Teams already know where cloud money leaks. The gap is acting on it safely across multi-cloud, hybrid and regulated estates.
Webinar 02 asked why cloud savings stall and whether AI agents can close the gap. Toshal Khawale, Managing Director at PwC, joined Rakesh Rajendran and Shiv Pratap Singh for an hour-long discussion. Toshal framed the problem in one line: the money is lost between knowing and doing.
Shiv said that where visibility already exists, as in cloud operations and Kubernetes, the gap is execution, and savings stall because the team that finds them does not own the change. The speakers came back to three ideas: an agent should raise a pull request for a senior engineer to review rather than change production directly, agents should be narrow and task-specific, and each agent should be measured, including what it costs to run, before it is trusted.
Mostly because the team that finds a saving does not own the change. Shiv said FinOps reports the saving while an SRE or DevOps team owns production. Rakesh said a cost issue is just a ticket, while a failed change lands on the person who made it. So teams take the lowest-hanging fruit and skip the rest.
Mostly an execution problem. Shiv said visibility already exists for cloud operations and Kubernetes, but teams hesitate because they fear what a change will do. Toshal said the money is lost between knowing and doing. Both point to the gap between seeing a saving and acting on it.
Open a pull request. Shiv said he would never recommend that an agent change production directly. The agent should identify the saving and raise a pull request, and a senior engineer should review and merge it. Toshal said the right balance is automation plus a human doing the impact analysis.
Usually the data is too short or too narrow. Shiv said the most common mistakes are reading only seven or 15 days of data when month-end or quarterly jobs run, and cutting memory and CPU while ignoring IOPS. Toshal added that old settings, such as a disruption budget set for one pod, stop fitting after growth.
Scope agents by task. Shiv said a single all-powerful agent widens the security blast radius, because nobody reads its logs until something breaks. He wants agents limited to one job, with the model making only the decision. Toshal argued for purpose-built agents with a human in the loop.
Measure the agent itself first. Shiv said to track what it recommends, how much it saves and how much it costs to run, and to check whether the output is useful. Toshal suggested starting in non-production environments to build confidence. Rakesh said each agent should be metered so it keeps improving.
The money is lost between knowing and doing.
If something goes wrong, it's directly on me. A cost issue is just a ticket.
55:17
Use these 6 guardrails to deploy AI agents to production, giving them write access in stages: read-only first, allow and deny lists, approval for changes, blast-radius limits, tested rollback and an audit trail.
55 min · recorded 23 Oct 2026
New sessions are announced on LinkedIn. Recordings and clips go up on the NudgeBee YouTube channel.