Are AI Agents Useless Failures or the Key to AI’s Future?

Are AI agents the future of artificial intelligence or an overhyped technology that nobody actually wants? I think the answer is both.

Let’s take a closer look at the tension happening right now.

The Agentic AI Tension

A recent WIRED article explored why normal people aren’t using AI agents. Despite enormous enthusiasm in the technology industry, agents have nowhere near the adoption of chatbots like ChatGPT.

The article’s argument is that the industry may be solving a problem that most people don’t have. For many everyday tasks, a chatbot is already useful enough. People may not need an AI system that can independently plan, use multiple tools, and take actions on their behalf. And when agents do offer those capabilities, they can introduce new complexity and friction that isn’t worth the payoff.

In other words, we’ve focused heavily on what agents can do without asking enough questions about what people actually want them to do. I think that’s a fair criticism, but it doesn’t mean the promise of agents is wrong.

Consider the perspective of Andrew Ng. A Stanford adjunct professor and pioneer in machine learning, Ng led the Google Brain team and served as chief scientist at Baidu. He’s also the founder of DeepLearning.AI, chairman and co-founder of Coursera, and founder of AI Fund, which has raised more than $370 million.

Ng sees agentic AI as the next big opportunity in putting AI to work. He argues that enterprises  and investors should focus less on building ever-more-powerful (and expensive) foundation models and more on building valuable applications using agentic workflows. Agents can break complex tasks into smaller steps, use tools, reemember and reflect on their work, and collaborate in ways that can make AI substantially more useful for real-world work. In a recent Masters of Scale interview, Ng said this:

A lot of the work that lies ahead is to take these amazing agentic AI capabilities and map them to real business workflows.

Both perspectives can be right. I see the problem as less whether AI agents are capable and more whether we’re applying those capabilities in the right places and giving agents what they need to be useful.

Let’s take a closer look at what needs to change.

3 Changes Needed to Realize Agentic AI’s Promise

I believe the approach to agentic AI needs to shift in three ways.

1. Focus on Useful Work for Organizations, Not Individuals

AI agents can’t take out my trash, fold my laundry, or feed my dog. I have limited personal need for an agent to perform most of the tasks in my life.

Organizations are different. Every organization has thousands of repetitive, complex, knowledge-intensive, or interconnected tasks. There are opportunities for agents to research, analyze, plan, coordinate, create, and take action.

Cleveland Clinic offers a good example. In a conversation with IBM about AI in healthcare, Cleveland Clinic CEO and President Tom Mihaljevic described several areas where the organization sees opportunities for AI, such as helping patients receive care more seamlessly and helping healthcare executives run an extraordinarily complex business more efficiently. He then offered a striking example of the problem agentic AI can help solve.

Close to one-third of our time when it comes to the practice of medicine is used in activities that have nothing to do with actual patient care. It is mostly either entering or trying to retrieve data from the electronic medical record. If we were to facilitate that with artificial intelligence, it is obviously better for patients. It is also better for providers.

That’s a very different starting point for AI than “What cool stuff can an AI agent do?” Start with a costly organizational problem and ask whether AI can give skilled people more time for work that matters.

Marriott CEO Anthony Capuano describes a similar philosophy. Speaking about AI and technology at a Morgan Stanley event, he emphasized why efficiency created by AI-enabled automation is valuable:

We look at technology as a creator of capacity for our associates. Every minute that they’re not spending on older technology… creates capacity for them to better engage our guests.

Let’s consider another shift for agentic AI.

2. Make Agents Easier to Set Up, Use, and Govern

Organizations need agents that are easier to set up, tailor, use, and manage. They need strong onboarding and user enablement. They need permissions, guardrails, and ways to understand what agents are doing.

ServiceNow Chairman and CEO Bill McDermott made this point at the company’s Knowledge 2026 conference, where he described the transition enterprises are facing as a move from “AI chaos to control.” He argued that as companies deploy more agents, they need the infrastructure to bring AI, data, workflows, and governance together.

What we’re seeing is bigger than AI, bigger than software. The world of work is being remade.

That statement came as McDermott was describing ServiceNow’s vision of an agentic business, in which AI agents become part of the workforce and take on increasingly autonomous work. But autonomy also increases the stakes. ServiceNow’s current approach emphasizes enterprise context, permissions, workflow constraints, and governance because a more capable agent can also create more significant risks when it operates without those controls.

And then there’s the issue of cost. While they may cost less to build and implement, AI agents can cost a lot to maintain. Agentic AI can be much more computationally intensive than a simple chatbot interaction. An agent may reason through multiple steps, call tools, maintain context, generate additional prompts, or even invoke other agents. As agentic workflows scale, so can costs like token consumption.

We’ve entered what some have jokingly called the “Tokenpocalypse,” but the underlying issue isn’t funny for organizations trying to manage AI budgets. Governance has to include economics as well as security and risk.

In fact, ServiceNow’s current AI Control Tower explicitly includes financial dashboards designed to help organizations monitor and control AI spending alongside agent behavior, risk, compliance, and performance. That’s the right direction.

Now let’s turn to one more shift needed in agentic AI.

3. Connect Agents to Deep Knowledge and Expertise

An agent can have impressive capabilities, but capabilities alone don’t tell it what matters to your organization. It needs context like your business knowledge, standards, processes, data, goals, and expertise.

Mayo Clinic Platform President John Halamka made this point particularly clearly in a June 2026 article about the need to rearchitect healthcare data for an AI-enabled future. Writing about the need to make healthcare data usable for both people and AI, he said:

It is increasingly apparent that we will not achieve this without significant changes globally to how we make healthcare data accessible in real time or near real time for both humans and AI agents. Every data point in healthcare should serve one purpose: improving outcomes and developing more cures for patients. As healthcare generates more data than at any point in our history and AI becomes more capable of reasoning across it, we must structure data so it is usable for discovery at scale, something not currently possible in most public or private healthcare systems.

That’s a powerful way to think about the challenge. AI may be the exponent, but the organization’s knowledge and data are the foundation.

The Home Depot is putting a similar principle into practice. At NRF 2026, the company announced new agentic AI capabilities designed to combine its home-improvement expertise with real-time inventory, product locations, and customer context. Jordan Broggi, EVP of Customer Experience and President–Online, described the goal this way:

We’re putting ‘Orange Apron’ expertise in the pocket of every customer and creating an AI experience that is personalized, contextual, and available wherever the customer is—whether that’s the home, the jobsite, or in the aisles of our stores.

The same idea applies across industries. In a 2025 McKinsey interview about AT&T’s strategy and its use of AI, CEO and Chairman John Stankey described the company’s opportunity to use AI while emphasizing that its proprietary knowledge is what can create differentiation:

These are some of the things that bring a much more satisfying experience to our customers, and our competitors can probably do these things equally well. Where I think we really need to up our game is using data that’s specific to AT&T for our strategic advantage. I think there are unique opportunities for us to use our data around pricing, our knowledge of markets, our awareness of what infrastructure is in place, and so on.

The foundation models may be broadly available to everyone. An organization’s proprietary knowledge isn’t.

This is where I think Content Science has a particularly important perspective to bring to the AI conversation. We see content and knowledge as data—and, in most organizations, a huge amount of that data is unstructured. It’s embedded in documents, websites, research, guidelines, presentations, processes, policies, and the accumulated expertise of people across the organization. That unstructured data may be messy and difficult for traditional systems to use, but it contains treasure of what the organization knows.

As AI agents become more capable of reasoning and taking action, the question becomes how well we can make that knowledge accessible, trustworthy, and useful to them. That’s why high-quality organizational content becomes increasingly valuable in an agentic world. Content isn’t just something people consume anymore. It can provide the context agents need to make better decisions and perform useful work. But that only happens when the knowledge is accurate, structured, governed, accessible, and connected to the organization’s expertise. Content is becoming crucial infrastructure for enterprise AI.

Putting the Principles of Change into Practice: The ContentIQ Example

The changes above aren’t just ideas I have about where agentic AI needs to go. They’re also the principles the Content Science team and I used to build ContentIQ, our new superagent developed in partnership with Rival.io. We designed ContentIQ specifically around the kinds of organizational problems where content, knowledge, strategy, and AI intersect.

First, ContentIQ focuses on organizational work. It is designed to help organizations address content strategy, systems, and operations. These interconnected activities that can consume enormous amounts of time and require people to navigate complex information and decisions.

Second, we’ve tried to make the experience approachable rather than assuming organizations already know how to work with agents. ContentIQ offers both step-by-step onboarding and a conversational interface, allowing people to choose how they want to engage with it. The goal is to lower the barrier to adoption while still giving organizations the structure they need to use AI responsibly.

Third, ContentIQ puts specialized knowledge at the center of the experience. It incorporates Content Science’s proprietary research, data, frameworks, methodology, and expertise. Organizations can also add their own specialized knowledge, standards, data, and other sources of truth.

Finally, we’ve considered governance and economics as part of the product rather than as afterthoughts. ContentIQ is offered as a subscription with no surprise charges because organizations need to be able to understand and manage the cost of putting AI to work.

If you’re curious about ContentIQ, join us for an upcoming webinar walking through what it is, why we built it, and how it works:
Register for Introducing Content IQ >>

If more agentic AI solutions take a similar approach, the promising future of AI agents can happen now.