Conversations and solo deep dives on building AI agents. No fluff — just what's actually working.
Guardrails, redaction, and fabrication checks all add latency. Anup Kamath points to the one lever most teams underrate: smart memory management is the rare move that improves the experience and wins back the time those safety layers cost you.
Is it conversational AI, or just a chatbot with a new label? Anup Kamath draws the line: real conversational AI understands intent, holds context, and reasons through ambiguity — a chatbot, an IVR, and a script don't.
The smartest AI isn't the one that answers everything. Anup Kamath makes the case for restraint: the system reads frustration signals, brings in a human at the right moment, and hands off so cleanly that the AI and the person operate as one team — not two disconnected agents passing a ticket.
Most people picture AI memory as one bucket. Anup Kamath breaks it into four working layers: session for the live conversation, workflow for the task in progress, longitudinal for a person's history over time, and an enterprise knowledge layer the whole system draws on.
At hundreds of millions of calls, every millisecond matters. Anup Kamath reframes latency as a human-experience metric, not just a technical one — the moment a pause reads as hesitation, the continuity, trust, and reliability the person is counting on start to slip.
The biggest mistake teams make with AI? Trying to automate everything. Anup Kamath breaks down when a system should stop and route to a human — clinical complexity, ambiguity, emotional cues, repeated failures, and moments of real vulnerability.
Everyone obsesses over the model. Anup Kamath argues the real work is in the guardrails — grounding every response in verified data, handling sensitive information as non-negotiable, and refusing to let the system fabricate an answer it doesn't have.
Conversational AI can now fool 9 out of 10 people into thinking it's human. So why are more callers than ever demanding a real one? Anup Kamath calls this the AI trust paradox — and solving it, he argues, is a design problem, not a technology one.
Will autonomous AI compress jobs in AML? Abhishek Mittal is candid: the tasks getting automated are the low-complexity, deterministic ones that arguably shouldn't have existed in the first place, the work that never tapped human expertise or creativity anyway.
With AI vendors shipping industry plugins every week, is deep domain expertise still a competitive moat? Abhishek Mittal says yes, but the role has changed: domain experts have moved from handing over requirements once to providing ongoing context and driving the actual decisioning.
If every company runs the same frontier models with the same features, what's left to compete on? Abhishek Mittal's answer: context. The ontologies, taxonomies, and prompts you build are becoming the real differentiator in an AI-commoditized world.
Regulators are pushing AML teams to stop counting checklist boxes and focus on the signal in the noise. Abhishek Mittal on why the number of SAR filings is the wrong measure of success, and how AI can finally move the 1% of financial crime we actually catch.
In anti-money laundering, an AI told to protect a bank's profits could bury the very transactions it's supposed to flag. That's alignment faking — and it's a real risk for AI in AML.
Most AI learns from known patterns. So how does it catch financial crime it has never seen before? Abhishek Mittal on why the best AML systems stitch together deterministic rules, probabilistic AI, and human expertise, dialing each up or down depending on the pattern.
$4.4 trillion flows through global financial crime every year — and current systems catch barely 1% of it. Can AI change that, or is it just flooding anti-money laundering (AML) teams with more noise?
@agentsandactionspodcast
Headline: Is Your AI Agent Compliance-Ready? 🤖🏦 The Future of Banking is "Agentic," but the Moat is Domain Knowledge.