Inside Our Final TDAI Council Meeting of 2025: When Agents Meet Reality
Table of Contents
We wrapped up 2025 with our final Trusted Data for AI (TDAI) Advisory Council meeting, and if there’s one thing everyone agreed on, it’s this: AI agents are finally escaping experimentation and entering production. But here’s the catch: trust, traceability, and evaluation are still playing catch-up.
Data + AI leaders from Western Union, Roche, Fox, Drata and others, all showed up with the same question: How do we make sure these agents actually work when it matters?
Table of Contents
What’s Actually Happening in Production?
Before we got to answering that critical question, council members kicked things off by opening up about how they’re really using agentic AI today. Spoiler: it’s less “robot takeover” and more “smart assistant with strict guardrails.”
In FinTech and Fraud Detection: Western Union shared their plans to deploy an agentic fraud workflow that combines outlier detection, deep historical context, and LLMs to surface things like negative news, jurisdiction risks, and unusual transaction patterns. The key constraint? PII guardrails and model controls aren’t optional. They’re non-negotiable. The goal is to reduce manual operations friction without putting compliance at risk.
In Pharma and Life Sciences: One member described using AI to draft 800-page regulatory reports, slashing first-draft time from weeks to hours. They’re even exploring reviewer agents to assist human sign-off. But make no mistake: humans remain the final decision-makers. Always.
In Enterprise Operations and Finance: Other teams are deploying internal AI assistants for everything from sales recommendations to schema tagging to CFO analytics. Some are experimenting with AIOps and self-healing data pipelines.
Despite the different industries and use cases, the themes were remarkably consistent. It wasn’t about model performance anymore. It was about accuracy, explainability, and whether the business actually trusts what the AI is telling them.
Enter Agent Observability
Monte Carlo’s CTO & Co-founder Lior Gavish came in with a preview of Agent Observability, our newest capability. It’s designed to answer one simple but critical question: “How do I know my agents are behaving correctly in production?”
Built on OpenTelemetry traces that stay in your data environment (not ours), Agent Observability gives you three superpowers:
- Track operational health for every step in an agent workflow. Latency, cost, errors, tool calls, all of it.
- Evaluate those fuzzy quality signals that matter but are hard to pin down (relevance, safety, tone) using LLM-as-judge evals or custom logic.
- Get alerted when behavior drifts, so you can dig into specific traces instead of playing guessing games after the damage is done.
The council emphasized what mattered most to them: keeping traces in-region and under their control, plugging seamlessly into existing observability stacks, and (this was big) eventually using these signals to improve prompts and policies proactively, not just reactively.
We’re Still Counting the Wins (But Expectations Are Rising Fast)
As Barr Moses, CEO & Co-founder of Monte Carlo put it, we’re still early enough in the AI era that we’re counting “hell yeah” moments. But expectations are rising faster than most teams can keep up with.
The TDAI Council is sticking around for 2026, and we’ll keep charting this path together. Because at the end of the day, the AI products we’re building need to be more than just powerful. They need to be reliable, explainable, and grounded in trusted data.
If you’re wrestling with similar challenges (whether it’s getting agents production-ready or just making sure your data pipelines don’t betray you at the worst possible moment), you’re not alone. And hey, that’s exactly why we built this community.
Our promise: we will show you the product.
