Generative artificial intelligence has been loud in promise but quieter in everyday office impact. Workers are experimenting with chatbots for jobs like drafting emails, and companies are running many trials, yet white-collar work has not been fundamentally remade.
The next step may not be a better chat window. It may be AI agents: systems that can use other software, reach outside services, and take action through a computer or the internet.
Why chat alone has limits
Chatbots such as Google’s Gemini and OpenAI’s ChatGPT are usually built around a simple exchange: a user types, and the system replies with text. That setup can be useful, but it leaves the model separated from many of the systems where office work actually happens.
Business tasks often require more than producing words. A worker may need to look up customer information, run a calculation, fill in a form, change a user interface, or move between different applications. A chatbot confined to text can advise on those steps, but it generally cannot perform them on its own.
That is why tool use matters. If an AI system can access the right software at the right moment, it can become more than a writing assistant. It can become a helper that participates in a workflow.
Anthropic’s push to give Claude tools
Anthropic, a competitor to OpenAI, announced a major new product that lets developers direct Claude to use outside services and software. The idea is straightforward: when Claude needs a tool to complete a task, developers can require or allow it to use one.
The examples show why this matters. Claude can use a calculator for math problems that large language models may struggle with. It can be required to access a database containing customer information. It can also be compelled to use other programs on a user’s computer when that would help complete the job.
Anthropic has also been working with companies on Claude-based helpers for workers. Study Fetch, an online tutoring company, has developed a way for Claude to use different features of its platform to modify the user interface and syllabus content a student is shown.
That example is important because it shows the difference between answering a question and changing a working environment. The system is not only generating text; it is affecting what a user sees and how a service behaves.
Google and Adept are moving too
Anthropic is not alone. Google demonstrated prototype AI agents at its I/O developer conference earlier this month, alongside other AI products. One prototype was designed to handle online shopping returns.
That return agent would search a person’s Gmail account for the receipt, complete the return form, and schedule a package pickup. Google has not launched that return-bot for broad public use, and the cautious pace reflects a central challenge: AI agents can make mistakes while trying to complete multistep tasks.
Adept AI is also focused on office work. The company was cofounded by David Luan, formerly VP of engineering at OpenAI, and has been honing AI agents for office work for more than a year. Adept has not shared many details about its customers or the specific work its agents perform, but its approach is centered on controlled deployment.
“Our agents are already in the 90s [percent] for reliability for our enterprise customers,” Luan says. “The way we did that was to limit the scope of deployment a bit. All the new research we do is to improve reliability for new use cases that we don't yet do well on."
Luan also says Adept wants its agents to understand the goal of a task and the steps needed to achieve it, rather than simply copying existing human behavior. In his words, “They need to understand the reward of the actual task at hand,” and “Not just have the ability to copy existing human behavior.”
The reliability problem
The main obstacle is not whether an AI agent can click, search, calculate, or fill in a form. The harder issue is whether it can reliably understand what it is being asked to accomplish and choose the correct chain of actions.
Large language models do not always identify a user’s goal correctly. They can also make wrong assumptions that break the sequence needed to finish a task. In a normal chat, that may produce a flawed answer. In an agentic workflow, it can mean the system takes the wrong action inside business software.
That is why early agents may be most useful when they are limited to a specific task or role. The source article compares this to physical robots, which are often deployed in carefully controlled settings to reduce the chance of failure. The same logic applies here: a narrow workflow can make the system easier to guide, test, and contain.
- Limited scope can reduce the risk of unexpected actions.
- Clear workflow boundaries can make performance easier to evaluate.
- Tool access can make agents more useful than text-only chatbots.
Why mundane work could be valuable
The most immediate opportunity may be routine office work. Some large companies already automate common tasks through robotic process automation, or RPA. That often means recording human workers’ onscreen actions, breaking them into steps, and having software repeat those steps.
AI agents built on large language models could expand what can be automated because they are not limited to replaying a fixed sequence. They may be able to interpret goals, use tools, and adapt to different software tasks inside a workplace.
The money involved is already significant. IDC, an analyst firm, says the RPA market is worth $29 billion. It expects AI to more than double that market to around $65 billion by 2027.
This is why mundane work matters. Tasks like finding a receipt, completing a form, checking a database, or using a calculator are not glamorous, but they sit inside real business processes. If AI agents become reliable enough in controlled settings, they could make chatbots far more useful in the places where office work actually happens.