What Is an AI Employee? Workplace Agents Explained (2026)

Updated November 7, 2026.

What is an AI employee?

An AI employee is a plain-language name for workplace software that takes a goal, uses tools, and returns work for a person to review. The more precise name on this site is a digital worker. The working system underneath is usually an agent: a language model plus a harness.

People also say “digital worker,” “AI teammate,” and “employee agent.” The words are ahead of the products. On November 7, 2026, Anthropic’s workplace product is Claude Cowork. OpenAI’s is ChatGPT Work, with shared workspace agents as a separate research preview. A chat reply that only explains the task is still chat. A script that always does the same five steps is still fixed automation.

The software is not a legal employee. It has no duties of its own, and a mistake is still the organization’s to own. Deloitte’s September 10, 2026 essay treats “agents as labor” and “agents as infrastructure” as two choices a company might make, and it tells leaders to watch for moments when agents are treated as the responsible party in a review. That is a warning about accountability, not a feature list.

This guide is for adult work and for university study. School-account policy is a different page: Google’s November 2026 K-12 switch.

What is a digital worker?

A digital worker is the same idea with the job in the foreground. Someone assigns a recurring task, such as a Monday readout or a first-pass contract memo. The software reads the sources it is allowed to read, does the steps, and leaves a draft.

OpenAI’s business guide, a PDF read on November 7, 2026, splits agents into a model, tools, and guardrails. The model interprets the goal and picks the next step. The tools gather information or take an action. The guardrails keep that action inside rules a person set. The PDF has no publication date on the copy read here. It still uses the product name “ChatGPT agent” for the pre-built example, and it names Operator and Deep Research as other examples. Those names are the guide’s. The current workplace page calls the product ChatGPT Work.

Deloitte’s December 10, 2025 Tech Trends essay uses “digital workers” and a “silicon-based workforce” for the same management question: how a company supervises software that can act, beside the people who still handle shifting requirements. The essay also warns about “agent washing,” its name for automation products relabeled as agents. A new badge on an old script does not make an employee.

The September 10, 2026 Deloitte essay is the later workplace piece. It cites a Gartner forecast that the average Fortune 500 company is likely to have 150,000 AI agents by 2028, and a Deloitte study of senior managers in which two-thirds expected multi-agent systems to change how work is done within two to three years, while about a quarter said their processes or workforce were prepared. Deloitte’s 2026 State of AI in the Enterprise survey, cited in that essay, says nearly three-quarters of surveyed organizations plan to use agentic AI at least moderately within two years. A separate Deloitte study cited there puts 93 percent of organizational AI investment toward technology infrastructure and 7 percent toward work and people. Those are Deloitte’s and Gartner’s figures, from that page. They are a reason to write the oversight down. They are not a headcount of digital workers you can hire.

How does a workplace agent differ from a chatbot and from fixed automation?

OpenAI’s business guide draws the line with two everyday systems.

A question-answering chatbot can return an answer and can call a tool. Asked for the travel policy, it can fetch the current file from an HR system. The guide says it still cannot plan the offsite: venues, costs, and an agenda, unless every step is programmed.

A workflow automation follows steps and rules, including simple “if this, then that” logic. The guide’s example is an expense system that stops when a new vendor field is blank, and a login rule that locks an account after three failures and a quiet day. The strength is that every branch was written down. The weakness is a condition the author did not list.

An agent, in that guide, starts from a goal. It plans, picks tools, and can change the plan when a source is missing or a person answers a question. The guide says the three can sit in one process: a fixed rule for the stable part, a model for one judgment inside that rule, and an agent where the path is not known at the start.

The same split is how this site uses the words. A chat reply is a model. An agent is that model in a loop with tools. The loop, the permissions, and the stop rule are the harness. Designing them on purpose is harness engineering.

A shopping agent is the same shape pointed at a purchase. That guide is Can an AI shopping agent check out for you?. A click on a real screen is computer use. Both are narrower than the phrase “AI employee.”

What can Claude Cowork do at work?

Claude Cowork, read November 7, 2026, is Anthropic’s product for research, analysis, documents, and other multi-step work that is not software engineering. The page says you give Claude a goal. It works in the folders and tools you choose. You come back to work for review. Claude Code remains the product for writing, debugging, and shipping code. Cowork uses what Anthropic calls the same agentic approach, aimed at knowledge work.

The page’s examples are ordinary office jobs: a scheduled Monday marketing readout pulled from Amplitude and a Drive tracker, a spreadsheet that flags variances, a first-pass memo on each contract in a folder, and a quarterly read of call notes with quotes tied back to the notes. Those are illustrations on the product page, not a promise that every connector is turned on in every account.

A few boundaries on that page matter more than the demos:

• Cowork is included with a paid plan. The page says it is rolling out on the web and on mobile. The desktop app is what can reach folders and applications on your own computer.
• Every conversation can start as chat. When a task needs files, Claude brings in Cowork’s capabilities. Until that unified experience reaches an account, the page says Cowork still sits next to Chat.
• Deleting anything needs your approval. A permissions setting can make Claude show its plan and wait before anything significant. You can redirect it or stop it.
• The desktop app has a built-in browser for Cowork, separate from your own browser, on Pro, Max, Ultra, and Team. Enterprise admins manage it under Organization settings, then Cowork. Cookie import is described for Chrome, Edge, and Firefox on macOS, and for Firefox on Windows and Linux.
• Computer use, which drives the screen, is a research preview for Cowork and Claude Code. The page says Claude tries connectors first, then a browser, and uses the screen last, asking before each new application. The click-by-click account is the computer-use guide.
• Tasks assigned from a phone keep running in the cloud after the laptop is closed, on Pro, Max, Ultra, and Team. Enterprise plans can opt in. Agentic tasks use more of the usage allowance than a short chat, because they coordinate tools and sub-agents.
• Enterprise admins can turn Cowork off, set access by role, set spend limits, and stream activity to a SIEM through OpenTelemetry. The page also says Claude can be reached through a Claude account or through Amazon Bedrock, Gemini Enterprise Agent Platform, or Microsoft Foundry.

Plugins, on the same page, bundle skills, connectors, and sub-agents so Claude starts with a role’s tools and instructions. A skill is the file format in the agent skill entry. A connector is one tool.

When does the unified Claude experience turn on?

Anthropic’s Enterprise article, read November 7, 2026, says the unified Claude experience is generally available for Claude Enterprise. Chat and Cowork are one conversation on web, desktop, and mobile. A question stays a chat. A task that needs code or files continues in a cloud environment Anthropic manages, and the person can pick the result up on another device. The article says chats, projects, connectors, skills, and plugins carry over.

The admin article, read the same day, is the one with the date. Chat and Cowork unified is off by default in existing organizations and turns on by default on November 16, 2026. An Owner or Primary Owner who has not changed the setting sees “Will be automatically enabled on November 16, 2026,” and can choose Enable now or Keep disabled. A choice already saved stays as it was. The switch also needs Chat, and cloud code execution and file creation, to be on. If either is off, nothing turns on by itself. After November 16, the new experience turns on once both are on.

The same article excludes organizations that use zero data retention for this setup, a HIPAA-ready configuration, or customer-managed encryption keys. Those organizations keep chat and Cowork as they are. The Enterprise article says the same exclusion. HIPAA support is described as expected after general availability. Customer-managed keys are described as a later release.

For organizations that do turn it on, the admin article describes the cloud session in concrete terms. Each conversation that needs agentic work gets its own virtual machine. After a few idle minutes the machine stops and its disk is saved as an encrypted snapshot. The sandbox expires 30 days after it was created. Network egress is off by default for Enterprise. Local files and a linked computer stay optional, and Remote control has to be on before web, mobile, or a scheduled task can use that computer.

Cost stays usage-based. The Enterprise article says Claude Opus 5.5 costs 40 percent less to run than Claude Opus 5 on typical workloads, and Claude Sonnet 5.5 costs up to 30 percent less than Claude Sonnet 5 for most work. Those are Anthropic’s comparisons, not a public price card. Admins can turn on Efficient mode, which the help article limits to Sonnet and Haiku in the unified experience, with new sessions starting on Claude Sonnet 5.5 at low effort. Model allowlists and spend limits at the organization, group, and member level still apply.

What is ChatGPT Work?

ChatGPT Work, read November 7, 2026, is OpenAI’s agent for a finished piece of work: a presentation, a report, an analysis, a spreadsheet, or a recurring update. The page says it is powered by GPT-6, and that GPT-6 introduces Astra as a tier for work that needs more judgment. It is available on every plan in the Mac and Windows desktop apps, and on Plus, Pro, Business, Enterprise, and Edu on the web and on mobile.

OpenAI’s getting-started guide, read the same day, tells you when to leave Chat. Use Chat for an answer, an explanation, a brainstorm, or a short draft. Use Work when the task has a clear outcome, uses more than one source or tool, and produces something you will review. On the desktop app, Work can also use local files, apps, and the browser when those are turned on. The guide says a person who has been using Codex for non-coding work can stay in Codex or move to Work. The same guide names three models in the selector: GPT-6 Astra for demanding work that needs careful reasoning or a polished file, GPT-6.1 Sol for complex work at a lower cost than Astra, and Luna for simpler tasks. Longer tasks can use more of the plan’s allowance because the product is doing more of the steps.

The product page adds the controls around that work. Plan mode gathers context, asks questions, and writes a step-by-step plan you can change or approve before the work starts. The page says more than 1,400 plugins can pull context from tools the team already uses. Scheduled tasks can run once, on a cadence, or when an event happens. Team tasks can delegate a shared job, such as a weekly report, through approved company tools, and teammates with access can edit the instructions. ChatGPT Space is the shared page for editing with the team. Sites turns a plan or a dataset into an internal web page. The desktop app has a built-in browser for work that has to cross tools in tabs.

The getting-started guide’s own instruction is the oversight rule in one line: say which sources to use, what good looks like, and when to stop for review. Its sample asks for a draft before anything is shared, and it tells you to follow the notice if a task pauses for a safety review. If the work needs the computer in front of you, that is computer use, documented for the desktop app on the computer-use guide.

What is a ChatGPT workspace agent?

A workspace agent is a shared agent a company builds inside ChatGPT, separate from one person’s Work task.

OpenAI’s business page, read November 7, 2026, says workspace agents are in research preview for ChatGPT Business, Enterprise, Edu, and Teachers plans. One person defines the job in plain language, connects approved apps, and shares the agent so other people run the same workflow. The page’s examples are recurring work: reviewing leads, summarizing support requests, generating reports, updating tickets, editing documents, and sending messages. Admins control who can build and run agents, which tools those agents can use, and the logs of what ran.

The help article, read the same day, adds the switches. Workspace agents are off by default at launch for ChatGPT Enterprise workspaces. An admin turns them on. People see and run only the agents their role allows. A run can start in ChatGPT, in Slack, on a schedule, or from an API trigger. Write actions, the ones that send, edit, post, or delete, default to Always ask. A builder can set Never ask, or a custom approval for specific writes. The article says to use those write approvals carefully. Action constraints can narrow an app further, for example limiting an email send to one domain. The constraint limits what the agent may ask the app to do. It does not filter a record the app returns through some other allowed call.

Files attached to an agent are limited, on that page, to 512 MB per file and 10 GB total. The article says very large collections can keep the agent from running, and it tells builders to add only the files the job needs.

Edu appears on both the Work availability line and the workspace-agent plan list. A university that buys ChatGPT Edu is buying a workspace with those controls available, not a promise that every student agent is turned on. The next section is what three universities have actually published.

When must a person approve an agent’s action?

A human in the loop is a person who can approve, edit, or stop a step before it counts. An approval gate is the configured stop for one kind of action.

The products checked on November 7, 2026 publish that stop in different places:

• Claude Cowork asks you to choose the folders and tools. Deleting needs approval. The permissions setting makes Claude show a plan and wait before a significant action.
• In the unified Enterprise experience, admins can turn off “Skip all approvals” and “Automatically approve.” Off means a person cannot hand Claude a standing yes for tools, file edits, and browsing, including scheduled tasks and the built-in browser. “Always allow” for a connector tool can be turned off so Claude asks every time. Remote control is a separate switch: when it is off, web, mobile, and scheduled tasks cannot use a member’s linked computer.
• ChatGPT Work’s Plan mode is a review of the approach before the work begins. The getting-started guide tells you to require a draft before a share, and to stay with a task that pauses for a safety review.
• Workspace agents default write actions to Always ask. The business page names sending a message and updating a record as the examples.
• OpenAI’s business guide says the pre-built agent it describes was trained to seek confirmation before many real-world actions, to ask for oversight before sensitive tasks such as email, and to refuse high-risk actions such as financial transfers. Treat that as the guide’s account of its training rules. Confirm the switch in the product you actually have.

Deloitte’s December 2025 essay uses Mapfre as a workplace example of the same pause. Agents handle routine claims tasks such as damage assessments. Customer communication stays with a person. Mapfre’s group chief data officer, Maribel Solanas Gonzalez, is quoted calling the design hybrid: the software changes what people spend time on, and risky work still goes through a human worker. That is Deloitte’s interview, not a product manual.

A gate you never look at is a rubber stamp. Deloitte’s September 2026 essay says to watch for the moment when approval speed becomes the main measure, because review then multiplies with the volume of agent output and stops being a check. The measurement for a real check is the same idea as a behavioral evaluation: name the action, such as “sent without a person,” and see whether the run still does it.

Desktop agents raise a second question, which is whether security tools can see the work. Palo Alto Networks’ November 2, 2026 post says Claude Cowork does this desktop work inside a small local virtual machine. That boundary also keeps a normal endpoint sensor from watching the agent. Palo Alto says a Cortex XDR integration will put a lightweight sensor inside that machine, and that the integration will be available in private preview in the coming months. The post’s own forward-looking notice says unreleased features may not ship on the stated schedule, or at all. Anthropic’s admin article, separately, says an endpoint sensor can run inside the virtual machine on a user’s computer and is in beta, and that this does not cover the cloud sandbox Anthropic manages. A cloud session and a desktop session are two places to log.

Autonomy is how many of these gates you remove. A suggest-only draft is the low end. A saved Always allow, or a write action set to Never ask, is the high end. Money is the case this site already treats as its own gate. The shopping guide and the computer-use guide both say a person stays on the payment step unless a product has a narrower checkout rule you already accepted.

How are universities using workplace AI?

The education pages checked for this guide are about adults in universities, and about the jobs those adults are preparing for. They are access programs, assessment rules, and staff training. They are not a hiring plan for digital workers, and they are not a school product.

OpenAI’s November 5, 2026 post describes collaborations with UNSW Sydney, the University of Sydney, La Trobe University, and Adelaide University. It also reports a survey Pureprofile ran for OpenAI: 1,018 university students in Australia, aged 18 and over, studying on campus or remotely, from August 3 to August 7, 2026. In that survey, 81 percent said they already use AI several times a week, 33 percent several times a day, and 28 percent said they had already built an AI agent or assistant for study, work, or personal life. Those percentages are the survey’s. They describe students who already have the tools, which is the access problem the universities say they are trying to even out.

The university pages, read November 7, 2026, are more specific than the slogan.

UNSW’s September 9, 2026 announcement says the partnership will bring ChatGPT Edu to more than 80,000 students and staff. Access that week was being extended to higher-degree research candidates and casual academic staff, on top of access that professional and academic staff already had. The university said it was working toward the whole student community ahead of the 2027 academic year. That is a staged rollout, not a statement that every student could sign in on September 9. UNSW’s assessment rule is Can, Can’t, Must: a course can permit AI, prohibit it where the student has to show the skill alone, or require it when the point of the assignment is to use it. The announcement also describes staff training, research uses such as literature review and coding, and intended links to internships and industry projects.

The University of Sydney’s September 14, 2026 announcement says students and staff will get free, supported ChatGPT Edu, introduced in stages, with more features turned on over time. University data will not be used to train models. The university is also putting $1 million into an AI Sustainability Research and Innovation Fund. Its assessment description is a two-lane rule: secure supervised work where students show what they can do themselves, and open work where they show how they use AI. Cogniti, the university’s own teaching tool, stays the preferred classroom system. ChatGPT Edu is the supplement. OpenAI’s November 5 post adds a date the university page does not: access for all students scheduled for Semester 1, 2027, after a trial. Use the university page for the commitment, and OpenAI’s post for that semester date.

La Trobe’s September 28, 2026 note, by Pro Vice-Chancellor Phil Laufenberg, says the OpenAI collaboration commits ChatGPT Edu access for all of about 45,000 students and staff, with the initial rollout already underway. OpenAI’s November 5 post is tighter on the calendar: all 3,500 staff already have ChatGPT Edu, and universal access for those 45,000 students and staff is planned for the first quarter of 2027. The same OpenAI post says La Trobe has made Codex and ChatGPT Work available across the institution, and it points to handbook subjects, including agentic software engineering, from 2027. Laufenberg’s own note emphasizes judgment: a graduate who can decide whether AI helps, check the output, and apply it. That is the human-in-the-loop skill, taught to adults, rather than a claim that the campus has hired agents.

Adelaide University appears in OpenAI’s November 5 post as a newer institution building on existing API use, including OpenAI Daybreak Blue for authorised defensive security work. This guide does not have a separate Adelaide page confirming that sentence, so it stays OpenAI’s account.

None of these pages says a university has replaced a staff job with an AI employee. The pattern they do describe is the workplace pattern in the table above: give people a capable chat and, on some campuses, agentic tools, then write down where a person still has to show the work.

What should you decide before calling software an employee?

1. Name the job in one sentence, and name the file or the action that counts as done. “A Monday deck from these two sources” is a job. “Be my employee” is a slogan.
2. Pick the narrowest system that can finish it. A chat answer, a fixed workflow, and an agent are three different grants. OpenAI’s guide says to use the agent where the path has to adapt.
3. List the tools. Claude Cowork and ChatGPT Work both ask you to connect folders, plugins, or apps. A tool you did not connect is a tool the agent should not have.
4. Write the approval gates. Deletes, sends, record changes, and payments are the stops on the pages above. A draft for review is the default a careful team can live with.
5. Decide who is accountable when the draft is wrong. Deloitte’s September 2026 warning is the useful test: if the notes of a mistake name the agent as the responsible party, the accountability has drifted.
6. If the agent runs on a desktop, decide who can see that session. Palo Alto’s November 2 post is the security version of the same question, and it is still a private preview.
7. If the buyer is a university, read the staging date on that university’s own page. “ChatGPT Edu” on a plan list is not the same sentence as “every student has it today.”

The how agents work topic is the loop in general. Task performance is where this site follows office work, computer use, and other jobs an agent finishes. Education is the shelf for learning and campus tools. New AI tools and SaaS is the shelf for the products. Harnesses is the workshop around the model.

Where can you read more on this site?

• This guide’s terms: digital worker, human in the loop, approval gate
• Nearby terms: agent, harness, guardrail, autonomy, tool use, computer use, agent skill, language model
• Guides: What is an AI agent harness?, How do you evaluate an AI agent harness?, Can an AI agent use your computer?, Can an AI shopping agent check out for you?
• Campus policy that is out of scope here: Google will turn off Gemini and Search for K-12 students unless schools opt in
• Directory: Claude, ChatGPT
• Topics: how agents work, task performance, education, tools and SaaS, harnesses

Frequently asked questions

What is an AI employee?

It is a label for workplace software that is given a job and returns work for a person to review. The glossary entry is digital worker. On November 7, 2026, the two product pages that match the label most closely are Claude Cowork and ChatGPT Work. The software is not a legal employee. Deloitte’s September 10, 2026 essay says to treat a review that names the agent as the responsible party as a sign that accountability has slipped.

Is an AI employee the same thing as a chatbot?

A chatbot answers in the thread. OpenAI’s business guide says that system can call a tool for a current document and still cannot plan a multi-step job unless every step is programmed. Claude’s Cowork page says a conversation starts as chat and picks up file work only when the task needs it. ChatGPT’s Work guide says to stay in Chat for a short answer and to use Work for a finished file. The table above is the comparison.

How is an AI employee different from RPA?

Fixed automation follows rules written in advance. OpenAI’s guide calls that a workflow automation and uses an expense approval and a login lock as the examples. People often call that pattern RPA. An agent starts from a goal and can change its plan when a condition changes. Deloitte’s December 2025 essay says some projects sold as agents are this older automation under a new name. A team can use a script for the stable steps and an agent only where the path is not known.

What is Claude Cowork?

It is Anthropic’s product for knowledge work that is not coding. The product page says you hand Claude a goal, it works in folders and tools you choose, and you review the result. Claude Code is the separate product for shipping software. Computer use on the desktop is still a research preview. Enterprise organizations move to one chat-and-Cowork conversation on the schedule in the next answer.

When does Claude’s unified experience turn on for existing Enterprise organizations?

On November 16, 2026, by default, if an owner has not already chosen. Anthropic’s help article says the setting is off until that date in existing organizations, and that owners can enable it early or keep it off. Chat and cloud code execution have to be on. Zero data retention, HIPAA-ready, and customer-managed encryption-key organizations are outside this switch.

What is ChatGPT Work?

It is OpenAI’s agent for a task with a reviewable result. The product page read on November 7, 2026 says it is powered by GPT-6, on every desktop plan, and on Plus, Pro, Business, Enterprise, and Edu on the web and on mobile. Plan mode asks you to approve an approach first. Workspace agents are the shared, admin-controlled version, in research preview for Business, Enterprise, Edu, and Teachers, and off by default for Enterprise at launch.

Does a workplace agent need a person’s approval?

For a send, a delete, a record change, or a payment, the pages checked here say the person should be asked. That pause is an approval gate. Claude can show a plan and wait. Enterprise admins can forbid modes that skip approvals. ChatGPT workspace write actions default to Always ask. A saved permission is a wider grant. The habit of reading the draft is the human in the loop.

Are university ChatGPT Edu rollouts the same as hiring an AI employee?

They are staged access for adult students and staff, with training and assessment rules. UNSW’s September 9, 2026 page, Sydney’s September 14 page, and La Trobe’s September 28 page each describe ChatGPT Edu for study and university work, not a digital staff roster. OpenAI’s November 5 post is the overview and the source for a few dates the university pages do not print. School-age account policy is the K-12 news post, and it is outside this guide.