What AI-Assisted AML Investigations Actually Look Like
Maya has been an AML analyst for three years. She works at a mid-sized fintech that processes about $2 billion annually, serving small businesses across the U.S. She knows the work matters. She also knows that most days feel like treading water.
This morning, she logs into the case management system and sees 47 alerts in her queue. Yesterday she closed five cases. The day before, four. The math is not in her favor. Her queue has grown by eight alerts since Friday, and it is only Tuesday.
The alerts are rule-based, which means they spike unpredictably. Last month, a holiday weekend triggered a surge of flagged wire transfers. The week before that, a batch of new merchant onboardings set off velocity rules across dozens of accounts. Each spike adds to the backlog. Each backlog adds to the pressure. Maya keeps a running tally in her head of how far behind she is.
The Alert
She clicks into the next case. The alert summary shows a small business account with a pattern of round-dollar wire transfers totaling $127,000 over six weeks. The funds went to three different LLCs, all registered in Delaware within the past 90 days. The account holder runs a landscaping company in Ohio.
On the surface, it could be anything. A business owner paying contractors through shell companies for a renovation project. A victim unknowingly moving money for a scam operation. Or something worse. Maya has seen all of it.
How Maya Would Investigate This Manually
She opens a blank Notes document and types the account holder's name at the top. This is where she will track everything she finds, because the case management system is clunky and she has learned the hard way that she needs her own record of what she checked and when.
First, she pulls the full transaction history. The internal database loads slowly. She exports six months of activity into a spreadsheet and starts scanning for patterns. The round-dollar transfers are obvious, but she needs to look at everything else too. Are there other suspicious counterparties? Any cash deposits that preceded the wires? She highlights rows, sorts columns, scrolls back and forth. This takes twenty minutes.
Next, she pulls the KYC file. The business was onboarded 18 months ago. The documentation looks standard: articles of incorporation, a driver's license, a utility bill. She checks whether the address on file matches what is on the license. It does. She notes this in her document.
Now the external research begins. She opens a new browser tab and searches for the landscaping company. The website is basic but functional. She screenshots the About page and the contact information. She opens Google Maps and finds the business address. It is a real building. She switches to LinkedIn and searches for the owner. She finds a profile that matches, with a work history that makes sense. She notes all of this.
Then she looks up the three Delaware LLCs. She navigates to the Delaware Division of Corporations website and searches each one. Two were registered by the same registered agent. She copies the agent's name and searches for it separately. This leads to a rabbit hole of other LLCs, some with similar naming conventions. She is now eight tabs deep and losing track of which entity is which. She pauses to update her notes.
She checks OFAC. No hits on the account holder. She runs the three LLCs through the sanctions screening tool. No hits there either. She checks the adverse media database. Nothing. She searches Google News for any of the names. Nothing.
She pulls up past alerts on this customer. There was one flag two years ago, unrelated to this activity. She reads through the narrative to understand the context. It was a false positive that got escalated due to a documentation gap. She notes this.
By now, ninety minutes have passed. Her eyes are tired from switching between windows. She has twelve browser tabs open, a four-page Notes document, and a growing suspicion that this account holder is a money mule who does not know it. But she is not certain. She needs to trace where the money went after it hit those LLCs, and that information is not in her system. She makes a note to request it from the downstream bank.
Her queue has grown by two more alerts while she worked.
How Maya Investigates with Roe's AI AML Investigator
Now imagine a different version of Maya's morning.
She arrives at 8:30 AM, sets down her coffee, and logs into the case management system. The 47 alerts are still there. But something is different. Overnight, while Maya slept, Roe's AI AML Investigator was working. It received alerts as they came in, and it started investigating them immediately. By the time Maya sits down, 43 of those 47 alerts already have completed investigation summaries waiting for her review.
The landscaping company case is one of them. The alert triggered at 2:47 AM when a batch process flagged the wire transfer pattern. By 3:02 AM, the AI AML Investigator had finished its work. Maya is looking at a case that was investigated six hours ago, while she was asleep.
Here is what happened during those 15 minutes.
The AI AML Investigator received the alert through an API connection to the case management system. Within seconds, it formed an investigation plan. It identified the relevant data sources, the risk checks to run, and the sequence of steps it would follow. The plan was based on the standard operating procedures that Maya's compliance team had configured, so it mirrored what a human investigator would do.
The agent started executing. It pulled the transaction history from the data warehouse and ran SQL queries to identify patterns: round-dollar amounts, frequency of transfers, new counterparties, velocity anomalies. It cross-referenced the recipient LLCs against the KYC data on file. It searched public records for each LLC, checking registration dates, registered agents, and any links to other entities. It ran OFAC and sanctions checks. It searched Google, LinkedIn, and state business registries. It reviewed past alerts on this customer.
This is where Roe's AI AML Investigator differs from a simple AI summary tool. It iterates. It does not run one query and stop. It runs a query, evaluates the results, and decides what to investigate next. If the registered agent search reveals a network of related LLCs, the agent follows that thread. If a public records search turns up a news article, the agent reads it and factors it into the analysis. On average, the AI AML Investigator goes through 8 to 15 iterations per investigation, executing 100 to 250 discrete steps. The process takes about 15 minutes per case.
While Maya was sleeping, the AI AML Investigator worked through dozens of cases. It does not take breaks. It does not get tired at 3 AM. It does not lose focus after the tenth alert in a row. By the time Maya's alarm went off, her queue was already triaged.
As the agent works, it bookmarks evidence. Every transaction it flags, every public record it finds, every sanctions check it runs gets tagged and stored. This creates an audit trail that Maya can review later.
What Maya Sees
Maya opens the landscaping company case in Roe's interface. The first thing she sees is the recommended action: the AI AML Investigator has flagged this case as suspected suspicious activity.
Below that is an executive summary. Each finding is a single line with an annotation pointing to the specific evidence. For example: "Recipient LLCs registered within 90 days of first transfer" links to the Delaware Division of Corporations records. "Round-dollar transfers totaling $127,000 across 6 weeks" links to the specific transactions in the data warehouse. "Registered agent associated with 14 other LLCs formed in the same period" links to the public records search results.
If Maya wants to go deeper, she can. She clicks into the full investigation log and sees every step the agent took: the queries it ran, the searches it performed, the sources it checked. She can see the agent's reasoning at each decision point. If she wants to understand why the agent flagged a particular risk check, she can drill into that check and see both the red flags and the green flags it identified.
This level of transparency matters. Maya is not being asked to blindly trust an AI recommendation. She is being given a detailed evidence package that she can verify, challenge, or build upon.
The Feedback Loop
Maya reviews the findings. She agrees with the overall conclusion: this looks like a money mule situation, and the account holder is likely a victim who was recruited through a job scam or romance scheme. The pattern matches cases she has seen before.
But one risk check catches her eye. The AI AML Investigator flagged the transaction velocity as high-risk based on a threshold that Maya thinks is too sensitive for this customer segment. Small business accounts often have lumpy cash flows, especially in seasonal industries like landscaping. She adjusts the threshold value for this rule so that future investigations will calibrate appropriately.
Later that day, she reviews another case. This time, she notices that the agent used a signal she had not seen before: it cross-referenced the recipient's business address against a database of known mail drop locations. This was not part of the original SOP. The AI AML Investigator identified it as a relevant check based on the patterns it had learned from previous investigations.
Maya decides this is valuable. She adds the mail drop check to the standard process so that the agent will run it consistently going forward. The system learns from her, and she learns from it.
The Outcome
The landscaping company case is suspicious. The account holder appears to be an unwitting money mule, recruited through a fake job posting and used to move funds for a fraud ring. Roe's AI AML Investigator drafts a SAR narrative based on the evidence it gathered, formatted to match Maya's team's template. Maya reviews the draft, makes a few edits for clarity, and submits it.
What would have taken her three hours of manual work took 20 minutes of her attention. She closes the case and moves on to the next alert. For the first time in weeks, she feels like she might actually get through her queue.
What This Means for Investigators
AI does not replace Maya. It handles the data gathering, the pattern recognition, the tedious cross-referencing across a dozen sources. Maya makes the final call. She reviews the evidence, applies her judgment, and decides whether the case is suspicious or a false positive. She provides feedback that makes the system smarter over time.
The work is still demanding. The stakes are still high. But the balance shifts. Instead of spending her day drowning in tabs and spreadsheets, Maya spends it on the decisions that actually require human judgment.
Roe's AI AML Investigator runs investigations in about 15 minutes on average, executing 100 to 250 steps and iterating 8 to 15 times per case. It connects to your existing data warehouse, transaction monitoring system, and KYC data through a flexible API. It follows the SOPs your team defines, and it learns from investigator feedback through a built-in memory system.
If your AML team is buried in alerts and losing ground every week, let's talk. Schedule a demo to see how this works with your data and your workflows.
