AI Agents Set to Revolutionize Workplaces by 2025, Says Experts

In the future workplace, human workers will closely collaborate with advanced AI agents.

According to McKinsey, generative AI and similar technologies could automate about 60% to 70% of tasks done by employees. Interestingly, around one-third of American workers are already using AI tools at work, often without their employers even knowing.

Experts believe that by 2025, these so-called "invisible" AI agents will step out of the shadows and play a more active role in business operations.

“Agents will likely fit into enterprise workflows just like specialized team members,” says Naveen Rao, VP of AI at Databricks.

AI agents are more than just simple chatbots. They are sophisticated assistants that use foundational models to perform complex tasks that were once thought impossible. These natural language-driven agents can manage multiple responsibilities and, with human guidance, execute them effectively.

“Agents are goal-oriented and can make independent decisions based on context,” explains Ed Challis, head of AI strategy at UiPath. “They will show different levels of autonomy.”

Ultimately, AI agents are expected to perceive, plan, act, reflect, learn from feedback, and improve over time. Raj Shukla, CTO of SymphonyAI, emphasizes this potential.

“At a high level, AI agents aim to achieve the long-awaited dream of automation in businesses,” he notes. Large language models (LLMs) will act as their “planning and reasoning brain,” enabling them to mimic human-like behavior.

However, AI agents are still in their early stages. Many use cases are still being defined and explored.

“It’s going to be a broad spectrum of capabilities,” says Rowan Curran, a senior analyst at Forrester.

The most basic level is what he calls “RAG plus,” or retrieval augmented generation. This system performs actions after retrieving initial data. For example, it can identify a maintenance issue in a factory, outline a procedure, create a draft work order, and send it to the user for final approval.

“We’re already seeing a lot of that today,” Curran adds. “It essentially amounts to an anomaly detection algorithm.”

In more complex scenarios, agents could gather information and take action across multiple systems. Imagine a wealth advisor saying, “I need to inform all my high net worth clients about a recent issue—can you help create personalized emails detailing the impact on their portfolios?” The AI agent would access various databases, run analytics, generate tailored emails, and send them out through an API call to an email marketing system.

Going further, we could see sophisticated multi-agent ecosystems. For instance, on a factory floor, a predictive algorithm might trigger a maintenance request that an agent sends out. This agent would evaluate different options based on cost and availability while communicating with a third-party agent. It could then place orders while interacting with various independent systems and machine learning models.

“That’s the next generation on the horizon,” Curran explains.

For now, though, agents are unlikely to be fully autonomous. Most use cases will still require human involvement for training, safety, or regulatory reasons. “Truly autonomous agents will be quite rare, at least in the short term,” he adds.

Challis agrees, emphasizing that “one of the key points to understand about any AI implementation is that AI alone is not enough.” Businesses will benefit most from a combination of traditional automation, AI agents, and human collaboration.

One common use case for AI agents is onboarding new employees. This process typically involves multiple departments, like HR, payroll, and IT. AI agents could make this process smoother and faster by handling contracts, collecting documents, and setting up approvals.

Consider a sales rep using AI. This agent could work with procurement and supply chain agents to establish pricing and delivery terms for a proposal, explains Andreas Welsch, founder and chief AI strategist at Intelligence Briefing.

The procurement agent would gather information about available products, while the supply chain agent calculates manufacturing and shipping times and reports back.

Or think about a customer service rep asking an agent to gather relevant information about a customer. The agent would consider the inquiry, customer history, and recent purchases from various systems. It would then create a response for a team member to review and edit before sending it to the customer.

“Agents carry out steps in a workflow based on a goal that the user has provided,” says Welsch. “The agent breaks this goal into subgoals and tasks and then tries to complete them.”

While agent frameworks are relatively new, some companies have been using what Rao calls compound AI systems. For instance, FactSet runs a finance platform that allows analysts to query large amounts of financial data to make timely decisions.

The company created a compound AI system that lets users write requests in natural language. Initially, they had a single monolithic LLM, trying to pack as much context as possible into each call. However, this approach hit a quality ceiling with about 59% accuracy and a 16-second latency.

To improve this, FactSet restructured its system into a more efficient AI agent that calls various smaller models and functions, each tailored for specific tasks. After some iterations, the company significantly improved accuracy to 85% while cutting costs and latency by 62% (down to 10 seconds).

Ultimately, Rao notes that “true transformation will come from leveraging a company’s data to build unique capabilities or business processes that give that business an edge over its competitors.”