What Is Agentic AI? How AI Agents Work, and Where They Fall Short

What Is Agentic AI?

Agentic AI is artificial intelligence that can act on its own to reach a goal. A normal chatbot answers your question and stops. An AI agent takes a goal, such as “find three flights under $400 and add the best one to my calendar”. It works out the steps, uses tools to carry them out, and checks its progress along the way.

The word “agentic” comes from agency, the ability to act independently. In practice, today’s agents have limited agency: they work inside the tools and permissions people give them, and the best-designed systems still ask a human before doing anything important.

Most AI agents are built on a large language model, the same technology behind chatbots like ChatGPT. What makes them agents is what is wrapped around the model: access to tools, memory of what has happened so far, and a loop that lets the model decide its next step.

Chatbot vs. AI Agent: What’s the Difference?

The difference is who does the work between steps. With a chatbot, you read each answer and decide what to do next. With an agent, the AI makes many of those decisions itself.

That last row is the important one. A chatbot that makes a mistake gives you bad text. An agent that makes a mistake might send an email, delete a file or spend money.

How AI Agents Work: The Agent Loop

Most agents follow the same basic cycle, repeated until the job is done.

1. Plan. The language model reads the goal and everything that has happened so far, then decides the next step.
2. Act. It calls a tool: searching the web, reading a document, running code, filling in a form or sending a message.
3. Observe. The tool’s result is fed back to the model.
4. Check. The model judges whether the goal is met. If not, it plans the next step and the loop continues.

Well-designed agents also have stopping rules, such as a maximum number of steps or a spending limit, so a confused agent can’t run forever.

Workflows vs. Agents

Not every AI system that uses tools is a true agent. AI developer Anthropic draws a useful line between two designs.¹ In a workflow, language models and tools follow predefined code paths that a developer has written: step A, then step B, then step C. In an agent, the model itself directs the process, deciding which tools to use and in what order.¹

Workflows are more predictable and easier to test. Agents are more flexible, which suits open-ended problems where you can’t predict how many steps will be needed.¹ Anthropic’s advice to builders is to start with the simplest solution possible and add complexity only when it is needed.¹ It notes that agentic systems often trade higher cost and slower responses for better results on hard tasks.¹

For a business, this is practical advice: many tasks marketed as needing an “agent” can be handled by a simpler, cheaper and more reliable workflow.

What AI Agents Can Do Today

Agents work best on tasks that have clear goals, many small steps and results that are easy to check. Common uses include:

• Coding assistants that read a codebase, make changes across several files, run the tests and fix what fails.
• Research assistants that run many searches, read the results and write a summary with sources.
• Customer service agents that look up an order, check the policy and process a simple refund within set limits.
• Back-office work such as matching invoices to purchase orders, updating records or drafting routine reports.
• Browser agents that navigate websites and fill in forms on a user’s behalf.

They are much weaker at tasks where goals are vague, mistakes are costly and hard to spot, or judgement and accountability matter, such as financial, legal or medical decisions.

How Many Businesses Are Using AI Agents?

Use is growing quickly, but most organizations are still early. McKinsey’s State of AI survey, fielded in May and June 2026 with 1,719 participants in 97 countries, found that about two in ten respondents said their organizations had reached the scaling phase with AI agents. Among large companies with more than $1 billion in revenue the figure was 40%, up from 27% a year earlier, while smaller organizations were flat at 22%.³

Research firm Gartner has made bold long-term forecasts. It predicts that 33% of enterprise software applications will include agentic AI by 2028, up from less than 1% in 2024.² It also expects at least 15% of day-to-day work decisions to be made autonomously through agentic AI by then, up from zero in 2024.²

The Hype Problem: “Agent Washing”

Not everything sold as an AI agent is one. Gartner coined the term “agent washing” for vendors that rebrand existing products, such as chatbots, robotic process automation or AI assistants, as agentic without real agentic capabilities. It estimated that only about 130 of the thousands of vendors claiming agentic AI are real.²

The same Gartner research predicts that over 40% of agentic AI projects will be cancelled by the end of 2027, because of escalating costs, unclear business value or inadequate risk controls.² That doesn’t mean agents don’t work. It means many projects start with the technology rather than with a clear problem worth solving.

Before buying, ask a vendor to show the agent completing a real task from start to finish, what happens when it fails, what it costs per task, and which actions require a human to approve.

The Risks of Agentic AI

Giving AI the ability to act creates risks that chatbots don’t have. The main ones:

• Compounding errors. A small mistake early in a long task can snowball, because each step builds on the last. Anthropic flags this as a key risk of agents that run for many turns.¹
• Prompt injection. An agent that reads web pages, emails or documents can be tricked by hidden instructions inside them, such as text telling it to forward private data. Prompt injection tops the OWASP list of security risks for applications built on language models.⁵
• Too much access. An agent with broad permissions to email, files or payments can do broad damage if it goes wrong or is manipulated.
• Cost. Agents can make many model calls per task, so costs can be much higher and harder to predict than a single chatbot answer.¹
• Accountability. When an agent acts, someone still has to be responsible for the outcome, and that needs to be clear before it goes live.

The US National Institute of Standards and Technology (NIST) publishes an AI Risk Management Framework that many organizations use to structure these questions.⁴ It covers who is accountable, how risks are measured and how they are managed over time.

How to Use AI Agents Safely

The same few rules apply whether you are trying an agent at home or deploying one at work.

1. Start narrow. Give the agent one clearly defined task with an easy way to check the result.
2. Grant the least access needed. Read-only where possible; no payment or deletion rights unless essential.
3. Keep a human in the loop. Require approval before sending, buying, deleting or publishing anything.
4. Test in a safe space first. Anthropic recommends extensive testing in sandboxed environments with appropriate guardrails.¹
5. Watch for hidden instructions. Be cautious about letting an agent act on content from untrusted web pages or emails.
6. Keep logs and limits. Record what the agent did and cap steps, time and spending.

These are the same habits that make any AI tool safer: treat its output as a draft until a person has checked it.

A Simple Example: An Agent Handling a Refund

Here is what an agent’s loop looks like on a real, everyday task. A customer emails an online shop: “My order arrived damaged, can I get a refund?” A customer-service agent with read access to orders and limited refund rights might work like this:

1. Plan: find the order linked to the customer’s email address.
2. Act: look up the order in the shop’s system; observe that it was delivered three days ago for $48.
3. Plan: check the refund policy for damaged items.
4. Act: read the policy; observe that damaged items under $100 can be refunded within 30 days with a photo.
5. Act: reply asking for a photo; when it arrives, prepare the $48 refund.
6. Check: because the refund moves money, pause for a staff member to approve, then confirm to the customer.

Each step is small and checkable, the agent’s permissions are limited, and the one action that spends money needs a person to approve it. That is what a well-designed agent looks like in practice. The example is illustrative; real systems vary.

Will AI Agents Replace Jobs?

Agents are more likely to change jobs than to remove them wholesale, at least in the near term. They take on chunks of routine work, such as gathering information, filling in systems and drafting first versions, and leave people to set goals, handle exceptions and make judgement calls.

Gartner’s forecast that 15% of day-to-day work decisions will be made autonomously by 2028 gives a sense of scale: meaningful, but still a minority of decisions.² The skills that grow in value are knowing what to ask for, checking the work, and understanding the business problem well enough to judge whether an agent’s result is right.

How to Try an AI Agent Yourself

The easiest way to understand agents is to watch one work on a low-stakes task. Many AI assistants now offer a research or “agent” mode that searches, reads and reports back. Try a task you could check yourself, such as comparing the features of three products or planning a day trip, and read the steps it shows you.

Notice where it does well and where it guesses. Keep personal, financial and login details out of these experiments, and don’t let an agent buy, send or delete anything until you trust how it behaves.

Key Terms Explained

These are the terms you will see most often in discussions of AI agents.

The Bottom Line

Agentic AI turns AI from something that answers into something that acts. That makes it genuinely useful for multi-step, checkable tasks, and genuinely riskier when it has access to things that matter. Expect real progress and plenty of hype at the same time.

Agents are built on generative models. Our explainers on generative AI and prompt engineering cover the basics.

Start small, keep a human approving important actions, and judge any agent on the work it actually completes. To understand the technology underneath, read how large language models work, or browse more AI tools explainers.

Keep Reading

Tech & AI

What Is Retrieval-Augmented Generation (RAG)? How It Works

Retrieval-augmented generation (RAG) lets AI look up trusted documents before it answers. Learn how RAG works,…

Tech & AI

What Is a Password Manager? How It Works and Is It Safe?

A password manager creates, stores and fills strong, unique passwords for every account. Learn how it…

Tech & AI

ChatGPT vs Claude vs Gemini: How the Big AI Assistants Compare

ChatGPT vs Claude vs Gemini compared on plans, prices, features, strengths and privacy in late 2026,…