The Queue Used to Be a Curriculum — A Compliance Analyst on What Agent Handoffs Actually Feel Like

Every workflow diagram has an arrow labeled "escalate to human." On one end of that arrow is an agent's confidence threshold. On the other end is a person with a queue, a regulatory clock, and a signature that converts whatever the agent assembled into an institutional record. Funke Adeyemi sits on the receiving end.

Funke is a composite, built from the operational patterns, regulatory structures, and specific frustrations that accumulate where agent-generated case packages meet human accountability requirements in AML compliance. The screen she describes is real. The queue dynamics are documented. The feeling of picking up a case with an age you have to calculate yourself — nobody puts that in the architecture diagram.

You receive agent-escalated cases every morning. Walk me through the screen.

Funke: It's not blank. I want to be clear about that, because people assume I'm complaining about getting nothing. The case management system shows me a risk score, the alert trigger reason, usually 90 days of transaction history, watchlist screening results, sometimes an entity-relationship map.¹ If the agent got far enough, there's a drafted SAR narrative sitting there too.² On paper, it looks like I'm reviewing a completed package.

But "completed" and "ready for me to act on" are different things. Nobody designed a screen for the difference.

What's the gap?

Funke: The agent's output is basically a persuasive document. Here's what I found, here's how I weighted it, here's my recommendation. Useful. But when I pick up a case, I don't need to be persuaded. I need to know where I am. Which is a completely different question.

Where I am means: has anything already happened to this customer's account? Did the agent send an information request? Is there a hold that's already live? What's the regulatory age of this case — not the alert date, the detection date, which might be different because a system-generated flag doesn't necessarily start the 30-day SAR clock?³ That's a determination I have to make, and to make it, I have to reconstruct what the agent did and when.

And here's the thing that keeps me up at night. The summary tells me what the agent concluded. It doesn't tell me what it tried and came back empty on. Absence of evidence and evidence of absence are completely different things in this work. If the agent queried a fraud intelligence network and found no matches, that's a data point. If it never queried it, that's a gap I need to fill. The summary looks identical either way.⁴

How much time goes to reconstruction versus actual review?

Funke: More than anyone wants to admit. Some mornings the overnight batch drops 40-something escalated cases into my queue, and each one needs me to figure out where it actually is before I can figure out what to do. A third of my time on a case goes to establishing current state. Not history. State. What's live, what's pending, what's irreversible.

There's a maker-checker requirement in your workflow. How does that interact with agent-generated packages?

Funke: This is where I want to flip a table. Every escalation, every restriction I propose, every SAR I draft requires a second qualified analyst to independently approve it. Their name goes on the record with a timestamp and a rationale.⁵ So the handoff from the agent doesn't just need to be good enough for me. It needs to be good enough for the person checking my work to form their own independent judgment.

And here's the perverse part. An agent-generated narrative that reads beautifully, that flows, that's persuasive? It can actually make the checker's job harder. Because they're supposed to independently verify, not agree with a well-written essay. They need raw evidence, not the argument.

I've had reviewers approve things they shouldn't have because the agent's draft was... compelling. A too-good handoff that short-circuits the very control it's supposed to support. That's not a quality problem. That's a quality problem wearing the costume of a quality improvement.

And when the examiner shows up?

Funke: They're looking at analyst notes, queue history, decision rationale.⁶ If my notes say "reviewed agent package, concur with recommendation," that's tissue paper. What holds up is: "independently verified counterparty against OFAC list, confirmed no prior SAR history, assessed transaction pattern against customer's stated business purpose." The agent may have done some of that work. But my signature says I did it.

Has the nature of your queue changed as agents handle more routine triage?

Funke: This is the subtle thing nobody warned me about.

When I started, maybe 70% of what hit my queue was obvious false positives. You'd look at it, confirm the alert was noise, close it, move on. Boring work. But it was also calibrating. You develop a sense of what normal looks like by seeing enormous amounts of normal.

Now the agents close the obvious stuff, which is genuinely great. But what's left is all the weird ones. Every single case in my queue is the one the agent couldn't figure out. So I'm making harder judgments with less baseline context, every day, all day.

Do you miss the false positives?

Funke: Every single day. Nobody believes me when I say that. They think I'm being nostalgic. I'm being serious.

What would a better handoff actually look like?

Funke: I'd settle for a panel that separates what the agent concluded from where the case is. Two distinct artifacts. One is the investigation summary: findings, reasoning, recommendation. The other is a re-entry brief: current account status, actions already taken, data sources consulted and not consulted, regulatory clock status, and what options I still have with their deadlines.

Right now those are tangled together in one narrative, and I have to untangle them under a clock that's already running.

Is there anything the current setup gets right?

Funke: The drafted narrative saves real time. One estimate puts SAR prep going from roughly a week down to under 30 minutes for straightforward cases, and honestly, that tracks with what I see.⁸ The transaction timeline assembly is genuinely good. And the risk scoring helps me prioritize the queue, which matters when you've got 40 cases and a clock on each one.

The tools aren't bad. They were designed to explain what happened. I need them to tell me where I am. Those sound like the same thing until you're the one whose signature goes on the record.

Funke Adeyemi is a composite character. The operational patterns, regulatory requirements, and system artifacts she describes are drawn from publicly documented AML compliance frameworks, case management architectures, and regulatory guidance.

Footnotes

1. Flagright, "What Happens When Fraud Gets Caught in Real Time." https://www.flagright.com/post/what-happens-when-fraud-gets-caught-in-real-time ↩

2. Unit21, "Agentic AI for AML Compliance: A Practitioner's Guide." https://www.unit21.ai/blog/agentic-ai-for-aml-compliance-a-practitioners-guide ↩

3. Compliance Cohort, "SAR Filing Deadline." https://www.compliancecohort.com/blog/timing-of-sar-filing ↩

4. Taktile, "AI agents in compliance and AML: A responsible deployment guide." https://taktile.com/articles/ai-agents-compliance-aml ↩

5. KLA Digital, "Completed FRIA worked example: AML alert-triage agent." https://kla.digital/blog/fria-worked-example-aml-alert-triage-agent ↩

6. IntelligentHQ, "The Alert That Starts the Case: How AI Transaction Monitoring Is Rewriting Corporate Fraud Investigations." https://www.intelligenthq.com/the-alert-that-starts-the-case-how-ai-transaction-monitoring-is-rewriting-corporate-fraud-investigations/ ↩

7. Norton Rose Fulbright, "FinCEN and banking agencies clarify FAQs on suspicious activity reporting." https://www.nortonrosefulbright.com/en/knowledge/publications/e8c1dfb5/fincen-and-federal-banking-agencies-issue-clarifying-faqs-on-suspicious-activity ↩

8. Unit21, "Agentic AI for AML Compliance: A Practitioner's Guide." https://www.unit21.ai/blog/agentic-ai-for-aml-compliance-a-practitioners-guide ↩