Substantially updated and fact-checked 17 July 2026: OpenAI introduced ChatGPT Work as a work-focused experience that can research, use connected context, create documents and spreadsheets, prepare reports, build shareable Sites, and run scheduled tasks. The July desktop update also brings local files, apps, browser work, and Codex closer together. This guide now covers that wider Work-and-Codex system while preserving its original analysis of plugins, Sites, and marketer workflows. For marketers, the opportunity is a governed workbench – not an unsupervised employee.
Short answer: use ChatGPT Work for multi-step knowledge work that crosses research, files, analysis, and deliverables. Keep ordinary ChatGPT for lightweight conversation and one-off creation. Use Codex when the work depends on a durable project workspace, code, terminal operations, or more explicit agent execution. Start Work read-only, grant the minimum apps and folders required, define acceptance criteria, and require human approval before publishing, spending, sending, deleting, or modifying a live system.
What is ChatGPT Work?
ChatGPT Work is OpenAI’s work-oriented product experience for completing multi-step tasks with company and local context. OpenAI says Work can research a topic, work with apps and files, create documents, spreadsheets, presentations and reports, and turn results into shareable Sites. Scheduled tasks let users assign work that runs later or repeats.
The official ChatGPT release notes describe the July rollout and later changes, including unified search across chats, projects, images, and documents. The ChatGPT Work product page is the current product entry point. Availability can differ by plan, workspace settings, region, device, app installation, and administrator policy.
This is a meaningful change for marketing because real work rarely lives in one prompt. A campaign decision may require a research brief, a spreadsheet, previous creative, analytics, stakeholder comments, and a review surface. Work attempts to keep those pieces in one execution context.
ChatGPT vs ChatGPT Work vs Codex
| Surface | Best fit | Typical marketing use | Primary control to add |
|---|---|---|---|
| ChatGPT | Conversation, quick research, ideation, explanation, and bounded creation | Rewrite a paragraph, explore angles, summarise supplied text, draft a small asset | Verify facts and avoid uploading unapproved sensitive data |
| ChatGPT Work | Multi-step knowledge work across files, apps, browser context, deliverables, and schedules | Weekly performance package, market scan, content research, campaign review, launch hub | Least-privilege access plus human approval before consequential actions |
| Codex | Durable project work involving code, terminal operations, repositories, technical artefacts, and agentic execution | Analytics implementation, site changes, data pipelines, SEO automation, reproducible audits | Workspace scope, mutation guards, tests, and version-control review |
The boundaries overlap. Codex can create documents, analysis, and Sites; Work can involve Codex on desktop; ChatGPT can use apps. Choose by the centre of gravity of the job rather than the newest product name. The useful pattern is the one described in DMT’s agentic AI in marketing guide: tools, context, actions, and verification must be connected without erasing human control.
How Codex plugins fit into the Work system
Codex began as a coding agent, but OpenAI’s June 2026 expansion positioned it as a broader project workspace for knowledge work. A Codex plugin is more than a connector. In OpenAI’s description, a role plugin can package four layers:
- apps that connect approved context or actions;
- skills that encode task-specific procedures and quality standards;
- instructions that preserve terminology, constraints, and expected behaviour;
- workflows that turn context into a repeatable deliverable or decision.
That structure matters because most marketing failures are not caused by a lack of fluent text. They come from scattered briefs, conflicting metric definitions, unclear approvals, untraceable files, and no repeatable review standard. Work provides a broader execution surface; Codex plugins can encode the operating procedure inside it.
OpenAI’s initial role-specific plugin announcement described dozens of apps and skills across data analytics, creative production, sales, product design, public-equity investing, and investment banking. A dedicated Marketing Strategy plugin was described as a future direction at that time, not a launched capability. Teams should distinguish a currently installed plugin from a named roadmap item and check the live Plugin Directory before planning around it.
Creative-production plugin pattern
For marketing, the valuable creative pattern is not unlimited generation. It is a controlled variant system:
- retrieve the approved brief, claims, audience, offer, brand rules, and channel constraints;
- generate a small set of deliberately different concepts;
- attach the hypothesis and intended audience to every variant;
- route work through brand, legal, and channel approval;
- export only approved assets;
- connect performance back to the originating hypothesis.
Without that chain, faster production creates information debt: more files, ambiguous approvals, and no learning loop.
Data-analytics plugin pattern
A credible analytics workflow loads governed metric definitions before querying, compares the current value with target and prior periods, separates observed facts from inferred drivers, tests alternative explanations, retains calculation receipts, and states what evidence is missing. An attractive chart or Site is not evidence by itself. Resolve duplicated conversions, inconsistent UTMs, consent gaps, and channel-specific attribution differences before automating the performance narrative.
Annotations and targeted review
OpenAI introduced annotations so a reviewer can point to a specific part of code, a document, spreadsheet, slide, or Site and request a targeted revision. The governance benefit is precision: “change this chart’s comparison period” is safer and more auditable than rerunning an entire deliverable with a broad prompt. Keep the original, annotation, revised artefact, and reviewer decision together.
Sites as a review surface
Sites can turn analysis into an interactive internal decision surface: a campaign launch hub, channel scenario planner, creative review gallery, or SEO opportunity dashboard. Every Site should show its owner, data sources, refresh time, definitions, assumptions, access policy, and unresolved questions. Treat it as a generated work product that requires verification, not as a new source of truth.
What changed in July 2026
A consolidated desktop work experience
OpenAI’s 9 July release notes describe a desktop “superapp” direction where Work can access local files and apps, browse, and collaborate with Codex. That reduces context switching, but it also concentrates permissions. A tool that can see a browser, local documents, and connected business systems should not start with blanket access.
Some early users have criticised the merged desktop workflow and reported disruptive changes. Media coverage captures that reaction, but complaints are qualitative evidence, not proof that every user or plan will behave the same way. Pilot the update with a reversible workflow before moving a critical operating process.
Scheduled tasks
Work can run scheduled tasks, enabling recurring research, monitoring, preparation, and reporting. Scheduling changes the risk profile. A weak one-off prompt wastes a few minutes; a weak daily task creates compounding errors, clutter, stale conclusions, or excessive usage.
A scheduled task needs:
- a fixed scope and source policy;
- a clear schedule and timezone;
- an output destination;
- an idempotency rule so it does not duplicate work;
- a success and failure notification;
- a stop condition;
- an owner and review deadline;
- a monthly decision to keep, revise, or retire it.
Plugin Directory and Sites
OpenAI’s July release notes point users toward a Plugin Directory and describe Sites as a way to publish interactive work products for sharing. For marketers, a Site can turn a static analysis into a campaign review hub, scenario planner, editorial dashboard, or creative gallery. It should still display data provenance, update date, owner, assumptions, and access boundaries.
Unified search and longer instructions
By 14 July, OpenAI said search could span chats, projects, images, and documents. On 15 July it expanded custom instructions to 5,000 characters for paid and workplace users. These changes improve retrieval and persistent context, but longer instructions are not automatically better. Keep the stable operating rules concise; route changing project facts to governed files or connected sources.
Atlas retirement
OpenAI’s release notes say Atlas is scheduled for retirement on 9 August 2026 as desktop capabilities move into the broader ChatGPT experience. Teams with Atlas-dependent workflows should inventory them now, map required files, browser actions, permissions, and outputs, then test an equivalent path before the retirement date.
Seven practical ChatGPT Work workflows for marketers
1. Source-backed market and competitor scan
Input: a market definition, approved source list, date window, customer segments, and decision question.
Work sequence: collect current primary sources; extract claims, dates, pricing, positioning, and evidence; flag contradictions; compare competitors against the same criteria; create a research table and a short decision memo.
Human approval: validate high-consequence facts and decide what the evidence means. Do not ask the system to scrape or reuse material in ways that violate access, privacy, or copyright rules.
Acceptance criteria: every important claim links to a source; facts and inferences are labelled; stale pages are identified; missing evidence remains visible.
2. Weekly marketing performance package
Input: governed metric definitions, analytics exports or approved read-only connections, targets, comparison periods, and known tracking issues.
Work sequence: validate completeness; calculate the agreed metrics; surface material changes; test alternative explanations; build a spreadsheet, narrative, and review Site; route it to the owner.
Human approval: the channel owner accepts or corrects the explanations and owns budget decisions.
Acceptance criteria: totals reconcile to source systems; observed facts are separate from inferred drivers; each recommendation includes an owner, action, and validation.
An AI-written performance story is only as reliable as the measurement layer beneath it. Resolve conversion definitions and attribution caveats before automating the narrative.
3. SEO refresh and information-debt review
Input: the existing article, Search Console performance, competitor pages, current primary sources, internal-link inventory, and the target intent.
Work sequence: identify stale facts, missing decisions, thin examples, unsupported claims, and new sub-intents; propose update/merge/redirect/leave-alone; draft a change plan; prepare revisions and a fact-check ledger.
Human approval: an editor approves the URL decision and all factual or positioning changes.
Acceptance criteria: the update preserves useful existing value, closes recency and information debt, avoids cannibalisation, and adds contextual internal links. DMT’s SEO content strategy guide provides the wider planning framework.
4. Campaign launch review
Input: approved brief, audience, offer, claims, landing pages, creative, tracking plan, budgets, channel policy, and owners.
Work sequence: compare every asset with the same launch checklist; identify mismatched messages, broken URLs, missing UTMs, policy risks, and absent approvals; produce a blocker list and owner view.
Human approval: channel, brand, legal, analytics, and business owners close their blockers. Work prepares the review; it does not launch media autonomously.
Acceptance criteria: every in-scope asset is accounted for; no unresolved critical blocker; landing-page and tracking checks have evidence.
5. Controlled content production
Input: approved research, keyword/intent brief, brand rules, source policy, internal-link candidates, and editorial acceptance standard.
Work sequence: produce a brief; map claims to evidence; draft; run a factual and information-debt review; create metadata; prepare an approval package; publish only after explicit approval.
Human approval: editor and subject owner verify the article and publication decision.
Acceptance criteria: direct intent satisfaction, original examples or analysis, primary-source support, clear uncertainty, accurate metadata, no duplicate URL, and a useful reader next step. DMT’s AI and SEO guide explains why scaled generic output is a weak search strategy.
6. Creative variant and learning loop
Input: approved claims, brand system, audience insight, channel specifications, and a testable hypothesis.
Work sequence: generate a small set of meaningfully different concepts; attach the hypothesis and intended audience to each; route for approval; prepare exports; connect results back to the original hypothesis.
Human approval: brand and channel owners approve production and launch.
Acceptance criteria: variants differ strategically rather than cosmetically; no invented claims; assets are traceable to approvals and performance.
This is more useful than “make 50 ads.” DMT’s AI marketing automation guide shows how to connect production with measurement and human control.
7. Scheduled change monitor
Input: a small list of official product, policy, pricing, regulatory, or competitor pages; the facts to monitor; and material-change rules.
Work sequence: check on a fixed schedule; compare the current page with the last verified state; ignore cosmetic changes; prepare a concise change receipt; alert the owner only when a defined threshold is met.
Human approval: the owner decides whether a content, product, or campaign update is required.
Acceptance criteria: no duplicate alerts; every alert includes old state, new state, source, date, likely impact, and recommended verification.
The permission ladder: start narrow
- Public web and supplied files: research and draft without business-system access.
- Read-only approved folders: retrieve context from a defined project directory.
- Read-only business apps: analyse selected data with governed definitions.
- Write to a review workspace: create drafts, reports, or Sites that cannot affect customers or production.
- Prepare actions: create a change set, message, campaign edit, or publish package for approval.
- Limited execution: perform an approved, reversible action within exact bounds and retain a receipt.
Most marketing workflows can create substantial value at levels two through four. Do not grant send, publish, delete, budget, billing, CRM-write, or production-site permissions merely because the connector exists.
A task specification that survives automation
Use this structure for one-off and scheduled work:
- Objective: the concrete outcome and decision it supports.
- Scope: included brands, markets, pages, campaigns, dates, and files.
- Sources: authoritative sources, permitted apps, and prohibited data.
- Definitions: metric, audience, funnel, and status definitions.
- Method: required steps and checks.
- Deliverable: format, destination, detail, and naming.
- Acceptance: tests that must pass.
- Permissions: read and write boundaries.
- Approval: who must accept the result before the next action.
- Failure: when to stop, alert, retry, or roll back.
- Schedule: time, timezone, frequency, and retirement date.
This is the operational version of good prompting. DMT’s GPT-5.6 prompting guide covers model and prompt design in depth; Work adds persistent tools, schedules, and permissions, so the execution contract matters even more.
Security, privacy, and reliability risks
Prompt injection from files and the web
A webpage, document, comment, or hidden instruction can attempt to redirect an agent. Treat retrieved content as data, not authority. Persistent rules should say that only the user-approved operating contract can change permissions or objectives.
Excessive access
Do not connect an entire drive when one project folder is enough. Do not connect a production ad account when a scheduled export can answer the question. Review active apps and remove access when the pilot ends.
Cross-client or cross-project leakage
Separate workspaces, folders, instructions, and output destinations. Add the client or brand identifier to every task and receipt. Never rely on the model to infer which similar file is safe.
Silent model, product, or limit changes
Record the product surface, model where visible, plan, execution date, and output version. Some users have reported that Work activity can consume plan usage faster than expected, but those reports are anecdotal and may be plan- or workflow-specific. Measure your own tasks, set alerts, and do not present a community report as official quota policy.
Confident but unsupported work
Require links, quoted evidence within copyright limits, calculation receipts, and visible uncertainty. For consequential facts, the reviewer should open the source. A polished Site is a delivery format, not proof.
Unbounded scheduled output
Every recurring task needs a maximum runtime or work unit, a deduplication rule, a destination quota, and a retirement date. Alert on failures; do not let retries multiply silently.
How to measure whether ChatGPT Work is actually useful
| Metric | Definition | Failure signal |
|---|---|---|
| Accepted-task rate | Share of outputs accepted with no or minor correction | High output volume with frequent rejection |
| Time to accepted result | Elapsed time including review and correction | Automation moves work but does not shorten it |
| Reviewer minutes | Human effort per accepted deliverable | Review cost exceeds saved production time |
| Critical error rate | Unsupported claims, wrong calculations, privacy breaches, or harmful actions | Any uncontrolled high-consequence error |
| Intervention rate | Tasks requiring a person to recover, redirect, or repair | Recurring failures are hidden as “human in the loop” |
| Reproducibility | Another owner can trace inputs, method, and result | The result cannot be audited |
| Total cost | Plan/API usage, tools, labour, and failure cost | Seat adoption rises while accepted work does not |
Avoid measuring messages sent, words generated, or number of scheduled tasks as primary success metrics. They measure activity, not value.
A 30-day rollout plan
Week 1: inventory and boundaries
- List current ChatGPT, Atlas, desktop, Codex, and manual workflows.
- Classify data and identify forbidden sources or actions.
- Choose one frequent, reversible workflow with a clear quality baseline.
- Define the owner, sources, acceptance rubric, and rollback.
Week 2: read-only pilot
- Run the workflow on supplied exports or a narrow read-only folder.
- Capture inputs, outputs, model/product state, errors, and reviewer time.
- Test prompt injection, missing data, conflicting data, and an unavailable source.
- Keep the existing process as the control.
Week 3: review workspace and schedule
- Allow output only to a draft document, spreadsheet, or internal Site.
- Add a schedule if the one-off workflow meets its acceptance threshold.
- Set deduplication, failure alert, owner deadline, and retirement date.
- Compare total cost and accepted-task time with the control.
Week 4: decide, document, and migrate
- Keep, revise, or stop the workflow based on evidence.
- Document permissions, sources, prompt, checks, owner, and recovery path.
- For Atlas-dependent work, test the replacement before 9 August 2026.
- Expand to a second workflow only if the first is stable and auditable.
This deliberately slow permission expansion is compatible with a broader AI in digital marketing operating model: prove decision quality first, then automate the repeatable parts.
When not to use ChatGPT Work
- the task is a deterministic calculation or data transformation better handled by code or a formula;
- the required data is not approved for the workspace;
- there is no owner who can validate the output;
- the task can directly spend money, publish, message customers, or alter production without an approval boundary;
- the process is so unstable that nobody can define an acceptable result;
- a connector would expose far more data than the task needs;
- failure cannot be detected or reversed.
In those cases, simplify the workflow, use a narrow tool, or keep the work human-led until the evidence and controls exist.
Frequently asked questions
Is ChatGPT Work a separate app?
OpenAI presents Work as a work-focused ChatGPT experience available across supported surfaces, with expanded desktop integration. Exact access depends on the current app, plan, region, and workspace configuration; check the live product and release notes.
Can ChatGPT Work access local files?
OpenAI’s July 2026 release notes describe desktop access to local files and apps. Grant access only to the smallest approved folder or file set, and test with non-sensitive content before relying on it.
Can ChatGPT Work run tasks on a schedule?
Yes, scheduled tasks are a core use case. Give every recurring task a scope, source policy, timezone, output destination, deduplication rule, owner, failure alert, and retirement date.
What is the difference between ChatGPT Work and Codex?
Work centres multi-step knowledge work across files, apps, research, schedules, and business deliverables. Codex centres durable agent execution in project workspaces, including code, terminals, repositories, and technical artefacts. They overlap and can work together; choose based on the job and required controls.
Does ChatGPT Work use the same usage limits as Codex?
Do not assume a universal answer. Plan behaviour and product policies can change, and community reports are not authoritative quota documentation. Check the current plan information and measure representative tasks in your own account before moving a critical workflow.
Is the Atlas browser being discontinued?
OpenAI’s July release notes say Atlas is scheduled to retire on 9 August 2026. Inventory Atlas-dependent workflows, export or document required context, and test the replacement desktop path before that date.
Can ChatGPT Work publish content or change campaigns automatically?
Connected tools may make actions technically possible, but autonomous publishing or campaign mutation increases risk. Start read-only, prepare an approval package, require an authorised person to approve the exact change, and retain a receipt for any limited execution.
Editorial note: ChatGPT Work, desktop behaviour, apps, models, plan limits, and availability are evolving. Official OpenAI sources linked above were checked on 17 July 2026. Secondary reports and user reactions are qualitative context, not universal product facts.
