ChatGPT Work: internal demos show meeting briefs to strategy decks

Cutting half-a-day of customer-meeting prep down to five steps

Alex, on OpenAI's sales team, says he used to spend half a day on a single customer call. Before the call, he'd dig through account history; after it ended, he'd write up notes, polish a follow-up email, and manually enter updates into the CRM. In a tutorial video OpenAI posted on August 18, Alex compresses that work into five steps and shows the whole thing playing out on his own screen.

That same day, OpenAI posted three videos three minutes apart. Alex from sales builds a customer meeting brief, Sahil from marketing drafts a launch blog post, and Arvind from strategy and business operations builds a strategy deck for leadership. All three break down their own process as "here's the order I use it in." Below are the steps shown in the three videos, the actual prompts typed in, and the full text of a separate set of sales prompts OpenAI released — kept in the same order so you can follow along.

When to use ChatGPT versus ChatGPT Work

ChatGPT Work is a work-focused mode OpenAI unveiled on July 9. It's not a separate app — it's a mode switch within the same ChatGPT, running on GPT-5.6 and inheriting technology from the coding agent Codex. OpenAI's own distinction is simple: quick answers, brainstorming, and short drafts are Chat. When ChatGPT needs to gather context from scattered sources and go through multiple steps to produce a reviewable finished artifact — a brief, an analysis, a deck, a spreadsheet — that's Work.

Three new mechanisms come with Work. Plugins connect to where your data lives — Slack and Microsoft Teams, Google Drive and SharePoint, email and calendar, CRM, and project management tools. The list OpenAI showed in its public demo included Google Drive, Outlook, Gmail, Teams, Slack, SharePoint, Salesforce, and Adobe Acrobat. Scheduled tasks let you choose among four options: run once, repeat on a schedule, trigger on a specific event, or watch for changes. Approval stays with the user. OpenAI wrote: "You decide what it can access, when it checks in with you, and when it needs your sign-off before taking action."

Access path	Availability
Desktop app (Mac/Windows)	Global rollout July 9. Chat, Work, and Codex available on all plans, including the free tier
Web/mobile	From July 9 for Pro, Enterprise, and Edu
Web/mobile (Plus, Business)	Rolled out a few days later

The cheapest way to start is the desktop app — Work mode is available even on the free plan.

How to try it — initial setup and your first task

All three videos share four common setup steps.

1. Switch modes. Open the desktop app or go to chatgpt.com and select Work from the mode picker. Alex starts on the web; Sahil and Arvind start in the desktop app. The same switch works on mobile.
2. Connect where your data lives. In the left sidebar, go to Plugins and add the tools your company uses one by one. ChatGPT only reads what's connected here — it can't see content in tools you haven't linked.
3. Install the plugin for your function. For sales, search Plugins for "Sales," open it, follow the setup instructions or connect your accounts, and check the list of connectors and skills. Sales plugins connect to Salesforce, HubSpot, Slack, Outreach, Clay, Rox, and Actively. If you use a call-recording tool like Gong, connect it separately, as Alex does.
4. Pick your first task. If you're not sure where to start, use the prompt below from OpenAI Academy's sales-team resource guide, posted August 5.

A prompt to help decide what to hand off first

This prompt scans your connected tools and surfaces three candidate tasks suited to ChatGPT Work.

Look through my recent activity in the tools I've connected. Pick three sales tasks that would be a good fit for ChatGPT Work. For each one, tell me what you could produce and recommend where to start. Don't begin any work until I've chosen which one to do.

That last sentence matters. Drop the condition telling it not to start until you've chosen, and ChatGPT will just run the first task immediately.

Sales — Alex's 5 steps to prep for a customer meeting

In the sales team demo video, Alex breaks his process into five steps.

1. Open ChatGPT and select Work. Same in the desktop app.
2. Connect the tools that hold customer information. Alex connects six: Calendar, Gmail, Slack, Gong, Google Drive, and Salesforce.
3. Specify in advance what's needed for each meeting. Alex specifies account history, unanswered open questions, and key people attending the meeting. On the go, he dictates briefs, slides, and CRM updates from his phone, and the format comes out consistent every time.
4. Describe what should happen after the call ends. Alex has it do four things — create meeting notes, draft a follow-up email, propose updates to enter into the system, and pull out next steps for the rest of the team.
5. Review and approve any customer-facing message and any significant change before it goes out.

With this set up, ChatGPT gathers account context before the meeting and produces a pre-meeting brief. The brief goes not just to Alex but to everyone listed on the internal invite. On the go, Alex's prompt is short:

Pull the account brief for this meeting. I'm heading there now so I can read it on the way.

Marketing — Sahil's 5 steps to draft a launch blog post

For Sahil, a launch blog post is the blueprint for a campaign. He used to spend days scraping information from Slack conversations, meeting notes, product docs, and user research, then translate all that into "why this matters to users." In the marketing demo video, he says he now finishes that first step on his commute. The material is his real work — he's writing the launch blog post for ChatGPT Work itself.

1. Open the desktop app and switch to Work. Also works on chatgpt.com and mobile.
2. Connect source material via plugins. He connects the messaging framework, product docs, research, and meeting notes. At this step, Sahil uses voice instead of typing — thinking out loud is faster, he says, and feels more like explaining something to a colleague.
3. Provide a template and examples. He feeds in the company's blog template along with a few previously published posts so the model understands format, structure, tone, and what counts as good writing.
4. Fill in interactive sections separately. Sahil's company blog isn't a block of text — it has components readers click through. He hands over customer interview transcripts he had stored on his desktop and specifies the shape of the quotes he wants and the narrative flow. He used to spend hours listening to transcripts to pick out usable lines; now, quotes matching his criteria come back in minutes.
5. Set it to keep watching the sources. He has it propose revisions to the blog post if launch details change.

At step 2, the voice prompt Sahil gives is one line:

Pull the latest context from these sources and put together a first draft of the blog post.

Strategy — Arvind's 5 steps from one question to a finished deck

The strategy demo video opens with a single question: "Which market in Latin America should we prioritize next?" Arvind says building a deck like this used to take days.

1. Open the desktop app and switch to Work. Also works on the web.
2. Pin down the business question you're trying to answer first. He writes the question down before gathering any material.
3. Connect tools via plugins and add the company's slide template. Arvind connects market research, customer data, and past strategy documents.
4. Get a plan and storyline before the deck itself is built. He has it propose the narrative direction before any slides are drafted.
5. Review the storyline, verify numbers and sources, refine the recommendation, and approve the final version.

Once the storyline is set, the last prompt Arvind enters directly specifies the table of contents for the final output.

Pull all of this together into a market-expansion readout. Include market context, customer analysis, supporting tables, and recommended next steps.

Six shared columns across the three demos

Different departments, different outputs — but the same sequence. Lining up all three videos side by side shows what repeats.

Column	Sales (Alex)	Marketing (Sahil)	Strategy (Arvind)
Where to open it	Select Work in ChatGPT (desktop also works)	Switch to Work in desktop app (web/mobile also work)	Switch to Work in desktop app (web also works)
What to connect	Calendar, Gmail, Slack, Gong, Google Drive, Salesforce	Messaging framework, product docs, research, meeting notes	Market research, customer data, past strategy documents
What to pin down first	Requirements for each meeting (account history, open questions, key people)	Request for a first draft (by voice)	A single business question
What framework to provide	Fixed format for briefs, slides, CRM updates	Blog template and prior post examples	Company slide template
What to review before final output	Pre-meeting brief	First draft with structure and copy in place	Plan and storyline
What the human does at the end	Approve customer-facing messages and significant changes	Review and revise the draft, reflect launch changes	Verify numbers and sources, then approve

Columns three and four are what determine output quality. Specifying the question and format up front fixes the shape of what ChatGPT will produce, turning every following step into just pouring material into that shape. Column five is where an intermediate output gets checked before the final one. Arvind sees the storyline before any slides are drawn; Sahil sees a first draft before the finished copy.

Five prompts ready to copy and use

Here are prompts OpenAI Academy released for sales teams, translated in full. They're more tightly specified than the short lines shown in the videos, and tend to fail less often on first use. Swap out anything in brackets for your own company or account names.

A deal brief to decide where to focus today

This prompt ranks which deals to look at today, complete with reasoning and next actions, packaged into a document. @Sales invokes the installed sales plugin, and the last two sentences act as a safeguard against contacting customers or modifying the CRM prematurely.

@Sales Review my open opportunities and determine where I should focus today. Create a Google Doc. Rank the opportunities, and for each one explain why it needs attention now, the supporting evidence, and the recommended next action. Only include information backed by sources. Then send a short summary and the document link to Slack. Don't contact any customers or update the CRM.

If the output feels flat, add one line to redirect it: "Add a 'what I need to verify' section for each opportunity." If you want the same brief every morning, follow up with: "Run this task every weekday morning at 7:30 AM" to lock it in as a scheduled task.

A strategic account plan for internal review

This prompt builds a presentation on a single account's status, opportunities, risks, and next actions, matched to the company's template. It's Arvind's five steps applied to a sales account — the key lines are specifying the template name and flagging items that still need verification.

Build a presentation for an internal-review strategic account plan for our [Acme] account. Explain what we know about this account, where the opportunity stands, what's at risk, and how we should move forward. Only include information backed by source material, and flag anything that still needs verification. Use the [Acme Strategic Account Plan template]. Make the narrative clear, use charts or tables where helpful, and include our recommended next actions.

A verification prompt to trace back the source

This prompt is for not taking a draft at face value — it targets a single figure and asks where it came from.

Where does this claim come from? [paste a claim from the presentation here] Show me the exact source and give me a link straight to that spot.

A rewrite prompt that applies company style all at once

Instead of repeating the same corrections every time, this prompt bundles the company's formatting and structural rules into a single list fed in at once. The items below are OpenAI's examples — swap in your own company's rules.

Rebuild the presentation to match these company rules:
- Drop the table-of-contents slide.
- Put the key recommendation first.
- Use the exact stage names from the CRM for open opportunities.
- Clearly flag assumptions and missing information.
- Put sources in footnotes.
- Use red only for genuine risks or blockers.
- In leadership reports, say "customer" instead of "account."
Keep following the account plan presentation template throughout.

Turning a workflow you've built into a reusable skill

This prompt saves the sequence you've just built as a named skill you can reuse on the next account. Once saved, you don't have to re-explain the template and rules every time.

Turn this workflow into a skill called strategic-account-plan-builder. Include the Acme account plan presentation template and the rules I gave you. Use this skill for all future strategic account plan presentations. Ground every claim in source material, and flag anything that needs verification.

When something goes wrong, the fix has a known location. If facts are wrong, go back to step 2 and check whether the data is actually connected. If the format is off, adjust the template in step 4. If the direction is off, the question in step 3 was too vague. When moving to a different department, only three things need swapping — the tools to connect, the template file, and the initial question — while the sequence stays the same. Worth carrying over too are the review items OpenAI attached: cross-check facts against the company's original records, keep customer-facing writing in draft form until approved, and check your company's AI, privacy, and security policies first.

Editor's view

OpenAI posting three departments side by side on the same day is a positioning move to brand ChatGPT Work as a department-specific work tool — the same script Microsoft Copilot and Salesforce Einstein ran first. But these videos showed sequence instead of a feature list. All three people number their screens 1 through 5 and narrate in that order.

Putting tools like this into actual practice tends to land on the same conclusion every time: the quality of the first output depends less on model performance than on connections and format. How much data got linked, and whether the shape of the output was specified in advance, determines everything else. That's why all three people in the videos connect tools and set the format immediately at step one. Skip the connection step and just write a long prompt instead, and the model fills in the blanks with generic knowledge, producing a plausible-looking document. The moment fixing that document takes longer than writing it by hand from scratch, the rollout has failed.

What teams here can try right now is the desktop app. Work mode is available even on the free plan, so one person can connect one task of their own without paying for anything. But connecting things at all raises the issue of exposing company data to an external tool first. Rather than linking CRM and internal documents wholesale right away, it's safer to run a week with just a personal calendar and email connected, and build a list of what turns out accurate versus off. Designing the approval step first matters for the same reason. All three videos end with human review as the final step — and outputs that go out without review are the most common failure in early AI rollouts.

More videos in the same format — for customer support, HR, finance — are expected in the coming weeks. Ahead of developer-conference season, OpenAI will likely keep stacking up department-specific use cases, and department-specific plugins, like the sales one, will keep getting added one by one. Learn the sequence with one department now, and when the next one arrives, all that needs swapping is the connections and the template.