An AI agent is not a smarter search box. It is software that can read a question, look up the right policy, take an action inside a business system, and send a response, all without a human touching it at any point. That is a meaningful difference from the automation tools most companies have used for the past decade.
Wipro put this to work for its own HR department. The company had policies, documents, and employee data spread across dozens of separate systems. When an employee had a question, the answer took 48 hours on average. After deploying a custom AI agent built with Ema, a US-based startup in which Wipro's own venture arm is an investor, response time dropped to five seconds. The agent handles 50 HR tasks that previously required a person.
That result is striking, but it also comes with a footnote worth reading. Wipro Ventures is a financial backer of Ema, meaning Wipro is both a customer and a part-owner. That does not make the outcome false, but it does mean external verification is limited.
The broader picture from independent research is more complicated. McKinsey finds that 62% of organisations are at least experimenting with AI agents, but only 23% are actually scaling them in any business function. Gartner projects that by 2028, about a third of enterprise software will have AI agents built in, up from less than 1% in 2024. The direction is clear; the timeline is not.
A Harvard Business Review survey of 600 business and technology leaders worldwide found that only 6% of companies fully trust AI agents to handle core business processes. Most restrict agents to routine, low-risk tasks and keep humans directly in the loop for anything that touches finances, customers, or sensitive data. Trust is the actual bottleneck, not capability.
The workforce side is where most companies are underinvested. An EY survey found that 84% of employees are eager to use AI agents in their role, but 56% worry about their own job security at the same time. That is not a contradiction, it is a management problem. When leadership does not communicate clearly about what changes and what does not, anxiety fills the gap. The same EY research found that 85% of desk workers are learning about AI agents on their own, outside of work.
Large companies have noticed this and are moving. Walmart has partnered with OpenAI to build a customised AI certification course for frontline and office staff, part of a commitment of nearly $1 billion to skills training. Danone is spending 100 million euros over six years to retrain all 100,000 of its employees for future roles, including AI skills. Salesforce offered free AI training and certifications publicly through 2025. These are not small gestures; they signal that the reskilling cost is real and that leading employers are treating it as a capital investment, not an HR project.
The nature of the jobs that remain human is shifting. Tasks that involve clear rules, repetitive steps, and known data sources are the first to move to agents. What stays with people is harder to define but easier to recognise: knowing which question to ask an agent, deciding what counts as a good enough answer, handling the exception cases an agent cannot resolve, and managing relationships where trust matters. Workers who understand this transition and position themselves on the design side of AI, rather than the execution side, will be in the stronger position.
For any organisation considering this, the practical sequence matters. Starting with a high-volume, low-stakes process, something like internal policy queries or invoice routing, gives a realistic picture of what an agent can actually do in your specific environment before you commit to wider deployment. Data governance and access rules need to be set before the agent goes live, not after. And staff need to know what is changing and why, in plain language, before the technology arrives.