What AI Agents Can Do: Real Examples and Tasks Still Hard to Delegate

The features and cautions in this article were checked against each vendor's official documentation and Japanese government guidelines on October 2, 2026. Specifications may change. Check the sources at the end for the latest details.

What an AI agent can do depends on the tools it can use. With a search tool, it can research. With a browser-control tool, it can look at a screen, click and type. It cannot do work for which it has no tool. To understand what an agent can do, look less at how smart the AI is and more at what it is connected to.

AI agents handle five types of work

The AI Guidelines for Business (version 1.2), issued by Japan's Ministry of Internal Affairs and Communications (MIC) and Ministry of Economy, Trade and Industry (METI), define an AI agent as an AI system that senses its environment and acts autonomously to achieve a specific goal. Anthropic describes AI agents as systems in which an AI model directs its own steps and tool use while it works.

Work you can hand to an agent is easier to organize when split into the five types below. Real AI agents combine them to complete a single job.

1. Research: find and gather information from the web or internal documents
2. Write: produce emails, reports, summaries and other text
3. Route: sort inquiries or documents by type and pass them to the next owner or process
4. Operate: drive browsers, files and external services
5. Review: check results from another angle and fix them before handing them over

As an example of routing, Anthropic describes sorting inquiries into general questions, refund requests and technical support, then sending each to a different process. Reviewing is a setup where one AI evaluates another AI's output and has it revised.

Anthropic explains that agents start work from a person's instruction or conversation, plan and act once the task is clear, and ask people for information or judgment when needed. Deciding where the agent should stop is the user's job.

Research: web search and internal document search

Research is a typical agent task. Anthropic treats an AI model augmented with search, tools and memory as the basic building block of an agent. It says current models can write their own search queries and choose which tools to use.

• Collect competitors' service details from several sites and line them up on the same criteria
• Find the parts of internal rules or manuals that relate to an inquiry
• Summarize a business partner's public information as material before a sales meeting

RAG (retrieval-augmented generation) is often used to search internal documents. The appendix to the AI Guidelines for Business says RAG is expected to reduce answers that are not factual (hallucinations) and to make the basis of an answer transparent by showing its references. It reduces errors but does not eliminate them. Have the agent output its sources too, and make sure people can open the original documents.

Write: emails, reports and summaries

Many companies already use AI for writing. According to MIC's 2026 White Paper on Information and Communications in Japan, the task type for which Japanese companies most often reported using generative AI was help with meeting minutes and emails, at about 70%.

With an AI agent, you can also hand over the work before and after writing. Anthropic gives examples such as drafting a document outline, checking that it meets the criteria and then writing the body, or writing marketing copy and then translating it into another language. Splitting the work into steps tends to be more accurate than asking for everything at once.

• Pull decisions and action items from a meeting transcript and turn them into minutes
• Draft a reply from an inquiry and its past support history
• Collect weekly figures and compile them into a report in a fixed format

Operate: browsers, files and external services

A key difference between an AI agent and a plain chat is that an agent can operate outside systems. Anthropic explains that tools let Claude interact with external services and APIs. Its examples include retrieving customer data and order history, processing refunds and updating tickets.

Operating a browser or computer

Claude's computer use tool takes screenshots and operates the mouse and keyboard. One tool set provides 17 tools, including screenshot, click, type and zoom. Actions run inside an environment that the company using it provides. For work that stays within web pages, the documentation recommends the browser use tool, which reads and operates pages directly.

• Enter rows from a spreadsheet one by one into an internal system with no API
• Open several sites and check specific items
• Open files, rename them and save them to a set folder

The appendix to the AI Guidelines for Business says AI agents can automate coordination and analysis that used to rely on people, by connecting with multiple systems and applications and making decisions based on the situation.

Work that is still hard to delegate, and why

As capabilities grow, both vendors and the Japanese guidelines spell out what needs care. Claude's computer use documentation lists limits such as the following.

• Speed: it can be slower than a person working directly. It suits work where speed is not critical
• Click position: it can get coordinates wrong
• Tool choice: it can act in unexpected ways. Reliability drops with less common apps or when handling several apps at once
• Spreadsheets: complex operations may take several attempts
• Social media and similar: actions such as creating accounts or creating and sharing posts are limited

The same documentation asks users not to use it without human oversight for tasks that require perfect accuracy or involve sensitive personal information. It recommends having a person confirm decisions with real-world consequences, such as payments, agreeing to terms of service or accepting cookies.

Unintended actions and data leaks

The appendix to the AI Guidelines for Business notes that an AI agent acting autonomously could order products or delete files that no one intended. It also warns that, while connecting with external systems, an attack could manipulate the agent's behavior and send internal data outside.

Claude's documentation also warns that the model may follow instructions written in web pages or images. As countermeasures, it lists running the agent in a dedicated environment with minimal privileges, not giving it sensitive data such as login credentials, and limiting the sites it can access. Anthropic notes that agents tend to cost more and that errors can compound, so it recommends extensive testing in sandboxed environments.

For how to prevent incidents, see AI agent risks.

How to map these capabilities to your work

Anthropic recommends starting with the simplest possible setup and adding complexity only when it clearly improves results. Apply the same idea to your work: start small and widen the scope.

1. Pick one job you want to delegate and write it out as steps
2. Decide which of the five types each step falls into
3. Set the points where a person checks, such as before anything leaves the company or money moves
4. Test on a small scale, separate from production data and accounts
5. Record how often you corrected the AI and how long it took, then decide how much to delegate

Anthropic says agents work best on tasks with clear success criteria, results that can be checked and fixed, and room for human oversight. It names customer support and software development as leading examples. For uses by department, see AI agent use cases.

You can also adopt a ready-made AI agent instead of building one. Employee Store is a marketplace where companies can adopt AI agents (AI employees) built by developers, with a one-time purchase or a monthly plan. Listing pages let you compare the job, supported tools, deliverables and price. For the idea behind AI employees, see What is an AI employee.