature, unstructured data, but if you go upstream to it, you make it the source of truth for all invoices and billing events. Smart.
🧠It also combines two megatrends: AI-driven Lawtech and AI-driven finance workflows. Both of these require a lot of unstructured data and expertise to solve issues. I've seen countless companies doing one side of this problem but not the other. Very, very cool.
2. Flex - The online checkout for HSA/FSA
Flex allows wellness companies, opticians, gyms, and mental health support services to allow their customers to pay with their HSA/FSA cards. This has driven a 20% increase in the average order values (AOV) and creates a simple user experience to spend the HSA/FSA dollars.
🧠This feels like a win/win. Consumers save 30 to 40% when they use their HSA/FSA, which gives them more spending power; what might have been a luxury health service is much more within reach if it's 40% cheaper. Making that as slick as any e-commerce experience for the growing market of wellness companies feels directionally spot on. Growth market, moving online, remove friction. Solid.
3. Alaan - The UAE's "#1" spend management card
Alaan offers corporate cards, spend management, payment automation, and 2% cashback on all transactions. It is adept at identifying available sales tax refunds and simplifying claims. They claim to save 16 hours per month per finance team and provide finance teams with a dashboard, controls, and insights into spending analytics.
🧠UAE and the MENA are a growing tech hub poorly served by traditional banks. This is a Y-com-backed start-up that will expand across the region. Whenever I talk to founders, they see the UAE as their wedge market, a little like getting traction in Silicon Valley before going wider. This is especially important when CFOs are your clients, who tend to value relationships and don't often attend conferences. If they're in the region right now, they're probably in the UAE.
4. Valyx - Finance ops for India's fast-growing businesses
Valyx automates invoicing, collections, and reconciliation for growing businesses in India. It helps build simple routines, waits for steps, chases emails for payment, and provides a cash flow dashboard. The service automagically reconciles payments against accounts.
🧠A great business is one whose primary competitor is companies managing internal comms on Whatsapp. Whatever workflow lives in Slack, Whatsapp, or elsewhere could be automated. How long until someone builds a Slack integration to identify tasks the team is doing that AI could automate? 🤔
Things to know đź‘€
From Jason Mikula’s LinkedIn: “In a letter sent yesterday, Democratic Senators Elizabeth Warren and Chris Van Hollen expressed concerns about bank/fintech partnerships, specifically naming companies like Stripe, Chime, Venmo, and, yes, Synapse.”
The Senators describe the current model as a threat to safety and soundness and request clear rules for the management of 3rd parties. Regulators have this authority under the Bank Service Company Act (BSCA).
🧠Almost every Fintech company (and I imagine most banks) would welcome clarity here. The current storm clouds make it hard to do business, and a blizzard of enforcement actions isn’t solving the issue.
🧠Next week we have the results of the RFI on 3rd parties. Everyone in Fintech is waiting to see what comes out of that because we’re hoping for some clarity and possible engagement. My guess is it will rule make for direct oversight of 3rd parties under the BSCA, but lets see.
🧠Often Fintech companies receive feedback from regulators second hand, after remediation has been agreed between the bank and regulator. This means Fintech companies have no direct relationship with regulators or insights on the concern for good faith actors, making compliance harder.
🧠We need more direct engagement between Fintech companies and regulators. I’m involved in work with the Association for Innovation in Regulation to organize roundtables to build this dialogue. Direct supervision is inevitable, so lets start the conversation about the real issues and how we solve them.
Klarna has announced it is shutting down Salesforce and will soon do the same with Workday. This is part of a wider initiative to shut down and consolidate SaaS providers. This follows a headcount reduction of over 50% in customer service and halving its overall headcount.
🧠Klarna is a case study in becoming AI-first. They were an early partner for Open AI and have been aggressive in the internal adoption of AI tools as a default.
🧠Focussed, intentional, and long-term effort is paying off. The CTO, CEO and entire team have leaned into to AI and stayed consistent with it, not just for the press, but with execution.
🧠The leverage is real. The companies doing best with AI are those using it as their operating system on a daily basis. It’s one thing to use ChatGPT to help with your day job; it’s another to rewrite your internal stack with AI.
🧠It bothers me how perfect this PR is. When putting slides together to explain AI to others, Klarna is now one of the best reference points because they’re telling so much of the story in the lead up to their IPO
Good Reads đź“š
I really enjoyed this breakdown of the tradeoffs between being a bank vs a Neobank. The core argument is that perhaps now the business case has flipped. At one point most Fintech companies would perform jujitsu not to be a bank.
But others are winning.
Tweets of the week đź•Š
7/ While stablecoins have historically been associated as a bridge to crypto-assets, when surveyed about their primary goals for using stablecoins, 47% of participants said to get dollar-access, 43% said to get better currency conversion rates, and 32% said to send money… x.com/i/web/status/1…
— Cuy Sheffield (@cuysheffield)
12:56 PM • Sep 12, 2024
That's all, folks. đź‘‹
Remember, if you're enjoying this content, please do tell all your fintech friends to check it out and hit the subscribe button :)
(1) All content and views expressed here are the authors' personal opinions and do not reflect the views of any of their employers or employees.
(2) All companies or assets mentioned by the author in which the author has a personal and/or financial interest are denoted with a *. None of the above constitutes investment advice, and you should seek independent advice before making any investment decisions.
(3) Any companies mentioned are top of mind and used for illustrative purposes only.
(4) A team of researchers has not rigorously fact-checked this. Please don't take it as gospel—strong opinions weakly held
(5) Citations may be missing, and I've done my best to cite, but I will always aim to update and correct the live version where possible. If I cited you and got the referencing wrong, please reach out