The first AI systems most people used were conversational. You asked a question, the model answered, and the work of acting on that answer stayed with you. The newest generation behaves differently. Given a goal, these systems can break it into steps, call software tools, check their own results and keep going until the task is done, or until they get stuck. Engineers call them agents, and over the past six months we spent time with companies in Bengaluru, Pune and Gurugram that are putting them to work.
What follows are field notes rather than a forecast: observations from operations floors, engineering stand-ups and a few frank conversations about what goes wrong.
Note 1: The first jobs are the boring ones
Nobody we met had deployed an agent to make strategic decisions. The early wins are in tasks that are tedious, rule-bound and expensive to get wrong by hand: reconciling invoices against purchase orders, chasing missing documents for a loan application, updating inventory records after a shipment arrives damaged. At one logistics company, an agent now handles the first pass on every delivery dispute, gathering photos, GPS traces and messages before a human makes the call.
The time savings are real but uneven. Simple cases close in minutes. Complex ones still land on a person's desk, though with the evidence already assembled. Managers said the biggest change was in morale: staff spend less time copying data between screens and more time on the cases that need judgement.
Note 2: Guardrails are the product
Every team we spoke to spent more effort on limits than on capabilities. Agents are given narrow permissions, a fixed set of tools and clear rules about when to stop and ask. Payments above a threshold need human approval. Emails to customers go through a review queue. Every action is logged so that someone can reconstruct what happened later.
“The model is the easy part. The hard part is deciding what it is allowed to touch, and proving afterwards that it stayed inside the lines.” — Neha Bhatia, head of automation at a Gurugram fintech (fictional, for this demo)
That emphasis on control is partly about risk and partly about trust. Employees are more willing to hand work to a system when they can see what it did and undo it if needed. Customers, regulators and auditors want the same reassurance.
Note 3: Work is being redesigned, not just automated
Putting an agent into an existing process rarely works well. The companies seeing results redesigned the process around it. A claims team at an insurer split its workflow into stages, with the agent handling intake and evidence-gathering and adjusters focusing on assessment and customer conversations. Roles changed; titles, so far, mostly did not.
This has implications for training. New employees used to learn a job by doing its simplest parts first. If agents absorb those tasks, organisations need other ways to build expertise. Some are pairing juniors with agents as supervisors, reviewing and correcting the system's work as a way of learning the domain.
The India angle
India's large services sector, built on processing documents, transactions and customer requests for clients around the world, is both exposed to this shift and well placed to lead it. Firms that understand these workflows deeply can build better agents than outsiders, and several outsourcing companies are already selling agent-based services alongside, or instead of, traditional headcount.
Note 4: The economics are still being worked out
Agents are not free. Each step can involve several calls to a large model, and a complex task may take dozens of steps. Teams track cost per completed task as closely as accuracy. Some have found that a smaller, cheaper model handles most steps well, with a larger one called only when the task gets difficult.
Pricing for customers is shifting too. Instead of charging per user or per hour, some vendors now charge per outcome: per invoice reconciled, per claim triaged, per document verified. That aligns incentives but requires both sides to agree on what counts as done.
When agents fail
Every team had stories of failure. An agent that confidently filled in a missing field with plausible but wrong data. Another that retried a failing step dozens of times, running up costs before anyone noticed. The common lesson was to design for failure from the start: cap retries, require evidence for every claim the agent makes and route anything unusual to a person quickly.
What India can do differently
The agentic economy is arriving everywhere at once, but its shape is not fixed. India has advantages worth pressing: a vast pool of people who understand complex service workflows, digital public infrastructure that makes identity, payments and documents machine-readable, and a growing ecosystem of models trained on Indian languages.
The risk is that agents are used only to cut costs, hollowing out entry-level jobs without creating new paths into skilled work. The opportunity is to use them to extend services to people and businesses that were too small or too remote to serve profitably before. Which of those outcomes dominates will depend less on the technology than on the choices companies and policymakers make in the next few years.
Kabir Ujjwalan