ChatGPT Work Features Go Beyond Chat Into Sites, Email, and Documents

OpenAI launched ChatGPT Work on July 9, turning ChatGPT from a conversational assistant into an agent expected to produce finished work. The first ChatGPT Work features cover websites, email workflows, document analysis, spreadsheets, presentations, reports, and scheduled tasks.

The change is larger than another model upgrade. OpenAI is separating quick conversations from longer jobs that involve multiple sources, applications, decisions, and deliverables. Work can continue for hours, divide a project into steps, request clarification, and let users redirect it while it runs.

That puts ChatGPT into closer competition with the software people already use to complete knowledge work. Google Workspace, Microsoft 365, website builders, automation platforms, and specialized AI assistants now face a broader interface competing for the same tasks.

The promise is straightforward: describe an outcome, provide the relevant context, and receive something ready to review. The harder question is whether one agent can reliably move across private data, editable files, external websites, and consequential actions without creating new risks.

ChatGPT Work Features Turn Prompts Into Deliverables

ChatGPT Work is designed around completed projects, not longer conversational answers.

OpenAI describes Work as an agent for research, analysis, and creating documents, spreadsheets, presentations, reports, and Sites. Codex remains the separate experience for software development, repositories, tests, and technical workflows.

That division gives ChatGPT three distinct working modes. Chat handles questions and brainstorming. Work handles longer projects and finished materials. Codex remains focused on software engineering.

The distinction matters because producing a deliverable requires more than generating text. An agent must locate source material, resolve conflicting instructions, preserve formats, perform calculations, and revise the result after feedback.

OpenAI says Work can break complex projects into smaller steps and complete those steps independently. Users can follow its progress, answer questions, change direction, and approve important actions during a run.

The company’s Work announcement says the agent can remain on a project for hours when needed. It can also carry context across a chain of related outputs.

A marketing task illustrates that difference. A normal chatbot might draft a campaign brief from a pasted summary. Work is intended to gather customer research, create the brief, produce campaign materials, and adapt them for separate markets.

This is a shift from response generation toward workflow execution. The agent is judged by whether the files, analysis, and actions form a coherent result.

OpenAI built Work with technology developed for Codex. That connection explains why the product emphasizes planning, tool use, file operations, and visible progress rather than a single polished answer.

The company says more than five million people use Codex weekly. It also says more than one million use it for work outside software development. Those company-reported figures help explain why OpenAI expanded the agent pattern beyond coding.

Work operates across ChatGPT web, mobile, and the desktop application. Cloud Work conversations can sync across those surfaces, allowing someone to start a task on a phone and continue reviewing it elsewhere.

The desktop application adds an important difference. With permission, Work can use local files and desktop applications that cloud sessions cannot directly access.

That local access creates broader possibilities for document-heavy jobs. It also requires careful scoping because a desktop agent can encounter private material beyond the immediate task.

OpenAI advises users to grant access only to the files a task needs. Local files and outputs remain on the computer unless the user explicitly moves or shares them.

Work can also operate inside a ChatGPT Project. Projects keep related conversations, files, and instructions together, giving longer assignments persistent context without requiring users to attach everything again.

The product therefore combines several existing ChatGPT ideas into one task-oriented surface. Projects supply context, plugins connect external systems, Scheduled Tasks provide recurrence, and the agent coordinates the workflow.

That combination is the central change behind ChatGPT Work features. OpenAI is no longer asking users to assemble separate research, writing, file, and automation sessions manually.

The agent now attempts to own the path from an initial request to a reviewable output. Whether that path remains accurate and controllable will determine how much work users actually delegate.

ChatGPT Sites Adds Building and Hosting

ChatGPT Sites moves Work into territory previously occupied by website builders, dashboard tools, and lightweight application platforms.

Sites lets users create interactive websites and small web applications from a description. A user can ask Work to build a website or invoke Sites directly within a prompt.

OpenAI lists dashboards, project trackers, launch calendars, prototypes, internal portals, and interactive reports as intended use cases. These examples emphasize functional work surfaces rather than conventional company homepages.

A user can add files, links, data, content, and design constraints to the request. ChatGPT generates a private preview, accepts revisions, and provides sharing controls when the result is ready.

The Sites documentation confirms that deployment produces a live Site URL. Users can keep a Site private, share it with selected people, expose it within a workspace, or publish it publicly when permitted.

This makes the claim that Work can build and host websites substantially accurate. However, Sites is a managed runtime with defined constraints, not a replacement for every hosting platform.

OpenAI says some frameworks, private networks, databases, background services, and hosting patterns may not work. Supported capabilities depend on the runtime and the features enabled for the account.

That qualification is significant. A launch dashboard or interactive report fits the product’s current scope more naturally than a complex commerce system or regulated customer portal.

Sites does not support financial transactions at launch. It also cannot process payment-card data or protected health information, according to OpenAI’s published restrictions.

Developers also cannot assume that every generated application receives conventional infrastructure controls. The code, storage, logs, deployment environment, and permitted integrations remain governed by the Sites platform.

Custom domains are available in some environments, but Sites does not register a domain. The user must already own it and must update its DNS records through the domain provider.

Enterprise administrators receive additional control over creation and publishing. Public publishing is disabled by default in Enterprise workspaces, and an administrator must grant access before members can publish externally.

That governance layer reduces accidental exposure, but it does not remove the need for review. A generated Site can still include source text, files, forms, links, or data that should not be public.

OpenAI explicitly tells users to inspect content, access settings, forms, authentication behavior, uploaded files, and interactive features before publishing. Generated output should not be treated as automatically safe.

This review requirement exposes the tradeoff at the center of Sites. The same system that lowers the effort needed to publish also lowers the effort needed to publish a mistake.

A project manager could turn a spreadsheet and meeting notes into a live status portal. That saves assembly time, but the manager still owns the accuracy of dates, responsibilities, and shared information.

A researcher could convert findings into an interactive report. The result might communicate evidence better than a static document, yet every citation, visualization, and generated interpretation still needs verification.

Sites can also update as underlying information changes. Combined with Scheduled Tasks, that creates a path from a static artifact toward a maintained operational surface.

For example, a team could monitor account activity and refresh a sales command center every morning. Another team could update a launch calendar when new tasks or schedule changes appear.

Those workflows pressure standalone dashboard builders because ChatGPT can create the interface using context already available through connected tools. The user does not need to manually transfer every data point.

However, specialized platforms retain advantages in permissions, testing, integrations, analytics, reliability guarantees, and complex data modeling. Sites currently competes most directly at the lightweight end of the market.

Its strongest early use may be replacing internal artifacts that never justified a full development project. Many teams still coordinate through spreadsheets, slides, and manually updated pages.

ChatGPT Sites gives those teams a quicker path to an interactive version. The crucial test is whether generated Sites remain dependable after their first impressive demonstration.

Email Management Depends on Connected Apps and Permissions

ChatGPT Work can manage parts of an email workflow, but users should not interpret that as unrestricted control over an inbox.

Work relies on plugins and their underlying apps to reach services such as Gmail and Outlook. These connections can provide messages, attachments, contacts, calendar information, and supported actions.

Once connected, Work can gather email context while completing a larger assignment. It might summarize a thread, identify unanswered questions, extract commitments, or incorporate attachments into a report.

The agent can also connect email with other sources. A sales preparation task could combine recent messages, calendar events, CRM records, presentation materials, and public company information.

That cross-source coordination is more valuable than a standalone email summary. The agent can use messages as evidence inside a broader deliverable instead of merely compressing them.

Scheduled Tasks extend the model further. OpenAI says Work can monitor new messages, update documents or slides, and share important changes with a team.

One example involves refreshing a presentation when new feedback arrives by email. Another involves converting new Slack activity into an updated meeting agenda on a recurring schedule.

These workflows turn the inbox into an event source. Email becomes one input within an ongoing process rather than the final place where work accumulates.

However, available actions vary by plugin, account, workspace policy, and region. Connecting an email service does not guarantee that every reading, drafting, sending, editing, or filing action is supported.

The plugin permissions separate access from approval behavior. A connection determines what information and actions are technically available. Permission settings determine when ChatGPT must ask before using them.

OpenAI’s default approach allows many reading actions while requesting approval for important changes. Sending or editing an email is listed as an action that can require confirmation.

Administrators can impose stricter controls in managed workspaces. They can disable apps, restrict actions, configure role access, and decide when members must approve changes.

Individual users may also choose more restrictive settings. Requiring approval before any change offers more control than allowing lower-risk modifications to proceed automatically.

The safer prompt is not “handle my email.” OpenAI advises against vague, open-ended instructions because they leave too much room for interpretation.

A better task names the account, time range, senders, desired classification, output format, and actions that require review. It should also state what the agent must never send, delete, or share.

For example, a user might ask Work to review customer messages from the previous week and draft replies without sending them. The task could require a separate list of ambiguous or sensitive cases.

That structure keeps judgment visible. It also limits the damage caused by a misunderstood instruction, missing context, or incorrectly classified message.

Email introduces another major risk: prompt injection. This occurs when malicious text inside external content tries to redirect an agent or make it reveal protected information.

A hostile message might instruct the agent to ignore the user’s request and forward private content elsewhere. Because Work reads messages as task inputs, it must distinguish data from commands.

OpenAI says its agents use confirmations, monitoring, refusal rules, and automated review to reduce this risk. The company also states that those safeguards do not eliminate it.

Users should therefore connect only the services required for the current workflow. They should inspect recipients, attachments, quoted text, and proposed actions before approval.

Email management also exposes the difference between drafting and operational authority. Producing a good reply is a language task. Choosing whether to send it can involve legal, financial, reputational, or interpersonal consequences.

Work can accelerate the first part and assist with the second. It does not transfer accountability away from the person or organization operating the account.

The practical near-term value lies in bounded workflows. Summarizing a defined mailbox, preparing drafts, extracting tasks, and updating a controlled document are easier to review than autonomous inbox management.

If OpenAI can make those bounded workflows dependable, email becomes a strong entry point for broader adoption. If users encounter misplaced messages or unsafe actions, trust will erode quickly.

Document Processing Becomes an End-to-End Workflow

The document story is not just summarization because Work can turn many sources into editable files, spreadsheets, slides, and reports.

ChatGPT already allowed users to upload files and ask questions. Work expands that behavior into a multi-stage production process that can preserve templates, revise outputs, and coordinate multiple formats.

Users can start from instructions, source materials, or an existing template. They can specify what must remain unchanged, including formulas, layout, branding, slide order, and table structure.

The file creation guide says Work supports editable documents, spreadsheets, presentations, reports, and analyses. Availability still depends on the file type, application, plan, and workspace configuration.

Native Google Docs, Sheets, and Slides workflows require the relevant Google Workspace app. Desktop integrations differ, and not every format receives identical support on every surface.

This means “summarize massive document collections” should be understood as a workflow goal, not an unlimited technical guarantee. Collection size, file compatibility, source access, and context quality still affect results.

The useful distinction is between retrieval and synthesis. Retrieval locates passages or facts. Synthesis combines them into a new structure, identifies relationships, and explains what deserves attention.

A strong Work assignment should define both stages. It should say which sources are authoritative, how conflicts should be handled, and what evidence must accompany each conclusion.

Consider a product team reviewing hundreds of interview notes, support tickets, and planning documents. A generic summary might produce broad themes while hiding differences between customer groups.

A better workflow asks Work to preserve source references, separate repeated complaints from isolated requests, and identify claims that lack enough evidence. The output can then become a prioritized report or presentation.

Teams already using a searchable knowledge base may benefit from keeping source discovery separate from final artifact generation. That separation makes it easier to inspect what entered the analysis.

Spreadsheets introduce additional requirements. A correct-looking workbook can still contain broken formulas, mismatched ranges, hidden assumptions, or charts built from incomplete data.

Users should specify required worksheets, formulas, columns, charts, and validation checks. They should also compare important outputs against the original data before sharing decisions.

OpenAI says its finance teams use Work to find source data, move it into Excel or Sheets, reconcile it, build slides, and verify results. According to the company, that process reduced some month-end work from days to hours.

This is an internal OpenAI example, not an independent benchmark. It shows the intended workflow, but it does not establish the same result for every organization.

The presentation workflow follows a similar pattern. Work can combine source files into a structured deck, use an existing master presentation, and revise the narrative after feedback.

The difficult part is rarely generating individual slides. It is deciding what belongs in the story, preserving evidence, maintaining visual consistency, and avoiding unsupported conclusions.

Work attempts to coordinate those decisions across the entire deliverable. Users can request revisions without manually moving text between a chatbot, spreadsheet, document editor, and presentation tool.

That coordination pressures specialized writing and presentation assistants. Their individual features matter less when a general agent can move context across several output formats.

Specialized tools still have room to compete. They can offer deeper controls, stronger template fidelity, domain-specific review, more predictable formatting, and clearer audit trails.

ChatGPT Work instead competes through breadth. Its advantage is the ability to start with mixed sources and produce several connected outputs within the same project.

Breadth also increases the verification burden. An error can move from a source summary into a spreadsheet, then into a slide deck, Site, or outgoing email.

The final artifact may appear polished even when an earlier interpretation was wrong. Visual quality can make weak analysis more persuasive, not more accurate.

Users should build review checkpoints into the workflow. Source extraction, analysis, calculations, narrative, and publication should each receive separate approval when the stakes justify it.

That review model resembles good human project management. Delegation works best when the expected result, constraints, evidence, and decision rights are clear.

For knowledge workers, the immediate benefit is reduced assembly work. Work can gather information and produce a first complete version while the user focuses on judgment.

The long-term value depends on whether OpenAI can preserve provenance across the entire chain. Readers need to know which sources support a claim and which passages the agent inferred.

Without that visibility, document automation risks creating polished uncertainty. With it, Work becomes a practical bridge between scattered information and reviewable action.

The Main Competition Is the Existing Work Stack

ChatGPT Work pressures suites and point solutions by competing for the coordinating role between them.

OpenAI is not merely adding another document editor or website builder. It wants ChatGPT to become the place where users describe outcomes and coordinate the tools needed to reach them.

That places Work against two established approaches. The first centers work inside productivity suites. The second combines specialized applications through automation platforms and manual handoffs.

Microsoft and Google already own many underlying documents, calendars, email accounts, meetings, and permissions. Their advantage comes from native access, familiar interfaces, and existing administrative controls.

ChatGPT approaches those systems as a coordination layer. Plugins let it retrieve context and perform supported actions without requiring users to abandon the applications holding their data.

Website builders face a similar challenge from Sites. OpenAI does not need to match every advanced feature if it can satisfy lightweight internal projects from a single prompt.

Automation platforms also face pressure because Scheduled Tasks can monitor changes and run workflows. The appeal comes from describing the goal conversationally instead of manually configuring every step.

However, conversational setup can hide complexity that visual automation tools expose. Triggers, failure handling, retries, permissions, and data transformations still exist even when users cannot see them.

Reliability therefore becomes the primary opponent to OpenAI’s broad promise. A flexible agent wins demonstrations easily, but operational work demands predictable behavior across repeated runs.

Traditional software encodes rules in fields, formulas, access controls, and explicit workflows. Agents interpret natural-language goals, which makes them adaptable but introduces ambiguity.

The central tradeoff is flexibility versus control. Work can respond to changing context without a new configuration screen, yet that freedom makes outcomes harder to reproduce.

Organizations will likely start with reversible tasks. Research, drafting, internal reporting, and private prototypes allow users to inspect results before they create external consequences.

Adoption will slow around financial commitments, regulated data, public communications, and destructive file operations. Those areas demand stronger validation and narrower permissions.

OpenAI’s own documentation repeatedly emphasizes review and approval. Users can monitor progress, redirect a task, and confirm actions that affect external systems.

That human control is not a temporary inconvenience. It is part of the product’s operating model, especially while the agent spans tools with different risk levels.

The best comparison is therefore not human worker versus autonomous machine. It is coordinated agent workflow versus fragmented manual workflow.

Work succeeds if it reduces searching, copying, formatting, and repetitive updates while keeping important decisions visible. It fails if users spend the saved time auditing hidden errors.

The product’s broad scope also changes how teams should organize context. Throwing every available file into one workspace can increase noise and expose unnecessary data.

A structured AI knowledge base can help separate trusted sources, working drafts, and restricted material. Clear boundaries make agent output easier to evaluate.

Enterprises have additional reasons to care about those boundaries. Administrators need to know which applications an agent can access, what actions it can take, and where outputs are stored.

OpenAI provides workspace controls for plugins, browser access, network access, roles, and important actions. It also offers compliance logging for Work conversations and actions.

Those features matter because one agent can cross boundaries that separate applications normally enforce. A harmless reading action in one system can supply sensitive context to a publishing action elsewhere.

The competitive advantage may therefore come from governance rather than raw model quality. Enterprises will favor systems they can constrain, inspect, and integrate with existing policies.

OpenAI has a head start through ChatGPT’s large user base and Codex-derived agent technology. Microsoft and Google possess deeper control over many workplace systems.

Specialized vendors retain domain expertise and more focused review processes. Automation platforms provide explicit logic and mature operational tooling.

ChatGPT Work tries to sit above all three groups. It offers one interface for goals while relying on their systems for data, actions, and final formats.

That position is valuable but politically difficult. Application providers can restrict integrations, strengthen their own agents, or reserve the best actions for native experiences.

The next phase will not be decided by which product can create the most impressive sample presentation. It will be decided by repeat usage inside real workflows.

What to Watch After the ChatGPT Work Rollout

Three signals will show whether ChatGPT Work features become daily infrastructure or remain an occasional production tool.

The first signal is completion quality across repeated tasks. One successful report matters less than whether a scheduled version remains correct after sources, dates, and formats change.

Users should watch for failures that compound silently. These include missing emails, stale files, broken formulas, incorrect permissions, and unsupported Site behavior.

Reliable systems also need useful recovery behavior. Work should identify blocked steps, explain missing access, preserve completed work, and avoid restarting an entire project unnecessarily.

The second signal is how OpenAI develops approvals and provenance. Users need clear evidence showing where claims originated, which files changed, and what external actions occurred.

Approval requests must arrive at meaningful decision points. Too many confirmations make automation tedious, while too few allow mistakes to escape review.

The cloud browser limits show how carefully OpenAI is staging web actions. At launch, cloud browsing works on supported public pages but cannot sign in or complete payments.

That limitation weakens some automation claims, yet it also narrows exposure while the system develops. Connected apps remain the preferred route when they can perform a task directly.

The third signal is competitive response from Microsoft, Google, automation vendors, and website platforms. Native providers can match agent behavior while offering deeper access to their own applications.

Watch whether those companies improve cross-application planning, editable deliverables, background execution, and permission controls. Their responses will reveal which parts of Work threaten existing usage most directly.

OpenAI’s rollout also needs close attention. ChatGPT Work and Sites are arriving gradually, and access depends on region, account, workspace settings, and product surface.

Sites is in public beta and has additional geographic restrictions at launch. Some organizations will not see the same combination of features described in demonstrations.

Users should verify which capabilities appear in their account before designing a workflow around them. The presence of Work does not guarantee every plugin, file format, Site feature, or action is available.

The right first project is one the user already understands. A familiar task makes it easier to judge missing sources, weak assumptions, incorrect formatting, and unnecessary steps.

Start with a bounded outcome and explicit review criteria. Name the required sources, final format, forbidden actions, and points where the agent must ask for approval.

A monthly report, meeting preparation package, research synthesis, or private project tracker can provide a realistic test. Each has a clear result without requiring immediate public or financial action.

Then compare the complete workflow, not only generation speed. Measure preparation time, correction time, source coverage, formatting quality, and the effort required to supervise the run.

ChatGPT Work features represent OpenAI’s clearest attempt to move from answering questions into coordinating knowledge work. Websites, email, documents, spreadsheets, and presentations now sit inside one agent-led workflow.

The launch does not make every workflow autonomous. It creates a new delegation layer whose value depends on access, accuracy, permissions, and review.

The most important question is practical: which recurring project consumes hours of searching, copying, reconciling, and formatting without requiring irreversible decisions?

That is the best place to test ChatGPT Work. Give it a familiar outcome, limit its authority, inspect every source, and judge whether the finished deliverable truly reduces work.