How India Inc Pairs People and AI Agents
As AI use widens, enterprise-level impact is still uneven, so the edge now is careful design and accountable oversight
News
- Half of Indian Treasury Teams Put Automation First, EY Report Says
- IIT Bombay's BharatGen Leads India’s Big AI Build
- TCS to Modernise Vodafone Idea’s BSS with AI in 5-Year Deal
- AI Turns Rivals Into Allies as Nvidia Bets $5 Billion on Intel
- WTO Says AI Could Lift Global Trade Nearly 40% by 2040
- Moody’s Puts AI Agents to Work Across Finance, Risk and Strategy
After a successful summit in New Delhi last month, MIT Sloan Management Review India is taking the Strategy Shift Forum to Bengaluru on 25 September. The gathering will bring together MIT professors, global AI experts, and business heads to equip Indian leaders with critical insights into navigating the next wave of AI transformation. For more details, speaker announcements, and to request an invitation, visit here.
Nearly four in five organizations now use AI in at least one business function, up from 72% in early 2024 and 55% a year earlier, with the heaviest use in IT, followed by marketing, sales, and service operations, a recent McKinsey survey showed.
The bottleneck is no longer adoption but converting deployments into outcomes that leaders can measure, a gap McKinsey says widens when companies stack models without resetting decision rights, guardrails, and data contracts.
That reality also framed the discussions last month at the Delhi edition of the Strategy Shift Forum, where panelists urged leaders to move from adding more models to redesigning how decisions get made between people and software agents. The message was less about tools and more about operating rhythm, accountability, and trust.
Neil Thompson, Director of the Future Tech Research Project and Principal Investigator at MIT’s Initiative on the Digital Economy and a visiting professor at Harvard’s Laboratory for Innovation Science, sketched the practical limits and promise of autonomy.
He described “autonomous” and “agentic” AI as a continuum that has been building for a decade, and then used a vivid office experiment to show how brittle fully unsupervised systems can be.
Anthropic hooked a vending machine to its model and let it run procurement and pricing. Before long, the system began ordering obscure items, including “blocks of pure tungsten,” and mispriced them so badly the mini-business “went bankrupt.”
The parable lands: autonomy multiplies decision points, so small errors compound unless there are guardrails and tools in the loop.
Thompson’s core caution is mathematical. If each step in a multi-step task is 95% accurate, two steps drop you to about 90% and three to roughly 86%. Longer chains decay fast, which is why agents should decompose problems, check their own work, and call external tools for brittle tasks like exact arithmetic.
That’s also why promises to automate month-long workflows should trigger skepticism, while ten-second jobs are within reach.
A recurring theme was sprawl control. Without a portfolio view and design standards, agents multiply across teams, behavior drifts, and value fragments.
The counter is simple, if not easy: classify agents by role, set decision boundaries up front, log what they do, and make autonomy legible so managers can see why an agent took a step and where guardrails kicked in.
That clarity matters most in high-stakes contexts such as medicine, finance, or safety, where a 3% error rate is unacceptable, even if a similar rate would be fine in customer service phrasing.
Across boardrooms the ask is consistent. Executives want actionable frameworks, concrete case studies, and tools teams can use tomorrow. That bias toward applied, measurable, repeatable results is what separates firms that dabble in agents from those that turn them into a dependable second line.
Make it work at scale
In practice, the McKinsey study said, it comes down to three steps. First, set the rules before you scale. Say what an agent can do on its own, what it can only suggest, and who signs off. Second, make your systems talk to each other. Use one layer that lets different agents share context, hand off tasks, and plug into company software without being tied to a single model or vendor.
And third, hire for the work. You need people who turn business needs into clear prompts, operators who run the agent portfolio and handoffs, and designers who decide when a human must step in. Then track the basics such as speed, right-first-time, and cost to serve, the study added.
On the customer front, Indian executives are already rebalancing what people do versus what agents should do.
Dilpreet Singh, Head of Loyalty CRM and Partnerships at ITC Hotels, said that as AI handles routine interactions such as bookings, FAQs and confirmations, the value of human teams rises.
“Our workflows are designed in such a manner that AI handles the mundane, but human touchpoints are reserved for moments that need empathy—delight, disappointment or decision-making,” he said.
If a frequent guest drops off the radar, the system flags it, but a team member calls to reconnect. When a guest faces friction, “no bot can truly soothe the experience.”
The point, he added, is not AI versus humans but a division of labor customers can feel, with agents raising throughput and people restoring trust when it matters.
Transparency underpins all of this. Agents respond and adapt; they do not always wait for explicit commands. Systems therefore need to show why an agent took a step, what alternatives it weighed and when it needs help.
Controls should anticipate hallucinations, data drift and policy changes, and they should evolve as agents learn and regulations tighten.
Data remains both the first mile and the last mile. Agents are only as good as the signals they consume and the actions they can trigger. The practical guidance is to move from use-case pipelines to reusable data products, extend governance to unstructured content and wire access and policy directly into the orchestration layer, the McKinsey study said.
