The experts weigh in

IBM spoke with four experts to get their take on agentic AI, and a better understanding of expectations versus reality. The consensus? AI agents are the hottest thing in the tech industry but confidence in their abilities varies widely. 

“There is the promise, and there is what the agent's capable of doing today. I would say the answer depends on the use case,” said Maryam Ashoori, Director of Product Management at IBM. 

We know that agents can work semi-autonomously. They can reason, prioritize, break down tasks, and make decisions on their own, but they don’t always get it right — especially when it comes to collaboration. Take Microsoft’s recent agentic experiment: a fake marketplace to test the agents’ ability to work together resulted in some issues and failures. 

“Right now, we’re seeing early glimpses — AI agents can already analyze data, predict trends and automate workflows to some extent. But building AI agents that can autonomously handle complex decision-making will take more than just better algorithms. We’ll need big leaps in contextual reasoning and testing for edge cases,” said Vyoma Gajjar, AI Technical Solutions Architect at IBM. 

AI agents in the wild

Despite the realities and setbacks of early-stage technology, companies (and governments) around the world are diving head first into agentic AI. Here’s a small glimpse into how AI agents are infiltrating every industry at the moment. 

AI agents in the IRS

The Internal Revenue Services (IRS) is rolling out Agentforce, Salesforce’s new AI agent system, to help staff with tasks like searching records and summarizing cases. Humans will stay in charge of all decisions, and the IRS says strict guardrails will keep the AI from handling anything sensitive like approvals or payments.

Windows agentic taskbar integration

Microsoft is turning Windows 11 into what it calls an agentic OS, embedding agents directly into the taskbar to run background tasks, access files, and send progress notifications. While users remain in control, this move signals a deeper shift toward seamless, always-available AI support in everyday workflows. 

3D object design with agentic AI

Researchers at MIT developed an AI agent that can use CAD software the way a human engineer would — turning a 2D sketch into a full 3D model by clicking and dragging tools inside the program. The key innovation: a massive dataset called VideoCAD with over 41,000 videos of human actions to help the system learn how to translate design intent into actual user-interface actions. 

Cybersecurity defence teams at Amazon

Amazon has launched a system called Autonomous Threat Analysis (ATA) that uses teams of AI agents acting as attackers and defenders to spot security weaknesses in its infrastructure. Humans still make the final decisions, but the multi-agent setup is already proving effective, achieving 100% detection rates!

Agentic AI collaborative workspace

Felo has launched LiveDoc, the world’s first collaborative AI‑agent workspace for documents, where multiple specialized agents like designers, translators, and analysts work together in real time. LiveDoc transforms static files into living, self-updating knowledge hubs, saving workers up to 30% of their time by automating repetitive tasks.

Is the hype really worth it? 

Let’s dig into what the actual research — not just the headlines — tell us about this new phase of artificial intelligence. 

Upwork’s Human + Agent Productivity Index found that AI agents are most effective when paired with humans. Productivity can increase by as much as 70% compared to agents working alone. This highlights a key point: agents are powerful tools, but they don’t replace human oversight.

When McKinsey looked at more than 50 real-world AI agent projects, they found that while agents can do a lot, many companies still struggle to get real value. The key is good workflow design and keeping humans in the loop. 

IBM’s benchmarking research reinforces this: agents are improving quickly, but they still fall short on complex reasoning and unusual scenarios. Building agents that can reliably handle edge cases will require more advanced contextual understanding and rigorous testing.

Broader research from McKinsey on the future of work emphasizes the need for human-AI collaboration. Most jobs won’t disappear, but the skills required will shift. Professionals who can work alongside AI, orchestrate hybrid workflows, and manage agentic systems will have a clear advantage.

The takeaway? AI agents offer real productivity gains, especially when humans and machines collaborate. The hype is justified in some areas, but the real value comes from leveraging agents thoughtfully rather than expecting them to run autonomously.

Help bridge the AI‑human connection

One key theme that pops up whenever AI agents are the topic of discussion is the human aspect of it all. From Microsoft’s marketplace experiment to industry experts’ insights, the need for an AI - human connection is clear. 

Sure, an AI agent might be able to complete tasks or process data faster than a human, but almost every company experimenting with agentic AI is using humans as guardrails. And this need for human oversight is creating exciting opportunities for AI trainers and enthusiasts to join the movement. 

Toloka, Mindrift’s parent company, recently launched Tendem, an innovative hybrid agent where human specialists deliver tasks with the help of AI agents. 

“How work gets done is being fundamentally redefined. We’re moving beyond the limits of human-only or AI-only workflows toward an integrated model – where AI handles scale and efficiency, and human intelligence ensures accuracy, context, and trust,” said Olga Megorskaya, CEO of Toloka. 

Curious? Meet the experts behind Tendem and learn how you can help power agentic workflows with human context and judgement.