For the last several years, much of the discussion around artificial intelligence has focused on what AI can create. We have watched systems learn to write reports, generate images, summarize documents, produce computer code and answer questions in seconds.
That alone has been enough to change the way many people work.
But the next development in artificial intelligence may be considerably more significant.
Instead of simply answering a question or generating something when prompted, a new generation of AI systems is being designed to take a goal and work toward accomplishing it. These systems can plan a series of steps, use different software tools, evaluate what happens along the way and make additional decisions without requiring a human to provide instructions at every stage.
They are generally described as AI agents, or agentic AI.
And while there is no shortage of hype surrounding the technology, agentic AI represents an important change in the relationship between people and computers. The computer is beginning to move from being a tool we operate toward something closer to a digital collaborator that can perform work on our behalf.
From Answers to Actions
Traditional software is fairly predictable. A person opens an application, selects a function and tells the software what to do.
Even the first wave of generative AI largely followed this model.
Ask a chatbot to summarize a document, and it summarizes the document. Ask it to draft an email, and it drafts an email. The AI might perform an extraordinarily complicated calculation behind the scenes, but the interaction itself remains fairly simple: a person asks, and the system responds.
Agentic AI changes that relationship.
Imagine asking an AI system to organize a business trip.
A normal chatbot might tell you which flights are available, suggest a hotel and create an itinerary.
An AI agent could potentially search the flights, compare schedules and prices, check your calendar, select options based on your preferences, book the trip, add the reservations to your calendar, create an itinerary and notify the people you are meeting.
The important difference is not simply intelligence. It is action.
Recent consumer systems are already moving in that direction. Meta, for example, introduced an AI assistant in September capable of performing tasks such as sending emails, arranging travel and carrying out transactions rather than simply providing information.
The same transition is taking place inside businesses.
The Digital Coworker
Some of the earliest applications of agents are appearing in software development.
An engineer can increasingly give an AI coding system a broader objective instead of asking it to generate a few lines of code. The agent can examine an existing codebase, determine which files need to change, write code, test the changes, identify errors and try again.
That is very different from autocomplete.
The same principle can apply to many types of knowledge work.
An AI agent could research potential suppliers, gather information from different databases and prepare a comparison for a purchasing team. Another could monitor customer support requests, determine which ones can be resolved automatically and route unusual cases to an employee. A financial agent might reconcile information from multiple systems and flag anomalies for a human to review.
Companies are already moving beyond small experiments. McKinsey’s 2026 global survey found that 40 percent of respondents at large organizations reported that their companies were scaling AI agents, compared with 27 percent a year earlier. Coding agents were among the most common applications.
The result could be a gradual change in what a typical workday looks like.
For decades, office workers have spent significant amounts of time moving information between systems, searching for documents, scheduling meetings, filling out forms, preparing routine reports and following up on relatively predictable tasks.
Agentic systems are particularly well suited to this kind of work because many of these activities are not really single tasks. They are chains of small tasks.
The ability to handle the chain is what makes agents interesting.
Changing Workflows, Not Just Jobs
It is tempting to look at every new automation technology and immediately ask which jobs it will eliminate.
That may be the wrong question.
The more immediate effect of agentic AI is likely to be the redesign of workflows.
Consider a sales team.
Today, a salesperson might research a potential customer, review previous correspondence, update a customer relationship management system, prepare material for a meeting, send a follow-up email and schedule the next conversation.
None of those activities necessarily requires a different employee.
But they require time.
An agent that handles much of the preparation and administration could allow the salesperson to spend more time speaking with customers, evaluating opportunities and negotiating.
Something similar could happen throughout an organization.
The employee does not necessarily disappear. The division of labor changes.
Humans may increasingly define objectives, make judgment calls, handle unusual situations and review results, while software performs more of the repetitive steps required to move work from beginning to end.
This could also change management.
Managers have traditionally coordinated work performed by people. In the future, some teams may consist of employees working alongside multiple specialized agents. Deciding which tasks should be automated, what access those agents should receive and when a human needs to intervene could become part of everyday management.
Gartner has argued that this transition will increasingly require cooperation between technology and human-resources leaders because the design of work itself begins to sit between the two functions.
That may turn out to be one of the more important implications of agentic AI.
The Technology Is Only Part of the Problem
There is, however, a substantial difference between demonstrating an AI agent and deploying one across a large organization.
Businesses have spent decades accumulating software.
Customer records may be stored in one system, financial information in another, employee information in another and operational data somewhere else entirely. Some of those systems may be relatively modern. Others may have been built twenty years ago.
An AI agent is only useful if it can access accurate information and interact reliably with the systems required to complete its task.
That is proving to be difficult.
McKinsey reported this year that nearly two-thirds of enterprises had experimented with agents, while fewer than 10 percent had scaled them to produce tangible value. Data limitations were cited as one of the largest obstacles.
Capgemini has described a similar problem. The company says many organizations are now beginning large modernization projects because fragmented legacy systems prevent AI from accessing consistent data across the business.
In other words, the future of artificial intelligence may depend partly on some very unglamorous work.
Companies need clean data.
They need systems that can communicate.
They need clear permissions.
They need reliable automation.
And they need to know where important information actually resides.
Buying access to the latest AI model does not solve those problems.
Who Is Responsible When an Agent Makes a Mistake?
As agents become more capable, another problem emerges: autonomy introduces risk.
A chatbot giving an incorrect answer is inconvenient.
An agent taking an incorrect action could be considerably more serious.
Suppose an AI agent has permission to communicate with customers, approve refunds, access internal documents or initiate financial transactions. A mistake is no longer confined to a screen.
This creates questions organizations will need to answer.
What information can the agent access?
Which actions can it take without approval?
When must it ask a human?
Who is responsible when something goes wrong?
How do you prevent one agent from sharing information with another agent that should never have received it?
The scale of the problem could become substantial. Gartner predicts that the average global Fortune 500 company could have more than 150,000 AI agents in use by 2028, compared with fewer than 15 in 2025. Gartner also reported that only 13 percent of organizations believe they currently have the governance required to manage agent deployments properly.
Whether those exact numbers prove accurate is less important than the underlying issue.
If companies begin creating thousands of software agents, they will eventually need systems for managing those agents just as they currently manage employees, devices, applications and user accounts.
Security will become part of the design from the beginning.
Software May Become Less Visible
Agentic AI may also change something most of us rarely think about: how we interact with software.
Today, work is organized around applications.
We open an email application to communicate, a CRM to manage customers, accounting software to review finances, a project-management platform to track work and perhaps several other systems throughout the day.
Agents could begin to sit above those applications.
Instead of opening four systems to complete a process, an employee might simply tell an agent what outcome is required. The agent could interact with those systems in the background.
If that happens at scale, software itself becomes less visible.
That could eventually challenge the economics of the software industry. Gartner estimates that as much as $234 billion in enterprise application spending could be exposed to what it calls “agentic arbitrage” by 2030 as agents reduce the need for humans to interact directly with multiple applications.
That does not mean business software disappears.
It means the interface may change.
The application remains underneath, but the AI agent becomes the layer through which the user accesses it.
The smartphone changed computing partly because people no longer needed to think about operating a computer in the traditional sense. Touch an icon, and the complexity disappeared behind the screen.
Agents could create a similar abstraction for business software.
We Are Still Early
For all the excitement, there are reasons to remain cautious.
AI agents still make mistakes. They can misunderstand objectives. They can follow incorrect information. They can struggle with long or complicated processes.
Even companies building some of the world’s most advanced AI systems have acknowledged that progress has not always happened as quickly as expected. In July, Meta CEO Mark Zuckerberg told employees that the development of AI agents had been moving more slowly than anticipated.
That is worth remembering.
Technological transitions often look obvious in retrospect and messy while they are actually occurring.
There will be impressive demonstrations. There will also be failed deployments, security problems, expensive experiments and products that never find a practical use.
The organizations most likely to benefit may not be those that deploy the largest number of agents the fastest.
They may be the ones that understand which processes actually benefit from autonomy and which ones still require human judgment.
A New Division of Labor
Artificial intelligence has already demonstrated that machines can generate information that once required considerable human effort.
Agentic AI asks a different question.
What happens when those machines can begin acting on that information as well?
The answer will probably not be a workplace in which humans simply disappear.
A more realistic possibility is a new division of labor.
AI systems will handle more research, coordination, administration and repetitive execution. People will remain responsible for goals, judgment, relationships, creativity, accountability and the many situations in which the correct decision cannot be reduced to a predictable process.
The technology still has a long way to go.
But if agents continue to improve, the important question for many organizations will no longer be whether employees use artificial intelligence.
It will be how humans and artificial intelligence learn to work together.