Risk and Compliance Operations

The groundwork before turning AI on: lessons from Kinecta and Crypto.com

Published
November 19, 2026
Read Time
8
mins
Gal Perelman
Gal Perelman
Product Marketing Lead, Unit21
Subscribe to stay informed
Table of contents

Most of the work in an AI rollout happens before anything gets switched on. At The AI Files, Unit21's virtual summit on AI in fraud and AML, Unit21 CEO Tyler Allen put it this way: the technology usually isn't the hardest part. Getting the organization ready is. That means making the case for budget, agreeing on how to measure return, and aligning on what AI is actually for.

‍

Two leaders who have done that groundwork walked through how it went.

‍

Jorge Cortes is VP of Enterprise Risk Management at Kinecta Federal Credit Union, a Southern California credit union with just under $7 billion in assets and about 250,000 members.

‍

Sercan Karapoyraz is VP of Compliance Advisory at Crypto.com, where he has spent five years and now focuses on global financial crime strategy and oversight. Crypto.com runs spot crypto, securities, derivatives, card programs and custody, so financial crime there is never a single lane: the team weighs on-chain and off-chain risk together.

‍

Before AI: stressed teams and too many tabs

At Crypto.com, one decision meant five or six tabs. Analysts weighed on-chain signals, like darknet exposure or stolen funds, against off-chain ones such as screening hits, device and geolocation anomalies and unusual transaction trends. Crypto.com uses the Unit21 risk engine for that off-chain side. Before every decision, an analyst still had to pull the pieces together by hand: the KYC profile, blockchain analytics, cross-referenced IPs.

‍

At Kinecta, the team was always racing the queue. When Jorge took over, the team was two fraud analysts and two investigators. It has grown since, but fraud grew more complex faster, and each alert took longer to research and investigate. Adding headcount to keep pace was close to impossible. "We had stressed teams," Jorge said. His people took pride in the work, but they had no real work-life balance.

‍

Making the case for AI

The two leaders made the case to their organizations in different ways.

‍

Jorge tied AI to fraud losses. To fund it, he committed to bringing losses down, so the investment would pay for itself. That made turning it on, in his words, a no-brainer. Giving his team its time back was part of the goal too. Tyler pointed out the link: analysts who aren't burned out do better work and find more fraud.

‍

Sercan needed proof, not promises. He was leading 70 analysts and investigators plus outsourced operations, and he had sat through countless machine learning vendors that fell short on transparency, auditability or flexibility. What finally moved him was an approach that gave his team full control of the AI. Crypto.com's leadership was AI-forward, but the mandate was concrete: show tangible operational value, prove it scales capacity without compromising compliance, and after launch, show measurable efficiency and rock-solid auditability.

‍

Picking the first workflow

Crypto.com started with level one alerts. That was the bottleneck: the highest volume, and the lowest-risk, highest-reward place to begin. Level one work runs on standard checklists and clear decision trees, so it needs far less subjective judgment than level two. Sercan's team designed an agent to spot red flags, group related patterns and draft the initial summary. Analysts verify and reconcile the data instead of gathering it. As any investigator will say, writing a narrative from scratch is the slowest and most tedious part of the job, so a first draft was a game changer.

‍

Next on Sercan's list: case-level agents that work like a second-level investigator, using AI to refine existing rules, and an EDD agent. That last one would read source of funds and source of wealth documents, where a detail on page 15 can be the core of a case, compare them with what the customer declared, and enrich the review with benchmarks like local GDP or typical salaries for an occupation.

‍

Kinecta started with small, recurring tasks. A Unit21 team ran an on-site workshop, which was the analysts' first introduction to AI agents. The starting prompt was simple: what is a recurring task you need help with? One investigator used to build transaction lists in Excel by hand. Now an agent compiles the list inside Unit21 and ties it to the case. Kinecta runs a number of agents like that today and keeps adding to them. Jorge said the main limit now is imagination: the more comfortable analysts get, the more they try.

‍

"Why can't we just use Copilot?"

Kinecta rolled out Microsoft Copilot to every employee, so Jorge heard the question. He gives Copilot credit: it summarizes data, tracks decisions, prepares presentations, produces trend reports and has even helped the team adjust some of its Unit21 rules. The difference is where the work happens. Unit21's agents work like an AI fraud analyst embedded in the platform, running against the institution's transaction data. Analysts no longer start every alert from zero.

‍

Jorge described the before and after plainly. Because of the volume, his analysts and investigators "were processors." With agents laying the groundwork, "they are truly analysts" and spend their time where it counts. Tyler made a similar point in his keynote: general tools like Copilot aren't designed for AML or fraud, and teams get more from going deep on one workflow. Read more in what 1M live AI alerts tell us about how fincrime teams adopt AI.

‍

Fix the data, not the prompt

Both teams learned that an agent is only as good as the data and procedures behind it.

‍

AI is almost too polite. In Crypto.com's first tries, the agent wasn't getting the most relevant data. Sercan said that when information is missing, AI happily fills the gap for you, and in AML that kind of helpfulness is dangerous: assumptions turn into wrong conclusions or missed risk. The lesson was to stop tweaking prompts when the data layer is broken. Working with Unit21, the team remapped fields, broke the data into more structured sections and pointed each agent at a specific dataset with targeted prompts. Once the data architecture was clean, performance changed significantly.

‍

Kinecta found gaps by watching agents underperform. As a heavily regulated institution, Kinecta had its procedures written down. But some data wasn't mapped correctly, and some knowledge lived only in people's heads. The agents exposed those gaps because they didn't perform as expected. After the team mapped the data and documented the processes, the results were significantly better.

‍

Tyler noted that this cleanup would have helped long before AI. The difference now is momentum. Teams that try to put AI on top of a messy data foundation hit a ceiling that no amount of prompt tuning can fix.

‍

Watch the full session with Kinecta and Crypto.com
Hear how both teams made the case for AI, picked a first workflow and got their data ready.
Watch on demand

‍

Capture what your best analysts already know

Crypto.com sat down side by side with one of its best investigators and mapped their subconscious checklist. The team turned that mental model into questions. What data points do you always check first? When does an anomaly become a red flag? What unwritten threshold triggers an escalation? Those routines became modular decision trees the agent follows step by step.

‍

This is the same gap Tyler sees across deployments: the SOP is out of date, and the real process lives in one experienced analyst's head. An agent can only follow what is written down.

‍

From skepticism to buy-in

Analysts at both organizations started out wary.

‍

At Crypto.com, the reaction was mixed. People wondered whether the agent would replace them or create more mess to clean up. That changed once they saw the first results. The agent wasn't making the final decision. It pulled the pieces together and drafted the summary, so nobody had to start from a blank page on every alert. Skepticism turned to buy-in.

‍

At Kinecta, the workshop was the turning point. Jorge remembers analysts asking what they were even doing and how to write a prompt. Once they saw what an agent could do, that flipped. Now they are excited and looking for the next thing to build. They take pride in finding the root cause of a scheme or recovering a loss for a member, and they see the agents as a way to look at more and catch more fraud.

‍

Advice for a peer about to start

Tyler closed by asking what each would tell a peer who calls about deploying AI.

  • Start now. Jorge's first answer: "Good luck. And it's about time. What are you waiting for?"
  • Start small. Both said to pick one repetitive, low-judgment pain point and build on early wins.
  • Map your data and put as much of it in as you can. The more data the agent can reach, the more holistically it can analyze, Jorge said.
  • Don't fall for the full automation dream. Sercan warned that promises to automate everything from alert ingestion to SAR filing usually end in a big, expensive disappointment.
  • Test, fail fast and retest. Sercan tested his first prompt at least 15 times and kept asking Unit21's team questions until he understood what the agent could do.
  • Design for a third party. Whoever writes the prompts will develop blind spots. Sercan's test: can an auditor or regulator understand why the agent flagged this or wrote the summary that way? If the answer isn't a confident yes, you aren't ready to scale. For what examiners will look for, see what examiners will ask about your AI.

‍

A credit union and a global crypto platform started in very different places. They ended up doing the same groundwork: make a concrete business case, start with one workflow, fix the data before the prompts, write down what your best people know, and build so a regulator can follow every step.

‍

For what happens after go-live, read how Rippling and Zepz prove agent quality every day. For the build vs. buy question, see how Circle splits its fincrime AI stack.

‍

Frequently asked questions

What should a team do before turning on AI for fraud or AML?

Make a concrete business case, such as lower fraud losses or more analyst capacity, then pick one repetitive workflow, map and clean the data the agent will use, and write down how your best analysts actually work. Kinecta and Crypto.com both saw agent performance improve significantly once the data and processes were fixed.

‍

Why not just use Microsoft Copilot for fraud alert reviews?

Kinecta uses Copilot company-wide for summaries, reporting and presentations. For alert reviews it relies on Unit21's agents, which work like an AI fraud analyst embedded in the platform and run against the institution's transaction data, so analysts no longer start every alert from zero.

‍

Where should you start with AI agents in financial crime?

Start with a high-volume, low-judgment task. Crypto.com began with level one alerts, where an agent spots red flags, groups related patterns and drafts the initial summary for an analyst to verify. Kinecta began with small recurring tasks its analysts identified, such as compiling transaction lists for cases.

‍

How do you get analysts to trust AI agents?

Show them early results and keep the analyst in charge of the decision. At Crypto.com, skepticism turned to buy-in once analysts saw the agent did the heavy lifting but did not make the final call. At Kinecta, a hands-on workshop moved analysts from hesitancy to building their own agents.

‍

Dig deeper
into the data

Join the Webinar→
The State of Agentic Financial Crime

‍

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.
AI Tasks
|
6
min

AI Task Spotlight | Edition No. 13: Human Trafficking, One Red Flag, One Rule, One Investigation

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

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

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