Why General AI Models Fail at Financial Fraud | Suddu Seshadri artwork

Why General AI Models Fail at Financial Fraud | Suddu Seshadri

MTS

October 17, 2026

Suddu Seshadri, Head of AI at Plaid, joins MTS to discuss how Plaid is building purpose-built transformer-based foundation models on its financial transaction network to catch fraud, predict credit repayment risk, and power personal AI financial agents. Turn ideas into software people love.

0:00

Suddu Seshadri

If you think about general models, they can actually really do an incredible job in explaining why a transaction is risky. But to be able to predict whether a transaction is at risk or not, requires for you to be able to train across historical data. But what we will need to do is still train them on historical data, and actually optimize and tune it for specific use cases. And we've been able to see some pretty incredible results, like fraud model, foundation model has been able to catch 40% more fraud than what we've been able to do historically.

0:31

SPEAKER_2

Hello, everyone, and welcome back to MTS. Today, we are joined live by Suddu Seshadri, who is the head of AI at Plaid. And so welcome to MTS.

0:40

Suddu Seshadri

Thank you for having me.

0:41

SPEAKER_2

Thank you for... I'm excited, excited to have you join. I know you guys recently announced some pretty exciting stuff with Plaid, especially when it comes to your own AI models. Tell me a little bit more about that.

0:52

Suddu Seshadri

Yeah, so when consumers and users think of Plaid, they're usually aware of how Plaid connects from their bank accounts to applications. But underneath that layer is a pretty rich network of users, identities and transactions. What we have been able to do is use that connection and network to be able to build foundation models which is purpose-built to solve financial problems.

So customer now can solve problems tied to whether a transaction is risky, or whether there's repayment risk tied to a credit or a loan, or is a fraud transaction risky. So what we have done is take those transactions and events to be able to model that out, which is purpose-built for financial AI.

1:33

SPEAKER_2

Purpose-built for financial AI. What does that mean?

1:36

Suddu Seshadri

Yeah. So if you think about it, as you look at the connections and the richness of the network, you have customers who are rebuilding a ton of problems which is tied to transaction risk or fraud risk. What you don't want to do is have customers rebuild all of that from scratch, but at the same time can use the diversity and the depth of Plaza Network to solve these questions which is tied to whether you need to be able to make a decision on a loan, whether you need to be able to allow a transaction to go through, or whether there is a risk of a fraudulent transaction here itself. So what you need to do is to be able to solve for both rigor and speed, and especially on speed, especially because fraud is moving so fast with AI. At the same time, people are getting much more access to financial data now, to be able to make their own credit and lending decisions. So what you want to do is solve for all of these use cases.

At the same time, build an infrastructure layer which enables you to solve all of these use cases. And that's what we mean by purpose-built for financial AI.

2:38

SPEAKER_2

You said fraud is moving really fast with AI.

2:41

Suddu Seshadri

Yeah.

2:41

SPEAKER_2

What's happening right now?

2:42

Suddu Seshadri

Yeah. So with AI, as you can see across whether it's identity, synthetic or first-party fraud, all of these are independent use cases across fraud. And with the AI, it's become easier for you to be able to replicate all of these scenarios across the board.

But at the same time, when you look at how Plaid is connected across 9,000 applications and 12,000 financial institutions with millions of users, especially about one to two American citizens are now on Plaid, the good actors have to connect with Plaid as well as the bad actors have to go through the same network and connections. So what we have been able to do is use that depth of that particular graph to sequence it and model it in such a way that we can identify anomalies and detect those anomalies to be able to identify fraud and prevent them before a transaction goes through itself. So fraud is really moving fast and as you can sequence events and activities to be able to understand where there is these anomalies happening, you can actually catch fraud faster too. So it's a continuous cat and mouse game which you're playing. I think it's only getting faster and faster as we think about it.

3:48

SPEAKER_2

But in theory, why can't general models work for that just as well?

3:52

Suddu Seshadri

Yeah, it's a great question. So if you think about general models, they can actually really do an incredible job in explaining why a transaction is risky.

Read the other 11 minutes of this episode here.

$15 gets you this episode and 99 more. Credits never expire. After checkout, one link brings you back here to read it.

Not what you expected? Full refund within 14 days.

Thousands of transcripts fetched by people building searchable podcast archives

Fetch transcripts with code

The same credits work from a script or an AI agent. The demo key returns a sample episode in full, no card needed:

request
curl -H "x-api-key: pt_demo" \
  https://spoken.md/transcripts/1000651996090

Markdown with the speakers named, for your notes, your knowledge base, or anything that makes HTTP calls.

From $0.10 per transcript. No subscription. Credits never expire. Prices exclude VAT, added at checkout for EU customers. Not what you expected? Email us within 14 days with 20 or fewer credits used and we refund the pack in full.

Using your own key:

request
curl -H "x-api-key: YOUR_KEY" \
  https://spoken.md/transcripts/1000796807518