Leaders from Revefi, Monte Carlo, Dollar General and Verizon Business said data and AI observability are becoming one discipline. They explained why unmonitored AI degrades, why observability can't be bolted on after deployment, how to benchmark readiness and why model swaps are a leading cause of drift.
Key takeaways
- One system, two lenses. Data observability makes inputs trustworthy; AI observability keeps outputs reliable, and panelists said the two are converging.
- Unobserved AI degrades. Monte Carlo found its own agent quietly drifting from its instructions only because it was monitored.
- Don't bolt it on. Start with an inference table that stores outputs, traces and telemetry next to the inputs.
- Measure data downtime. The number and duration of data quality incidents shows how long agents may be reading bad data.
- Model swaps cause drift. Upgrades to commercial models can change outputs overnight, so test new versions against old ones.
Data observability and AI observability are converging into one discipline, according to a panel of practitioners at the Data Science Connect webinar "From Black Box to Glass Box: Observability for Scalable AI Systems." Without monitoring, they said, AI systems only get worse.
The panel featured Girish Bhat, SVP at Revefi; Michael Segner, who works in product strategy at Monte Carlo; Sachin Mor, Director of Application Engineering at Dollar General, who leads AI for supply chain and merchandising; and Don Spalding, Senior Innovation Adviser at Verizon Business. Brian Mink, co-founder and president of Data Science Connect, moderated the session, which was sponsored by Monte Carlo, Revefi and SAP.
What is the difference between data observability and AI observability?
Mor gave the definitions the rest of the panel built on. Data observability covers the health of the data itself: freshness, volume, schema and quality. AI observability watches what models do, including drift, bias, fairness and performance as data grows. The overlap comes where data feeds the model. Data observability, he said, "ensures the inputs are trustworthy," while AI observability makes sure outputs stay reliable.
Bhat said AI observability extends classic data and application observability to models and LLMs, but tools remain siloed. "If you don't do the data observability part that includes cost, quality, performance, AI observability will just fail," he said. Spalding described the two as a feedback cycle in which model outputs drive actions on the data.
Segner said Monte Carlo's co-founder coined the term data observability in 2019, and the company now calls itself a data and AI observability platform. Where to start depends on the organization, he said. Teams still migrating systems should get data AI-ready first, while teams already shipping agents can monitor outputs and work back to the data. "It's one system," he said.
Why does AI observability matter at scale?
"If you don't observe a system, there's only one direction it goes. It degrades over time," Segner said. He said Monte Carlo has multiple agents in production with 30,000 users. Monitoring of one, an agent that recommends data quality rules, showed a workflow meant to suggest multi-field rules was suggesting none, or only simple ones. Left unchecked, he said, that would have eroded adoption and the feature's value.
Spalding described a budget airline whose chatbot told a user to bring movies because long delays were likely. That answer made its way into the airline's FAQs. Risks run from customer experience to GPU and token costs, he said.
For Mor, the issue is trust. When forecasting and replenishment models scale across distribution centers, business users ask how a recommendation was reached. Unless teams reveal what's inside the black box, he said, the business won't trust it. Bhat added that for customers underwriting loans, routing supply chains or approving patients for care, a black-box approach is a liability.
How do you build observability in from the start?
"The wrong way to do observability, whether it's AI or others, is to bolt it on after deployment," Bhat said. He called for instrumenting pipelines, feature stores, models and inference endpoints, and said logging practices that are long established elsewhere are still treated as novel for LLMs.
Segner said many teams test agents carefully in development against golden datasets, then lose that visibility in production. His first step: "make sure you have an inference table," storing outputs, telemetry and traces in the warehouse or lakehouse next to the inputs, so every decision point an agent makes is visible.
Mor compared it to writing test cases before development starts. He recommended defining which LLMs and governance practices a project will use up front, and running a data audit before building. He described an architecture review board that sits in the development workflow and checks technology choices, including whether an LLM is authorized.
How do you know you're ready for AI observability?
Segner suggested measuring data downtime: how many data quality incidents occur and how long they take to resolve, since that's the window when agents may read bad data. "It's impossible for your agent to have the right output if the context or input it's given is incorrect," he said.
Spalding focused on mechanisms more than outputs: layered telemetry at the data, model and business layers, guardrails between users and models, versioning and lineage, sandbox testing, SLAs with vendors and people kept in feedback loops. Bhat started with the business case and said cross-functional stakeholders, including legal and compliance, should frame the observability roadmap.
On regulation, Bhat described observability as a technical mechanism for compliance through logging, traceability and post-market monitoring. Segner cited a financial services customer in Australia that must report every incorrect data point, explain where its systems broke and show how it fixed the problem. He expects similar requirements for AI.
How do teams detect model drift today?
Spalding and Mor said drift detection is still largely reactive and based on outputs. Mor said much of the drift he has seen traces to data arriving late.
Segner said the most common drift he hears about is a wholesale model swap. One organization's customer-facing agents were thrown off for two to three days after an upgrade from GPT-4 to GPT-4.5, he said. He recommended tracking which model runs each workflow and testing embedding models regularly for drift. Bhat advised evaluating new model versions against previous ones rather than assuming the latest is best.
In the closing round, Segner warned against treating "AI readiness as a gate rather than a goal," and Spalding urged teams to build observability into the architecture: "Don't wait for it to become an add-on."
Keep the conversation going
Data Science Connect has 24 free live webinars on its 2027 schedule in four series: the AI × Enterprise Panel Series, AI Demo Days, Agentic AI Showcases and the Plug & Play Series. See the webinar schedule for upcoming sessions.
Frequently asked questions
What is the difference between data observability and AI observability?
Data observability monitors the health of data, such as freshness, volume, schema and quality, while AI observability monitors model behavior, such as drift, bias and performance, according to Dollar General's Sachin Mor.
What is an inference table for AI agents?
A table in the warehouse or lakehouse that stores an agent's outputs, telemetry and traces next to its inputs, which Monte Carlo's Michael Segner called the first step in observing agents in production.
How do you benchmark AI observability readiness?
Segner suggested tracking data downtime, the number of data quality incidents and how long they take to resolve, since that is when agents may access incorrect data.
What causes model drift in production AI?
Panelists said drift often comes from late or bad data, embedding problems, and wholesale swaps or upgrades of commercial models, and that detection today is still mostly reactive.
Watch: From Black Box to Glass Box: Observability for Scalable AI Systems on YouTube
Event: Webinars · More coverage: Webinars insights
