Read this first

Most people use ChatGPT as a smarter search box. That is roughly two percent of what is sitting in the account they already pay for. The gap is not intelligence — it is plumbing, permissions and habit.

This guide is organised the same way every section is: what the thing is, why it matters, and how you turn it on. Privacy comes first, because if you run a real business you should decide where your data goes before you decide what to automate.

The sixty-second version

ChatGPT is now three products in one — Chat, Work and Codex. Most people have only ever used Chat.
You can turn off model training and use Temporary Chat. Business and Enterprise workspaces are not trained on by default.
You can give it permanent knowledge — Projects, files, custom instructions, memory, and company knowledge across your connected tools.
You can connect it to the systems you already run — Gmail, Drive, Slack, SharePoint, GitHub, HubSpot, Asana and more.
You can connect it to your own internal systems using MCP, including servers that never touch the public internet.
You can let it act — browse signed-in websites, work on local files and folders, produce finished documents and spreadsheets.
You can put it on a schedule, or on a trigger. New email arrives, it runs. New Slack message, it runs.
You can package a workflow as a Skill or a Plugin so it runs the same way every time, for everyone on your team.

Three honest warnings before you start

One. Availability differs by plan, by workspace setting, and by surface. Something documented as available may not be live in your account yet — OpenAI rolls features out gradually. Two. Connecting tools expands your attack surface. Independent researchers have documented data-exfiltration risks in connector flows. Treat any third-party connector like a production dependency, not a browser extension. Three. The model can be confidently wrong. Nothing in here removes the need for a human to check anything that touches money, contracts, insurance or safety.

The map — Chat, Work, Codex

Before anything else: know which product you are in. OpenAI now splits ChatGPT into distinct experiences, and the one most people live in is the least capable one.

	WHAT IT IS	WHERE IT RUNS
Chat	Fast, conversational assistance. Questions, search, brainstorming. This is the one everyone knows.	Web, mobile, desktop
Work	An agent built for longer, multi-step work and finished deliverables — research, analysis, documents, spreadsheets, presentations, reports, Sites.	Web and mobile on eligible paid plans; desktop where included
Codex	Dedicated to software development and technical work — code, tests, terminal commands, repositories.	Desktop app only. Not selectable on web or mobile
Voice	A live voice interface into Work or Codex. Speak, interrupt, coordinate tasks by talking.	Desktop app on macOS and Windows

On desktop, you switch with the top-left menu (ChatGPT or Codex), then the Chat / Work toggle at the top of the page. On mobile it is the dropdown at the top of the screen. Cloud Work chats sync across web, mobile and desktop. Codex history stays separate — but you can reach supported desktop Codex sessions from the Remote tab in the ChatGPT mobile app.

The single most useful sentence in this document

In the desktop app, when your plan and workspace allow it, Work can use local files and desktop apps with your permission. That is the line between an assistant that talks about your business and one that operates on it. Work on web and mobile cannot reach files on your computer.

Privacy — decide this before you automate anything

Turn off training

WHAT: A single account-level toggle called Improve the model for everyone, in Settings → Data Controls.
WHY: On personal plans — Free, Plus, Pro — it is on by default. If it is on, your conversations can be used to train OpenAI's models. On Business, Enterprise and the API, OpenAI does not train on your inputs or outputs by default. That is the real reason to buy a workspace seat, not the feature list.
HOW: Settings → Data Controls → toggle Improve the model for everyone off. It applies to every device signed into that account. You can also opt out through OpenAI's privacy portal — "do not train on my content".

Temporary Chat

WHAT: A chat that leaves no trace.
WHY: It does not appear in your history, it does not use or create memories, and it is not used to train the models. OpenAI deletes temporary chats from its systems after 30 days.
HOW: Start it from the chat screen each time — unlike the training toggle, it is per-conversation, not a one-time setting. Use it for anything you would not want persisted: a client dispute, a payroll question, a legal draft.

The exceptions people miss

Feedback overrides opt-out. If you hit thumbs up or thumbs down, the entire conversation attached to that feedback may be used for training — even if you opted out.
Codex is governed separately. Codex has its own controls for training on full environments, managed in Codex Settings. Changing your ChatGPT setting or the privacy portal does not change those.
Shared Projects are all-or-nothing. On personal plans, shared project data is only used for training if every contributor and owner has the toggle on.
Connected apps are treated differently. Data accessed directly from connected Google apps is not used to train the generalized models, and if your training toggle is off, connected-app data is not used to improve models even when it shows up in your conversation.

The practical rule for a real company

Customer records, pricing, contracts, employee data, insurance claims — put them behind a Business or Enterprise workspace, where the no-training default is contractual rather than a switch an employee might flip back on. Standard Business seats are $25 per user per month, or $20 billed annually, with a two-seat minimum. That is the cheapest governance you will ever buy.

Knowledge — stop re-explaining your business

Every time you paste your pricing in again, you are paying for the same context twice. There are four layers of permanent knowledge and most people use none of them.

Projects

WHAT: A container that keeps related chats, files and instructions together.
WHY: Everything inside inherits the same files and the same rules. One project per function — estimating, hiring, collections, supplements — means you stop briefing it from zero every morning.
HOW: Create a project, upload the source files, write the project instructions once. From a project you can start either a Chat or a Work session using that project's context. Access to projects is included with every qualifying plan — no extra charge.

Company knowledge

WHAT: A mode that searches across your connected business tools and answers with citations back to the source.
WHY: This is the one that turns it from a writer into an analyst. It works across supported apps like Slack, SharePoint, Google Drive, GitHub, HubSpot and Asana, and it respects each user's existing permissions — people do not suddenly see things they could not see before.
HOW: Business, Enterprise and Edu workspaces. Admins enable apps in Workspace settings; users authenticate their own accounts. Then open a chat and select company knowledge under the message composer.

Reference files and templates

WHAT: Two different things that get confused. A reference file is an example for the current task. A template is a reference file plus reusable instructions.
WHY: A reference file is for a one-off — "match this deck's master slides, replace the content". A template is for the thing you do every week and want identical every time.
HOW: In Work, attach the file and say explicitly what must stay unchanged: formulas, layout, tone, branding, slide order, table structure. Then save the pattern as a template.

Memory and custom instructions

WHAT: Persistent facts and standing rules that follow you across conversations.
WHY: Memory is the lightest-weight layer and the easiest to leave misconfigured. Note that Temporary Chat deliberately does not read or write memory.
HOW: Managed in Settings. Review what is stored periodically — it is the most common place stale or wrong context quietly poisons later answers.

Connections — apps, connectors and MCP

In December 2025 OpenAI folded "Connectors" into Apps. The term now covers both interactive apps and the connectors that let ChatGPT search and reference your information. Underneath the new naming, the standard is MCP — the Model Context Protocol, an open spec now used across the industry.

Off-the-shelf apps

WHAT: Prebuilt connections to services you already use.
WHY: This is the lowest-effort, highest-return step. It replaces the download-then-upload loop entirely.
HOW: Settings → Apps. Browse the directory, connect the account, complete the OAuth handshake. On Business plans apps are enabled by default; on Enterprise and Edu they are disabled by default and an admin must enable them first.

Custom MCP — your own systems

WHAT: A connector you build, pointed at your own database, API or internal tool.
WHY: This is where it stops being a general-purpose chatbot and starts being your software. Your job costing system, your parcel data, your scheduling engine.
HOW: Enable developer mode in Settings → Apps → Advanced, then add a custom app with your MCP server URL and authenticate. ChatGPT connects to remote MCP servers over HTTPS with user-consented OAuth.

Plan gating — read this before you scope a project

Pro: can connect custom MCP servers with read and fetch permissions only, in developer mode.
Business, Enterprise and Edu: full MCP, including write and modify actions — kick off workflows, create tasks, update your CRM, orchestrate across several apps.
Admins review and publish connectors to the workspace. Only admins and owners can enable developer mode or publish an MCP app after testing.
Company knowledge supports custom MCP apps for search and fetch only. Apps with interactive UI are not currently supported there.

Your server does not have to be on the public internet

This is the question everyone asks and almost nobody gets a straight answer to. If your MCP server runs on a private network, on-premises or on a developer machine, OpenAI documents Secure MCP Tunnel to connect it to supported products without exposing it publicly. Consumer ChatGPT does not reach local servers any other way. If you have heard people talk about running everything over a private mesh network — that is a real technique, but it belongs to your own infrastructure, not to ChatGPT. The supported path in is the tunnel.

Doing things — from answers to output

Finished deliverables

WHAT: Work creates and edits documents, spreadsheets, presentations, reports and analyses — not a description of one, the actual file.
WHY: The output is the artifact. You hand it source material and a format, and it returns something you can send.
HOW: Open Work, describe the deliverable and where it should be created, attach the sources, and state what must not change. Work supports native Google Docs, Sheets and Slides when the Google Workspace app is enabled. The desktop app also supports Microsoft Excel through the ChatGPT for Excel add-in. PowerPoint is not part of the Work desktop flow at launch.

Browsing that survives a login wall

WHAT: A cloud browser that handles supported web tasks, including on sites where you sign in.
WHY: Most of the useful data in any business sits behind a login — supplier portals, carrier systems, county records. Read-only scraping of the open web was never the bottleneck.
HOW: Available in ChatGPT as a documented feature. Note that OpenAI's Atlas browser was retired on 9 August 2026 and browser-based agentic work moved into ChatGPT and Codex. For deeper browser work, use the desktop app.

Local files and folders

WHAT: In the desktop app, Work and Codex can open a local folder and operate on what is inside it.
WHY: Ten thousand job photos with useless filenames. Twelve years of PDFs in one directory. A spreadsheet nobody has been able to clean since 2019. This is what that capability is for.
HOW: Open a local folder in Work or Codex and grant access only to the files the task needs. Be aware: messages and task context may still be stored in the cloud even when the work itself runs locally.

Codex — for when it needs to build the tool

WHAT: A dedicated environment for writing and running code, tests, terminal commands and repositories.
WHY: Non-programmers dismiss this. That is the mistake. If the tool you need does not exist, Codex is how it gets built — a script that renames the photos, parses the PDFs, or reconciles two systems nobody integrated.
HOW: Desktop app, selected from the top-left menu. Codex works with local folders, repositories, terminals and developer tools, and keeps its history separate from ChatGPT.

Running it without you in the room

Everything above still needs you to start it. This is the part that does not.

Scheduled tasks

WHAT: Work that runs once, repeats on a schedule, or monitors for changes.
WHY: The report that lands before the morning meeting instead of being requested during it.
HOW: Set it up in Work and manage it under Scheduled. You can review progress, redirect it, and approve important actions as it goes.

Event-triggered tasks

WHAT: A webhook-based task that fires when something happens in a connected app.
WHY: This is the difference between automation and a reminder. The documented supported events are new Gmail messages, new Slack channel messages, and GitHub pull request activity in an authorized repository.
HOW: Every task has three parts you can open and inspect: Trigger, Condition, Prompt. For a Slack trigger you must add @ChatGPT to every channel the task monitors. Create and edit triggers on web or supported mobile — the desktop app can display them but cannot edit trigger conditions.

Eligible on Plus, Pro, Business, Enterprise, Edu and ChatGPT for Healthcare. Not available on Free or Go, or in FedRAMP workspaces. Enterprise, Edu and Healthcare admins must enable "Allow event-triggered scheduled tasks" first. In Healthcare workspaces these tasks are not covered by a BAA — do not put protected health information through them.

Workspace agents

WHAT: A named, reusable agent your whole company can run — the closest thing to hiring without hiring.
WHY: You build it once, test it, publish it, and the team invokes it. This is the ceiling of what is available without writing your own platform.
HOW: Business and Enterprise workspaces. Choose its model and reasoning effort, connect it to apps and tools, share it with the workspace, run it in Slack, put it on a schedule, or fire it from an API trigger so an internal system can start it. Run it in chat by typing @ and the agent name. Off by default at launch for Enterprise; an admin enables it.

The governance that makes this defensible

Admins get a consolidated view of every agent in the organisation: Agent ID, recent activity, connected apps, memory files, schedules, and analytics including unique users and runs over time. If you are ever asked how a number was produced — by a carrier, an auditor or an attorney — that record is the answer. Build with it on from day one, not after the first incident.

Repeatability — skills and plugins

A prompt is a performance. A skill is a process. The gap between a company that plays with AI and one that runs on it is entirely in this section.

Skills

WHAT: Reusable, shareable workflows. A skill can contain instructions, examples, and code.
WHY: Once installed, ChatGPT can pick up one or more skills automatically when they are relevant — you do not have to remember to invoke them. Your estimating standard, your tone, your carrier-letter format: encoded once, applied consistently.
HOW: Available to eligible Business, Enterprise, Healthcare and Edu users, subject to workspace settings. Permissions control who can create, use, upload, share, publish or install them. Admins get a Skills page showing each skill's owner, access, users, invocations over 30 days, and dates.

Plugins

WHAT: A bundle. A plugin packages skills, app integrations and MCP servers into one installable unit.
WHY: This is how a workflow travels — between your own machines, across your team, or out to a client. It is also the distribution model if you ever want to sell a workflow rather than perform it.
HOW: In the Codex app, open Plugins to browse and install. In the Codex CLI, run /plugins. Plugins can be published through a marketplace source such as a team repo. Workspace admins manage installation in Workspace settings → Plugins.

Remote control

WHAT: Running and supervising Codex work from your phone while the actual work happens on a host Mac.
WHY: The reason this matters for anyone who runs a field business: you are not at a desk. You should not have to be, to approve a step.
HOW: Codex in the ChatGPT mobile app lets you answer questions, redirect work, approve actions and review outputs on a connected environment. The mobile view reflects the live state — project context, approvals, screenshots, terminal output, diffs and test results. Codex also supports access tokens so approved automation can run non-interactively through scripts, schedulers or private CI runners under a workspace identity.

The setup checklist

In order. Each step is worth doing even if you never do the next one.

Settings → Data Controls. Turn Improve the model for everyone off. Two minutes.
Decide what belongs in a Business workspace rather than a personal account. Anything with a customer's name on it.
Learn the Chat / Work / Codex switch. Stop doing multi-step work in Chat.
Create one Project per function. Load it with your real files — price sheets, contracts, SOPs, templates.
Connect two apps. Whichever two you touch most. Usually mail and file storage.
Enable company knowledge if you are on Business or Enterprise, and ask it a question you would normally ask a person.
Save one format you produce weekly as a template. Run it once from the template to prove it holds.
Put one recurring report on a schedule.
Add one event trigger. Start with inbound email — it is the highest-volume, lowest-risk surface you have.
Identify the one internal system nothing talks to, and scope a custom MCP connector for it. That is the real project.

Translated: a contracting or roofing business

None of these are features you switch on. They are things you would wire using the pieces above. Listed so the capability map has somewhere concrete to land.

THE PROBLEM	WHAT YOU WOULD USE
Every estimate is written from scratch	Project + your historical estimates as files + a saved template
Adjuster scopes take a day to review	Work session with the carrier PDF and your estimate attached, output as a supplement letter draft
Nobody knows what was promised on a job	Company knowledge across mail, file storage and CRM, answering with citations
Leads sit in the inbox overnight	Event-triggered task on new Gmail messages, with a condition that filters to real inbound
The Monday numbers take an hour to assemble	Scheduled task that runs before the meeting and produces the file
Job photo archive is unusable	Desktop Work or Codex on a local folder — rename, sort, extract
Your CRM is a system nothing else talks to	Custom MCP connector, Business or Enterprise for write access
Every rep writes customer updates differently	A Skill encoding your voice and format, published to the workspace
Collections are inconsistent and awkward	Workspace agent with your AR data and your escalation rules, run on a schedule

Where this breaks

Anyone selling you this without the following section is selling you something.

Rollouts are uneven. OpenAI ships gradually. Availability can differ between Chat, Work and Codex, between plans, and between web, mobile and desktop — for the same account on the same day.
Names change under you. Connectors became Apps in December 2025. The Atlas browser was retired on 9 August 2026. Documentation that is six months old is not documentation.
MCP reliability is not uniform. Developers report connectors that complete OAuth but never surface tools, and apps that intermittently vanish from directories. Tool metadata is frozen at approval — a live server that drifts from the approved snapshot can start erroring on calls.
Connectors are an attack surface. Independent security researchers have demonstrated data-exfiltration vectors against connector flows. Least privilege by default. Read-only unless write is genuinely required. Vet third-party servers the way you would vet a vendor with database access, because that is what they are.
Setup is a technical task. Marketing says "connect your tools". The reality is closer to "deploy and vet an integration" — a remote HTTPS server, OAuth configuration, and scope review.
Agent mode does not use custom apps. Deep research can, but for read and fetch only. Know the boundary before you design around it.
The model is still fallible. Anything touching insurance, contracts, payroll, safety or a regulator gets a human signature. Use company knowledge specifically because it cites its sources — then check them.

Sources

Every claim in this guide traces to OpenAI's published documentation, read in September 2026. Primary references:

help.openai.com — ChatGPT Work and Codex
help.openai.com — Creating and editing documents, spreadsheets, and presentations with ChatGPT Work
help.openai.com — Data Controls FAQ
help.openai.com — How your data is used to improve model performance
help.openai.com — Projects in ChatGPT
help.openai.com — Apps in ChatGPT
help.openai.com — Company knowledge in ChatGPT (Business, Enterprise, and Edu)
help.openai.com — Developer mode and MCP apps in ChatGPT
help.openai.com — Skills in ChatGPT
help.openai.com — ChatGPT Workspace Agents for Enterprise and Business
help.openai.com — Scheduled tasks in ChatGPT
help.openai.com — ChatGPT Business General FAQ
help.openai.com — ChatGPT, Business and Enterprise release notes
developers.openai.com — Codex plugins

Use of this document

Free to share, forward, print and hand to your team. No gate, no email capture, no attribution required. If it is wrong somewhere, tell me and I will fix the next version — the whole point is that it stays accurate.

Download the PDF