AI Supper Club Bangalore: after the agent ships
AI agents are easy to demonstrate. Keeping them reliable once they run inside an organization is harder. That was the subject of the AI Supper Club in Bangalore, and the conversation moved quickly into AI that is already in use, and what goes wrong once it is.
Guests came from enterprise software, automotive, support automation and internal AI enablement. People brought examples of things that had broken, and they disagreed with each other about how to deal with them.
One question connected most of the evening: how do you know an AI agent is still doing the job you originally assigned it? The infrastructure can run, the scheduled jobs can succeed and the outputs can look plausible, while the agent has quietly changed what it treats as important.
What stayed with us
Agents can drift from their job
One guest described a monitoring agent that ran four times a day. After about two weeks it stopped reporting one metric, because nobody had interacted with it. The metric still mattered and nobody had told the agent to drop it. Every run succeeded. The agent had changed its own priorities.
Requirements drift too
Data drift and model drift are familiar. Requirement drift is harder to see: the system stays consistent while the business starts needing something different. Passing yesterday's evaluations does not prove an agent meets today's requirements. The same applies to memory and persistent instructions: what an agent should keep, update or forget when the business changes.
Match the guarantee to the consequence
An automotive perspective framed reliability by the cost of being wrong. A bad lead suggestion is rejected in seconds. An irreversible action needs a defined operating envelope and a person who answers for it. Not every decision needs the same threshold.
Autonomy depends on the task
For fixed, tightly controlled processes, the table preferred explicit steps and constrained execution. For open-ended work such as lead discovery, loops and autonomy make more sense, because missing a good opportunity costs more than a few weak candidates.
Usage is not the same as value
We compared offline evaluations, golden datasets, guardrails and behavioral signals such as copying, re-prompting and repeat use. An agent that people keep using can still miss a requirement. Usage, satisfaction, correctness and business value are separate measurements. None of them replaces the others.
Access to tools is not adoption
Several organizations already give employees AI tools and still struggle with adoption. One analogy from the table: people do not eat because food is available. They eat because they are hungry. Adoption starts with a concrete problem worth solving, not with a call to become AI-native.
From the evening
