#103 - Fraud vendors have a Jev problem

“The real promise, or threat, depending on which side of the fence you’re on, is the new semi-deterministic models like Jev.”

Are fraud vendors still needed when you can build everything with AI now?

Can you build everything with AI?

These were the questions I posed to Netanel Kabala, co-founder of Duly and a PayPal and Nuvei alum, in the last TSFS pod episode.

I spent most of the conversation with Netanel on how the build versus buy discourse changed in the past year, and how it might develop going forward.

But decision models kept pulling me back, and I didn’t get to finish my thoughts on air.

So here’s my take.

What’s actually new here?

Decision models - or System One models, if you prefer - are now hitting the market. Jev launched a couple of months ago, and the large LLM providers are already moving to release their own.

In what aspects are they superior to LLMs or financial foundation models (or System Two models, if you prefer)?

Decision models are much cheaper, faster, and they produce more consistent outputs.

But hold on. Does that sound familiar?

Machine learning models have given us fast, cheap and consistent scoring for two decades. So if that’s all decision models bring to the table, what’s the big deal?

I think the more interesting answer has to do with who can now use them, and how easily.

I see three use cases here. Everyone is talking about the first one, but I believe the other two will matter much more in the coming years.

The obvious

The first use case is the most straightforward one.

When a decision costs almost nothing and comes back in milliseconds, you can afford to make many more of them. Including decisions that, on their own, aren’t worth much.

Beyond accounts and payments, you can point a decision model at the small stuff.

Feature calculation is the obvious example, especially ones that revolve around string analysis.

Some examples: advanced address normalization, bogusness (e.g., asdasd@gmail.com), name nationality, email-to-name match, and many more.

Calculating them with an LLM is too slow for real-time operations, and training an SLM is too costly. But now you can do it pretty much out of the box.

Not only that, the output is also fit for purpose. It’s structured and reliable enough to feed your models, your rules, and your team’s UI.

That’s what most of the industry is excited about right now. And I feel like we’re only scratching the surface.

But it’s still just an upgrade to what we can already do today.

Do you still need a data science team?

Here’s the thing:

To deploy your own machine learning model today, you need a team.

You need infrastructure, features, and dozens of experiments until you understand how the model should behave and how to tune it.

Between the talent, the resources and the infrastructure, that’s months of work and a 7-figure budget, at minimum.

Most teams, especially those which are more operationally focused, never get there.

Now compare that to hooking a decision model into your production stack and simply asking it to score the likelihood of fraud.

To be clear - I’m obviously being simplistic here. You’ll want to work on your question definition, and it’ll require some effort.

But that’s exactly the point - ML is now a prompt engineering problem, not a data science problem.

All it takes is an API key and a prompt, with no ML platform to erect from scratch and no hiring round.

That has the potential to truly democratize risk detection. And it’s going to be big.

Side note: Building is the easy part. Maintaining what you build is what trips most teams up - I wrote about that build trap a few weeks ago.

Do you still need data?

This is something I’m still chewing on.

The reason why a model like Jev doesn’t need your team or infrastructure to work out of the box is the same one as for Astra or Opus. They come pretrained.

A well-prompted decision model can give you pretty decent fraud scores out of the box. No historical labels, no millions of rows.

This means that you don’t need to spend time (and fraud losses) on accumulating this data.

In fact, you could go live the day you launch your product with a cheap, working fraud model.

Is it going to beat specialized, customized models? Absolutely not.

Data covers more than historical labels. There are features a decision model will never have on its own - device ID, IP piercing and behavioral biometrics, just to name a few.

And there’s consortium data, the patterns from across your vertical that you can’t see from inside your own business.

But you don’t necessarily need best-in-class at first.

And in the meantime, the time-to-market for an intermediary solution has dropped to hours instead of months.

On the opposite end of the spectrum, I wonder if a decision model trained on payments or account activity would have better performance.

I mean, it’s pretty much certain. The only question is how it’ll compare to "traditional" ML models.

But whatever the results would be, ROI calculations are going to be disrupted. Hard.

Availability beats accuracy

I believe that decision models are a real game changer for fraud detection.

Speed and cost are the upgrades everyone sees. But needing no team and no training data are likely to have an even bigger long-term impact.

Time to market collapses, and smaller teams get a middle option that lets them rely on their own capabilities for much longer before they have to buy.

Think about a single-person Shopify seller.

Tomorrow, they can deploy a pretty decent real-time fraud model that costs them almost nothing.

And that world didn’t exist just three months ago.

Are you already testing decision models in production? Hit reply, I’d love to compare notes.

In the meantime, that’s all for this week.

See you next Saturday.

P.S. If you feel like you're running out of time and need some expert advice with getting your fraud strategy on track, here's how I can help you:

Free Discovery Call - Unsure where to start or have a specific need? Schedule a 15-min call with me to assess if and how I can be of value.
​Schedule a Discovery Call Now »​

Consultation Call - Need expert advice on fraud? Meet with me for a 1-hour consultation call to gain the clarity you need. Guaranteed.
​Book a Consultation Call Now »​

Fraud Strategy Action Plan - Is your Fintech struggling with balancing fraud prevention and growth? Are you thinking about adding new fraud vendors or even offering your own fraud product? Sign up for this 2-week program to get your tailored, high-ROI fraud strategy action plan so that you know exactly what to do next. ​
Sign-up Now »

Enjoyed this and want to read more? Sign up to my newsletter to get fresh, practical insights weekly!

<