Fintech 🧠 Food - From Outsourcing to AI Sourcing

Plus; Ramp's down round is the best down round ever, and Coinbase acquires a minority stake in Circle

Welcome to Brainfood, where I take the week's biggest events and try to get under the skin of what's happening from a Fintech perspective. Join the 32,850 others by clicking below, and to the regular readers, thank you. 🙏

Hey Fintech Nerds 👋

What if we can disrupt the outsourcing industry with Generative AI? It's worth $260bn in the US alone and growing at 9.4% CAGR. That's your 📣 Rant this week: From Outsourcing to AI Sourcing.

As always, Fintech's top stories in Things to Know and a Good Read for those who want to go deeper.

PS. No newsletter next week, travel + mini vacation 

Here's this week's Brainfood in summary

📣 Rant: From Outsourcing to AI Sourcing

💸 4 Fintech Companies:

  1. Finfo - Modern Treasury meets Airwallex for Canada

  2. Communion - Building your FU Fund.

  3. Stable Money - The sensible wealth platform for India

  4. Betterfront - The VC and PE Operating System 

👀 Things to Know:

📚 Good Read:

Weekly Rant 📣

From Outsource to AI Source

Some business process tasks are best performed by humans. 

This is for three reasons:

  1. The process was impossible to change. (e.g., The process exists around dependencies like legacy IT infrastructure, regulation, or 3rd parties that cannot be changed)

  2. The process was highly variable that it was near impossible for software to automate (e.g., providing a comprehensive risk assessment of a business)

  3. Humans significantly outperformed software in the tasks for quality and cost (e.g., recognizing the context of a family photo of a young child vs. child exploitation for social media giants)

The answer was always to outsource and offshore. 

Business Process Outsourcing (BPO) is a $260bn market growing at 9.4% CAGR. Big players like Accenture, Capgemini, and Wipro help companies run and manage the complex processes around their IT infrastructure. They also significantly reduce the cost by offshoring much of that work. 

Generative AI (GenAI) presents an opportunity for software to eat BPO.

Here's my argument in summary.

  1. Legacy IT infrastructure runs the world, but it's near impossible to change.

    1. The code base is massive and hard to comprehend

    2. The code wasn't designed to be changeable or easy to use

    3. The platforms were built before modern APIs and design thinking existed

    4. This creates a risk, cost, and process lock-in problem that companies solve by outsourcing and offshoring

  2. Some tasks humans are better at than software (pre-GenAI)

    1. Humans were better at tasks involving unstructured data

    2. Humans were better at context-specific and highly variable tasks

    3. But humans have emotions like boredom or disgust

    4. Humans are also error-prone

  3. GenAI is strong at these tasks 

    1. A mental model for Generative AI is the smart intern

    2. Large Language Models (LLMs) can outperform humans in these tasks in most cases

    3. LLMs can learn domain-specific tasks if given context

  4. Entrepreneurs are off the outsourced use cases one by one

    1. Open AI believes AI can screen social media content

    2. Some parts of Governments have also started using "Gen AI Bots" for narrow use cases

    3. Customer communication and marketing in regulated industries

    4. Compliance processes in regulated industries

    5. "AI-Agents" running internal processes in regulated legacy tech industries like financial services

  5. Enterprise has been slow to adopt because 

    1. They fear a lack of privacy

    2. Possible future risks around Intellectual Property

  6. What comes now

    1. Don't overlook the distribution advantage incumbent outsourcers have?

    2. Can a startup or existing company become the canonically AI-sourcing company?

    3. Can we create certainty over IP ownership?

1. Legacy IT Infrastructure runs the world but is near impossible to change.

a) The code base is massive and hard to comprehend. The code base is millions of lines, hard to read, and in some cases, nobody alive fully understands it. Everything you care about runs this way.

Mainframe code written in the 70s runs 90% of credit card transactions, 68% of the world's IT workloads across Government departments (IRS, National weather service), 44 of the top 50 banks, and all top 10 global insurers. Mainframes are amazing because they can handle massive workloads and survive earthquakes. But the code written on them was written before the advent of modern software engineering practice. 

Consequently, very few people are alive who understand how most of these mission-critical software systems work. Equally challenging are the 90s platforms and databases at the heart of many core business processes. 

b) The code wasn't designed to be changeable or easy to use. When these platforms were implemented, they were designed to perform core functions like storing account records, moving Money, or updating customer information. It wasn't envisioned that they'd need to be changed or adapted regularly. Historically implementations happened quarterly, and once a system was running, it would rarely go down except for small upgrades. 

c) The platforms were built before modern APIs and design thinking existed. The infrastructure is now relied upon by 1000s of internal and external systems. The admin console for extracting data, or the agent platform for sending payments, was designed for an internal department, often by a software engineer. 

d) This creates a risk, cost, and process lock-in problem. The cost of changing this software can run into the billions. The analysis required is staggering if nobody really knows how the software works or what systems it depends on. All of these external systems need to connect perfectly to a new system as they always did. Those systems, too, may not be well designed. On top of that, the underlying data structure needs to be understood, converted to a new format, and then worked when the system is cut over.

Changing these systems has a low success rate, resulting in spending billions without accomplishing the task and CEOs being fired and pulled in front of the Government to explain what went wrong. As the business reacts to competition, it adds new systems on top (like a website or mobile app), compounding the issue.

So the terrible and long-winded processes that appeared around the edges of the legacy platform remain.

Internal and on-shore staff are expensive options to run a process. Those processes are mission-critical and cannot be changed but are not value-added.

So those processes get outsourced and offshored.

2. Humans were better than software at some tasks

a) Humans were better at tasks involving unstructured data. In the last decade, robotic process automation (RPA) and workflow tools have begun to take work from humans and have software do it. UIPath makes $1bn annual revenue and is valued at $8bn. They became very good at moving between multiple internal and external systems, handling variables, and beginning different processes.

The internet is full of unstructured data like websites and images, many of which don't fit neatly into a process but need to be considered as part of a process. When giving an account to a new business, understanding if they're legitimate involves structured data (their formal business registration). But it also involves images of passports, PDF documents, and even their public-facing website. Traditional software had difficulty pulling this data in and making sense of it.

b) Humans were better at context-specific and highly variable tasks. Even when software can read unstructured data and see that an image has a cat, it struggles with the context. Was that a family cat or a cat subject to abuse? Every picture is different.

Or, take our merchant onboarding example, a simple question like "is this business legitimate" is very hard to answer. For one, their website might be very helpful (like they're a big, well-run company, and the website shows as much). For a younger company, that isn't a helpful signal. How to dial up and dial down importance is context-specific. Humans innately get that

c) But humans have emotions like boredom or disgust. Imagine if you had to spend all day checking a merchant's website or if its owners appeared on a sanctions list. Then if there's something wrong you have to write a report about why. That's dull. Worse, imagine you have to look at images that could be animal harm, child exploitation, or violence. Offshore humans are still very human

d) Humans are also error-prone. Citi accidentally sent Revlon nearly $900m when it had meant to send $7.8m. 3 people reviewed the payment before someone pushed send, but the person who pushed send ticked the wrong box on the digital payment form. In their defense, the UI was confusing. But there are 1000s of similar examples around the industry, where some poor, underpaid workers erred. 

The variable and context-specific processes are outsourced and offshored. 

But what if GenAI could do these tasks?

3. GenAI is good at these things

a) A mental model for Generative AI is the smart intern. A large language model can produce basic essays or descriptions even with basic prompting. Highly variable tasks like researching an industry, understanding the context of an image, or creating a story are all possible. The first output of a GenAI model may not always be as good as an expert human, but chances are it's better than most humans. 

b) Large Language Models (LLMs) can outperform humans in these tasks in most cases. Generative AI can pass the law exam (the bar) with a passing score higher than the human average. It can pass medical licence exams as well as countless others. It can be more creative than humans in some cases. It is infinitely patient, fast, and always on. It's more like if you had an army of interns who could perform tasks incredibly quickly if given the right context. Another way to think about GenAI is as Robotic Process Automation (RPA) for highly variable processes.

Where it often fails is understanding non-public domain data. LLMs like ChatGPT and Claude are trained on text scraped from the internet, so they struggle in subjective and domain-specific tasks. To get an LLM to follow an organization's policies or processes, it first needs to understand those processes. This can be achieved with sufficient prompting but often becomes slower or less efficient than an outsourcing provider.


c) LLMs can learn domain-specific tasks if given context. Your experience using a large language model likely involves using the chat interface. The best results come as you give it more context and help it learn, but this is long-winded. However, internal or non-public domain data can be available to LLMs to improve performance. 

A simple example is storing a company's FAQs and help docs in a vector database. When asked a question like "Why did this API call an error" or "Why did this payment fail" the LLM can answer from the help docs saving customers from hours of searching. Even more complex questions like "is this marketing copy compliant with regulation and my internal processes" can be well answered. Or open questions like "What risks might appear if we use this marketing copy." 

4. Entrepreneurs are picking off use cases one by one

a) Open AI believes AI can screen social media images. Content moderation can take months and leave humans psychologically scarred. In a recent interview Open AI creator Sam Altman said he believes GPT4 can bring that closer to a few hours. He also believes it will prove to be more accurate at labeling content. 

b) Some parts of some Governments have also started using "Bots." The use of GenAI is extremely sensitive in government because of the need to be private, ethical, and secure. However, there are some instances where exceptions can be made. I heard one example where a Government department had citizens in high-stress situations, a small fraction of which were a risk of suicide. Managing the backlog of comms meant messages from citizens had a 6-week turnaround. But GenAI is used to screen for suicide risk and ensure those needles in the haystack are dealt with sooner.

c) Customer communication and marketing. The example of marketing copy from above is done by a company called HariusParthean is a financial customer education / private banker AI. Many bigger companies like Ramp, Notion, and Shopify are also beginning to use GenAI in their product. Adoption is slower in traditional enterprisesAI features like this displace the role of specialist departments, consultants, and offshore talent.

d) Compliance processes in regulated industries. Coris is an LLM for merchant risk screening and ongoing monitoring. Understanding the true industry is highly context specific. This task would have bored a human, but is something a GenAI model has no issue doing and performs well.

e) "AI-Agents" running internal processes in regulated legacy tech industries like financial services. The most domain-specific task is running an internal process on an internal system. Generative "AI agents" can displace that outsourced and offshored work with an AI agent. Typically this agent can work and use. Cascading makes "AI Agents." that run legacy internal system processes.

5. But enterprise is slow to adopt 

a) They fear a lack of privacy. The FTC is investigating OpenAI over a "lack of transparency" on its privacy and data controls. While users can "turn off" their history, preventing their input from being used as future training data. ChatGPT for business also doesn't use inputs for training data by default. 

OpenAI hasn't done a good job managing the fears of enterprises and Governments here. The mere fact that it's default opt-out instead of default opt-in is a poor posture. And the lack of clarity about where training data came from could still be a huge issue. Given this context, it's unsurprising to see Microsoft release "Private ChatGPT" as an enterprise app in Azure (although they appear to have removed this now).

b) Possible future risks around Intellectual Property. After employees at Samsung used the consumer-grade ChatGPT to write up sensitive meeting notes, businesses are concerned about what sensitive data could show up as training data elsewhere. The writer's strike in Hollywood and the New York Times blocked OpenAI, and now suing them, could the models have to pay massive fines and royalties to copyright holders? If you used Generative AI to create a new asset as an enterprise, do you own that asset, or does it require royalties to be paid?

The reality is even the folks at OpenAI, Anthropic, and Stability labs don't know exactly how the model comes up with its output. So it's near impossible to prove who should get royalties.

6. Who wins, what are the moats, and what do we need?

The GenAI moat topic is hotly debated. Some argue the models are the platform, and others point out that cloud vendors and GPU manufacturers are selling shovels in a gold rush. That's all great. But the bigger question is what happens between that and the customer.

Do the models become the front page of the internet like Google?

That's the wrong question. If the models are network typologies (like ethernet and token ring), above them will emerge a spectrum of use cases like search engines, email, the browser, and SaaS companies. If that's true, we need to zoom in on the outsource vs. AI sourcing topic to determine who wins there.

a) Don't overlook the distribution advantage incumbent outsourcers have. The most likely buyer of AI Sourcing already has a massive contract with most of the major outsourcing providers. Those companies also specialize in technology integration and outsourcing and will build GenAI tools for their clients already. I heard of one beginning to charge a day rate for those GenAI "bots." 

Getting a contract with enterprise companies is hard; it's even harder in regulated industries. However, there is the cannibalization risk these companies will face at some point. 

b) Can a startup or existing company become the canonically AI-sourcing company? The last Y-Combinator batch of startups was nearly 50% companies doing a use case in GenAI that's too context-specific for a generic model like ChatGPT or Claude. But are they so niche they become a vertical-SaaS player at best? 

There needs to be a zoomed-out version. For example, I recently spoke to a company building a "compliance officer AI." Combining the existing models with a new model trained on law and regulatory texts. Your AI compliance officer must accurately interpret the law's intent and the prose. That is a nice abstraction between the generic models and the use case focussed startups.

Today's outsources are body shops. But the key here is not to stack them high and sell them cheap. It is getting distribution and manages 1000s of different related contexts.

c) Can we create certainty over IP ownership? Given we don't know what data a model used, it would be ideal if we had some sweeping legislation that allowed GenAI to exist anyway if we see it as a net positive for society. Ideally, we would create legislation as we had for the internet with section 230 (that defines websites as not being publishers.) Although the current political atmosphere around GenAI may not make that possible.

Another idea is to have future models tag input data content with a hash. The problem is that an output from the LLM results from a calculation that touches every weight in the network. If there was a way to track which weights were touched by how much, then this could be used to pay royalties if required, ideally in an automated fashion. But that's more science fiction than fact (at least based on my research so far).

There's so much to solve here. That's why we need entrepreneurs, and we need the hype cycle. That might fade as people get fatigued. 

Every day is 1000 more startups.

But let's admit that creativity is off the charts.

I wasn't excited about GenAI until recently.

With NVIDIA single-handedly holding up the stock market and the sheer volume of creativity out there. It feels like it's getting a lot less "me too" and more "holy crap, that's cool."

If you enjoyed this, hit reply and send me the most interesting thing you've seen.


4 Fintech Companies 💸

1. Finfo - Modern Treasury meets Airwallex for Canada

Finfo provides global currency accounts, FX, and automated payables for growth businesses in Canada. It offers more competitive rates than incumbent banks on FX, local accounts to get paid, and invoice approval workflows. Users can also connect multiple bank accounts to their platform and automate processes across them. 

🤔 This is a nice UX on Currency Cloud, but that's not bad. Most companies that do what Finfo does start in the US and stay there for a long time because there's so much room to grow in their home market. The great white north of Canada gets overlooked at first. It then gets underserved because it has a concentrated banking market that's harder to launch a Fintech company in. That will change as the BaaS providers start to arrive. But the idea of building on Currency Cloud is elegant. It gets you the dashboard across multiple accounts and lets you handle multiple currencies as a wedge for businesses that transact across borders. 

2. Communion - Building your FU Fund.

Communion mixes playful marketing with a savings mobile app to engage users in savings habits. Users are guided through a 10-day savings "ritual" that requires a £1 ($1.30) deposit for 10 days to build the FU.Fund and "unlearn harmful beliefs." The app has several months' worth of daily 90-second videos. Users get a 3.66% savings rate that rises to 5.66% if they recruit up to 5 friends.

🤔 Saving is the right thing to do, yet so few of us do it. The closest thing to a "CrossFit for savings" is FIRE (Financial Independence Retire Early) which is a bit niche. But there's no daily step count tracker for savings. The founder and CEO, Dan, also founded the mobile mortgage broker Habito and saw firsthand what held users back from wealth creation. I like the idea of taking a more extreme approach to marketing and engagement. It's almost screaming at you. JUST F*CKING SAVE; KEEP DOING IT. We all need that reminder. I hope it works, and he can monetize it. The world needs this.

3. Stable Money - The sensible wealth platform for India

Stable Money combines a financial education platform with fixed-income assets like Bonds, mutual funds, and fixed-term deposits. It also features "Stable Expert," a WhatsApp bot that can help answer common questions and provide real-time information like bank savings rates.

🤔 The anti-day-trading and anti-FOMO trend is strong. Some of that is because interest rates have reversed from zero to high by recent standards. I welcome this change, but I wonder if it's an attractive enough message to attract users. Stable Money is tremendously well executed.

4. Betterfront - The VC and PE Operating System 

Betterfront helps VC and PE firms raise capital from institutional investors. It builds a data-driven funding story, manages NDAs and due diligence documents, manages investor pipeline with a CRM, and lets funds white label the platform.

🤔 It's a tough time to be fundraising right now, especially in Fintech. Track record is a critical component of due diligence, and anything that can be done to help the mid-sized or smaller funds compete with the tier 1's is a welcome addition. 

Things to know 👀

Ramp, the rising star of Fintech, has raised a new round led by existing backers Thrive and Sands Capital following their March 2022 equity raise of $200m at an $8.1bn valuation. While the company hasn't revealed revenue figures, they have said it's in the "several hundreds of millions." The CEO noted in March that their revenue grew by 4x last year.

🤔 They're a case study in being customer and product obsessed. Their blog shared that they've saved customers $600m and 8.5m hours. The average customer spends 3.5% less after adopting Ramp. That's ROI.

🤔 Growth fixes everything. Even as valuations are normalizing and nature is healing, Ramp is an incredible growth story. In a market where Stripe saw their valuation half, Adyen had a market rout, and Klarna went from ~$40bn to $8bn, a 28% haircut is a good result. 

🤔 There aren't many companies growing this fast. The nature of venture is to pour fuel on the fire of those companies that are growing to keep growing. Ramp hasn't hit profit yet, but it isn't done growing. That's why it's hiring. 

🤔 More than just spend management. Seeing Ramp as a spend management company is missing the big picture. The product suite includes corporate cards, accounts payable, working capital, and accounting automation. If you add up what they offer, it's about 10 different Fintech companies or bank departments.

🤔 Changing the landscape. Ramp is attacking the comfortable transactional profit pools of large institutions. If you're at AMEX, Capital One, and Chase Sapphire, it will be nearly impossible to compete on product experience. This will force them to compete on price and spend more developing products to reach feature parity. I wouldn't bet against any of those companies doing so; my broader point is Ramp (and their cohorts) have massively changed the game. 

Coinbase is acquiring a minority stake in Circle Internet Financial, the maker of the USDC Stablecoin. The Center consortium, which used to govern USDC, will be shut down, with Circle now becoming the sole issuer of the Stablecoin. USDC is also coming to 6 more Blockchains. Both companies will continue to share the net interest income received from Stablecoins held on their platforms. The move follows the launch of the PayPal Stablecoin PYUSD, which the CEOs say "grows the pie" of the Stablecoin market.

🤔 Regulatory clarity for Stablecoins is coming. Europe, the UK, and now Singapore has Stablecoin regulations. The US has a proposed bi-partisan bill on clarity for Stablecoins; generally, the lack of negative sounds from the regulators is deafening. That says a lot.

🤔 Stablecoins are the most at scale use case. You might not see it in your business, but I hear from founders that the volumes on "Stablecoins, the cross border payments rail" continue to increase. The on and off ramp is fading into the background, becoming another way to move dollars. Think of it like a global dollar rail, with instant liquidity for the freelancer and global SMB economy. That's why PayPal is so interested in the space.

🤔 The US Dollar Stablecoin format war? There are so many US Dollar Stablecoins. Having to convert a PYUSD for a USDC is stupid. They need to work as one to gain much broader adoption. That needs a standard that goes way beyond ERC20. Imagine wrapping the Stablecoins in something like the Visa network (vUSDC, vPYUSD). Instead of it earning yield, it would come with consumer protections and work anywhere Visa is accepted as a payment method. 

🤔 Circle The SPAC that never was. Is this a path to acquisition by Coinbase? Circle had announced, then delayed a listing via a SPAC numerous times. It's clear they were looking for some sort of exit, which makes me wonder is the publicly listed COIN equity is partly that exit for staff or investors? Also, wouldn't USDC fit into Coinbase like a glove? 

Quick hit

🥊 India has announced a UPI for CreditThe idea is "frictionless" personal loans, credit cards, and micro SMB lending. The service will work with both UPI and the digital identity platform Aadhaar. The project is a pilot for now.

🤔 Just as UPI has become a default case study, if "UPI for Credit" goes mainstream, that would be a game-changer. One feature would allow digital access to land records. Meaning you could imagine an almost 100% digital mortgage process.

Good Reads 📚

AI is more than a tape recorder that regurgitates other information. It remixes it in new ways to create novel ideas. AI can pass every test we've created for human "genuine creativity," even when academics make further tests, AI sometimes wins. AI was better at structure and feasibility in domains like story writing or product development, while humans had more raw novel creativity. The best outcomes were when humans and AI collaborated. The caveat, however, is that the most creative humans were helped the least by AI.

The takeaways: 

  1. Recombining ideas is creativity

  2. AI as a co-pilot for creativity will produce better outcomes

  3. Giving AI constraints and examples will produce better outputs

🤔 Fintech use cases use AI to do schlep work. We've mostly hired AI Agents as co-pilots for engineering, data entry, or manual processes, like summarizing if a merchant is a high or low risk based on their website or pulling together text data like SARs or chargeback dispute summaries.

🤔 But Fintech can be creative. As I wrote last week, the hardest part about financial services is building genuinely new products and the level of problem-solving in tasks like risk, fraud, or compliance. 

🤔 GenAI in Fintech has been mostly hidden, but it won't stay that way. We can use it to help solve some of our most gnarly problems if we use it as our thinking partner.

That's all, folks. 👋

Remember, if you're enjoying this content, please do tell all your fintech friends to check it out and hit the subscribe button :)

Disclosures: (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 (3) Any companies mentioned in Rants are top of mind and used for illustrative purposes only. (4) I'm not an expert at everything you read here. Some of it is me thinking out loud and learning as I go; please don't take it as gospel—strong opinions, weakly held.