Risk and Compliance Operations

Build, buy, or both: how Circle splits its fincrime AI stack

Published
November 17, 2026
Read Time
8
mins
Gal Perelman
Gal Perelman
Product Marketing Lead, Unit21
Subscribe to stay informed
Table of contents

Every team putting AI into fraud and AML hits the same fork: build it in house, buy it, or run some mix of the two. The answer isn't the same for every team, and it usually isn't the same for every capability inside one program.

‍

At The AI Files, Unit21's virtual summit on AI in fraud and AML, Unit21 CPO Kunal Datta asked Siddharth (Sid) Talreja, Senior Director of Compliance Technology and Risk Management at Circle, how Circle draws that line. Sid's group owns all in-house development and vendor integrations for Circle's risk functions. His short answer: build where your needs are unique, buy where they aren't, and remember that writing the code is the cheapest part.

‍

How Circle splits its stack

Circle's financial crime stack is deliberately mixed:

  • Detection is built. An in-house tool works through Circle's transactions and context and generates alerts.
  • Case management is bought. Those alerts flow into Unit21, which Sid called Circle's centralized case management workflow engine.
  • Sanctions screening is bought. Screening and adverse media alerts also land in Unit21.
  • KYC is built. Circle's onboarding and due diligence workflows are nuanced enough that no off-the-shelf tool fits.

‍

Over the past 18 months, the team has also built in-house tools to make reviews more efficient, such as context stitching, while staying open to what the market offers.

‍

The build vs. buy framework

Circle runs a formal vendor selection framework. The factors include the feature set, time to market, ease of integration and pricing. One of the heaviest is engineering dependency. Engineering time is shared between building Circle's core products and powering the risk function, so how much the team can build depends on where the company is at the time.

‍

The deciding question is usually uniqueness. When a need is specific to Circle, the team leans toward building. When it isn't, buying wins. Case management is the clearest example: a workflow engine isn't unique to Circle, and as long as a vendor offers the features and configurability the team needs, there's no reason to rebuild something that already exists.

‍

Code is cheap. Enterprise grade isn't.

Kunal pushed back: with AI, anyone can build anything. If you could build a workflow engine in a day, why buy one at all?

‍

Sid agreed that code has become cheap. Work that once took weeks can now take a day. What hasn't become cheap is everything around the code that enterprise systems provide: audit trails, data storage, uptime and guardrails. His test is simple. If Circle's screening and KYC workflows depend on a homegrown engine and it breaks at 2 AM, who gets the call? AI-assisted coding hasn't matured to the point where that question has an easy answer, and for compliance tooling, it matters a lot.

‍

Kunal shared that Tyler Allen, Unit21's CEO, once asked him whether he would build his own Slack. He probably could, and it wouldn't take long. But the software is the tip of the iceberg. Under it sit the processes, maintenance, ownership and engineers on call with laptops on weekends, keeping reliable workflows running. For AML tooling, that layer is where things get serious.

‍

The data is the moat

Circle does have one system it has built and kept growing for more than a decade: an internal source of record that started as a single screen for analysts. It now connects KYC, customer care and other services, and is becoming a 360-degree view of Circle's customers.

‍

Sid wouldn't replace it, and the reason says a lot about where AI is heading. In a world where small teams can quickly build apps for their own needs, the data signals are the moat, not the interface. Rather than overbuilding the central platform, Circle gives teams governed access to those signals so finance, marketing, customer care and compliance can build what they need on top, with access controls and guardrails in place.

‍

Kunal framed it as a split between single-player and multiplayer tools. As agents become the main interface for a lot of work, a personal tool that pulls in data signals makes sense to build yourself. Shared, serious workflows like case management, heavy analytics and anything that has to run in real time make more sense to buy.

‍

Watch the full conversation with Circle
Sid Talreja and Kunal Datta on what to build, what to buy and what changes when analysts can build their own tools.
Watch on demand

‍

When analysts become builders

Circle has given people across the company the tools to build their own solutions. It took motivation at first. There was real skepticism, including fear that AI access meant job losses, but that has eased as people saw what they could do.

‍

Sid shared one example. Transaction monitoring reviews at Circle meant working across four or five separate systems at once. Sid said that problem has existed on every compliance team he's been part of: nobody has ever had everything in one place. One analyst built an app that connects to those systems and lays out the information in the order of Circle's procedures, step one, step two, step three, so everything needed to decide an alert is on one screen. The team is now exploring a skill on top of it that analyzes the transactions and recommends a decision, which will go through validation and quality assurance first.

‍

Where a prototype has to become production

That success raises the next question. If every analyst relies on the app and one Monday morning it doesn't work, the team can't depend on the one person who built it. So Circle is building a path with its tech ops team for apps that sit on a critical path to get enterprise-grade infrastructure and support.

‍

Sid's threshold is about who uses it and for what. A marketer exploring content ideas or a customer care analyst checking complaint sentiment can keep a personal app. Once a tool expands to a team and a critical function like compliance reviews, it needs production-grade support.

‍

Kunal noted the catch: that brings back the same engineering bottleneck. Unit21 saw it in its own experiment letting non-engineers push code. Features built by people outside engineering looked great locally, but moving them to production meant engineers reviewing AI-generated code. Sid's view is that it's still a big win. By the time a tool reaches engineering, the builder has done 60 to 70% of the work, and the old problem of translating compliance requirements for engineers mostly disappears.

‍

Who owns an AI agent you bought?

Kunal's last big question: when you buy an AI product rather than build one, where does ownership sit?

‍

On the risk side, Sid said, accountability doesn't move. If an AI agent does customer due diligence, the head of KYC is still accountable for it being done properly. What changes is the role of the second line. Because a vendor's AI isn't built by Circle's own engineers, Sid expects second-line teams to spend much more time validating what the institution buys, through pilots and testing, before getting comfortable with it.

‍

Vendors can help. Sid said Circle can test Unit21's AI agents thoroughly itself, and that Unit21's independent validation of those agents adds a second layer of assurance. That double check helps the institution get comfortable with ownership, even though ownership itself doesn't change. For more on what that validation looks like in practice, see how Rippling and Zepz prove quality every day and what examiners will ask about your AI.

‍

Build again vs. never build again

Kunal closed with a lightning question: what would Sid build again in a heartbeat, and what would he never build again?

  • Build again: a customer 360. A single view of customer information, alerts, transactions, screening results, customer care contacts, limit requests and entitlements. In compliance, signals can come from anywhere, so the holistic picture matters, and Sid said it has never been easier to build than it is today. It's a game of data connections.
  • Never build: sanctions screening. Plenty of vendors do it well. Circle adds its own layer on top to help with false positives, but building, validating and owning screening itself is too much risk compared with integrating a proven tool.

‍

The pattern holds across Circle's stack: build where the value is unique to you and the data is yours, buy where the capability is common and reliability matters most. For more on choosing tools that teams can shape to their own needs, read why configurable AI is the only AML AI compliance teams will trust.

‍

Frequently asked questions

Should compliance teams build or buy AML software?

Usually both. Circle builds where its needs are unique, such as detection and KYC workflows, and buys where the capability is common and needs enterprise-grade reliability, such as case management and sanctions screening.

‍

If AI makes code cheap, why buy software at all?

Because the code is the tip of the iceberg. Enterprise systems also need audit trails, data storage, uptime, support and someone to call when something breaks at 2 AM. Those are still expensive to build and own.

‍

When does an analyst-built AI tool need to become an enterprise application?

When it moves from one person's use to a team and a critical function such as compliance reviews. At that point it needs engineering support, accuracy checks and production infrastructure, because the team cannot depend on the one person who built it.

‍

Who owns an AI agent bought from a vendor?

Accountability stays with the institution, for example the head of KYC for a due diligence agent. Second-line validation of vendor AI becomes more important, and independent validation by the vendor adds a second layer of assurance.

‍

Dig deeper
into the data

Join the Webinar→
The State of Agentic Financial Crime

‍

This session was part of The AI Files, Unit21's virtual summit on how fraud and AML teams are putting AI to work.

‍

Gal Perelman
Gal Perelman
Product Marketing Lead, Unit21

Gal Perelman is the Product Marketing Lead at Unit21, where she spearheads go-to-market strategies for AI-driven risk and compliance solutions. With over a decade of experience in the fintech and fraud sectors, she has led high-impact launches for products like Watchlist Screening and AI Rule Recommendations.

Previously, Gal held marketing leadership roles at Design Pickle, Sightfull, and Lusha. She holds a Master’s degree from American University and a Bachelor’s from UCLA, and is dedicated to helping banks and fintechs navigate complex regulatory landscapes through innovative technology.

‍

Learn more about Unit21
Unit21 is the leader in AI Risk Infrastructure, trusted by over 200 customers across 90 countries, including Sallie Mae, Chime, Intuit, and Green Dot. Our platform unifies fraud and AML with agentic AI that executes investigations end-to-end—gathering evidence, drafting narratives, and filing reports—so teams can scale safely without expanding headcount.
ATO
|
12
min

Account takeover fraud: the complete guide to prevention, detection, and response

Gal Perelman
Gal Perelman
Product Marketing Lead, Unit21
This is some text inside of a div block.
Risk and Compliance Operations
|
8
min

What examiners will ask about your AI now that model risk guidance leaves it out

Gal Perelman
Gal Perelman
Product Marketing Lead, Unit21
This is some text inside of a div block.
Risk and Compliance Operations
|
8
min

Running AI agents in production: how Rippling and Zepz prove quality every day

Gal Perelman
Gal Perelman
Product Marketing Lead, Unit21
This is some text inside of a div block.
See Us In Action

Boost fraud prevention & AML compliance

Fraud can’t be guesswork. Invest in a platform that puts you back in control.
Get a Demo