Rani Radhakrishnan says AI managed services must measure whether systems deliver their intended business outcomes, rather than only keeping them running.
2
Agents need standardized processes, better data, centralized governance, careful evaluation, and human oversight before organizations can scale them safely.
3
Rani recommends starting AI work in forgiving back-office processes, then expanding into core products as teams build skills and confidence.
Summary
Rani Radhakrishnan describes managed services as the support work that begins when software reaches production. At PwC, her team applies automation and AI to that work, using models to reduce alert noise, predict infrastructure problems, and support proactive operations. She says agents add a harder layer because organizations must standardize processes, improve data quality, evaluate outputs, and decide where humans approve every result or only exceptions. AI managed services therefore includes tuning models, reducing hallucinations and bias, and measuring business outcomes after deployment. Rani argues that scaling depends less on writing code than on standardizing processes, involving business experts, managing change, and maintaining a feedback loop. She also gives sustainability a wider meaning that includes environmental costs, infrastructure and review costs, accessibility, localization, transparency, and explainability. Her advice is to begin with back-office use cases, invest in training, use centralized governance, and measure results against a baseline.
AI managed services begins when software reaches production
Rani defines managed services as taking responsibility for systems after a project goes live. That includes support for the hardware, software, and user experience. Her AI managed services team adds data analytics and AI to this operational work. She says the goal is not only to fix incidents and keep systems available. The team must also check whether an AI system delivers the outcome it was meant to deliver. That requires people who can tune and optimize models, including MLOps engineers, data scientists, and advanced analytics specialists.
Models can turn noisy IT alerts into proactive operations
Rani describes organizations receiving large numbers of alerts, many of which are noise. A trained model can identify the smaller share that deserves attention, making incident triage easier and reducing mean time to resolution. She also gives a predictive infrastructure example. Resource use can follow cycles, so a monitoring system might identify an upcoming spike and prompt a team to add memory before a server crashes. She sees this combination of MLOps and AI operations as a practical way to use operational data before an incident occurs.
Agents require standardized processes and usable data before they can scale
Rani says agents are difficult to introduce when employees follow different processes or when an organization's data quality is poor. Data modernization and application modernization can create a better foundation. Once processes are documented, teams can choose the ones used most often and get more value from automating them. She supports centralized agent management so departments do not independently adopt incompatible technologies. At the same time, she says parallel development can make sense when several departments have common, frequently used processes.
Agent development spends more time on design and evaluation than coding
Rani says low-code tools and integrated development environments make it easy to create an agent. The larger effort is deciding what the agent should do, designing its workflow, evaluating its outputs, and measuring its impact. Teams need to look for hallucinations and bias after the agent is built. She compares the current stage of agent adoption with early MLOps, when teams were still trying to put models into production and prove their value. In her view, most organizations are still building and evaluating rather than operating large numbers of agents.
AI managed services must tune systems after deployment
Traditional managed services focuses on break-fix support and incident management. Rani says AI managed services has a different responsibility because an agent can execute a process while still producing poor results. Human oversight remains necessary until a use case has shown that it can work without it. She distinguishes human-in-the-loop approval of workflows, human-on-the-loop review of exceptions, and human-out-of-the-loop operation. Highly regulated industries are more likely to approve every result.
Business experts must shape the agent's context and decisions
Rani uses healthcare authorization as an example of why engineers cannot define an agent alone. A system may help with a manual pre-authorization process, but a business specialist must check whether an AI denial follows the right policy. Engineers tend to see system calls and data flows, while business users understand the real-world meaning of a decision. She also mentions Epic Systems adding AI features such as drafting patient emails. Those drafts can save time, but people still need to edit them for tone and context.
Scaling depends on process adoption and measurement more than code
Rani says coding is usually the smallest part of developing an agent once the design is clear. Reusing an automated process in another place can be straightforward, but scaling from one process to ten requires a clear understanding of how the organization works. Standardization helps, while business context still matters because the same process can differ across departments. She says adoption and change management are harder than coding. Evaluation must continue during replication so teams do not spread an agent that produces inconsistent outputs.
Sustainability includes cost, people, and model responsibility
Rani divides sustainability into environmental, economic, and social concerns. Environmental questions include energy use, carbon emissions, water consumption, and electronic waste. Economic questions include the cost of building agents, human review, storage, retrieval, and choosing a smaller model when a large one is unnecessary. Social questions include accessibility, localization, transparency, explainability, and responsible AI. She says organizations are moving from asking where to start with AI toward asking how to scale and sustain what they have already deployed.
Organizations should start with forgiving use cases and a measured feedback loop
Rani recommends beginning in the back office, where mistakes are usually more manageable, before moving AI into core products and services. Organizations should stabilize their current systems, modernize data and applications, standardize processes, train and upskill employees, and keep humans involved. She supports pilots in several departments alongside centralized governance and guardrails. Her final instruction is to measure repeatedly. A baseline makes it possible to compare productivity, cost, and output quality instead of assuming that replacing human work with agents automatically creates value.
"I think the majority of the time needs to be spent designing the agent, then a little bit of time writing the agent and then the rest of the time in evaluating."Rani Radhakrishnan10:26
Who should watch
You are responsible for taking AI or agents from pilots into production and need a clearer view of the operational work that follows.
Your organization has many manual processes, noisy IT alerts, or inconsistent workflows and you are assessing where automation could help.
You need to justify AI spending with business, environmental, or social measures rather than adoption numbers alone.
