A US family finance app's AI now settles 90% of its security alerts alone
Greenlight built an AI system that checks most of its 3,000 to 4,000 monthly security and fraud alerts in seconds. People, or rules they approved, still carry out every fix.
The short version
- Greenlight, a US company that makes debit cards and savings tools for families, built an AI system called Luna to handle its security alerts.
- Luna now resolves more than 90% of up to 4,000 cases a month without a person stepping in.
- People or pre-approved rules still carry out every fix, which met the company's compliance requirements.
What Greenlight built
Greenlight handles sensitive money data under the same rules as major banks. Its security team gets 3,000 to 4,000 alerts a month. Some are customer cards found for sale on the dark web. Others are odd login patterns, system warnings or suspicious payments.
Each alert needs checking against several systems before anyone can decide. Is the card still active? Was it reported before? Did the customer's app sessions look strange? Is there an open support ticket? A human analyst needs 15 to 30 minutes per case.
Greenlight built Luna on Amazon's cloud service for AI agents. It is organized like a security team. A first-line AI supervisor sorts each case and calls on more than 30 specialist AI helpers. Each helper checks one source, for example card records. Routine cases get a decision in seconds.
About 10% of cases are harder. They go to a second-line AI that works through 3 to 5 data sources. It writes a full report, with evidence, for a person to review.
The part that made it work
Memory made a big difference. Luna records how past cases ended and learns from them. In one example, a Greenlight system runs routine maintenance every Tuesday at 3 AM. Luna flagged it the first time. Once it learned the pattern was expected, it stopped raising it.
It also keeps lessons for longer. One is that cards reported by a particular dark web source are usually false alarms. When Luna resolves a case the same way three times, it records that as a confirmed pattern.
Memory, together with the second-line AI, lifted Luna's resolution rate from 83% to more than 90%.
By hand, Greenlight's monthly caseload would take up to 2,000 analyst hours. The company estimates Luna saves 900 to 1,800 of those hours each month. Analysts now spend that time on cases that need judgment.
Why people still sign off
Luna checks cases and recommends what to do. It does not fix problems on its own unless an approved policy allows it. A person, or a pre-approved automatic rule, carries out the fix.
That split satisfied the company's compliance needs. It also helped the security team trust the system. The two levels matched how their own analysts already worked. Juniors handle volume, and seniors take the hard cases.
Once it worked for security, Greenlight reused it. Within two months, Luna was also checking financial risk cases, mostly through new specialist helpers.
What it means for companies in Indonesia
Indonesian banks and online sellers face the same flood of security alerts.
Start with one task where a right answer is clear. Let AI sort the routine cases quickly. Give it a memory of past decisions so it stops repeating false alarms. Keep people, or written rules, in charge of every fix. That last point matters for anyone answering to a financial regulator.
What to do next
- Count your monthly security and fraud alerts, and how long each one takes.
- Pick one alert type with a clear right answer, such as leaked cards, for a first trial.
- Make sure the AI can learn from past decisions, so known false alarms stop coming back.
- Keep a person or a written rule in charge of every fix.
- Once it proves itself, reuse the same setup for fraud or risk checks.