Lesson 08 - Integrations
ChatGPT Work + your company apps
AI does not work in a vacuum. It can use the tools you already have - if there is a connector, an API, or safe automation and the right admin approvals.
Slack
Reads channels, summarizes threads, sends alerts.
Google Workspace
Auto-replies in Gmail, Sheets analysis, generating Docs.
Make.com / Zapier
Form -> ChatGPT -> CRM -> email to the salesperson. No code.
HubSpot / Salesforce
Auto-updating records, lead scoring, follow-ups.
Notion / Confluence
Auto-generating documentation, summaries, a knowledge base.
GitHub / GitLab
Code review, tests, and preparing changes. Leave merging or deployment for human approval.
Ready-made recipes - what to connect ChatGPT Work with
Four combinations you can implement in your first month. Each recipe: tools + what you get out of it.
Morning inbox review
Every day at 8:00, ChatGPT Work reads new emails, pulls tasks and deadlines into a table, and drafts replies for urgent messages. You just approve sending.
Friday sales report
On Friday at 16:00 the agent pulls data from the CRM, calculates pipeline and conversions, generates charts, and posts a summary to #sales. Zero manual Excel.
Automatic code review
Every new pull request gets a review: bugs, risks, suggested fixes. The merge decision always stays with a human.
Meeting notes -> tasks
After a meeting you paste in raw notes. ChatGPT Work creates a Notion page with decisions, a task list (who, what, by when), and suggests calendar entries.
Tool cookbook
Six briefs you can use at work today today
Start with an outcome, name the sources and finish with a safety boundary. Replace the text in brackets and copy the brief.
Pre-meeting client brief
Follow-up and opportunity update
Cost and variance review
Source-backed market research
Weekly project summary
New-lead qualification
- 1. Grant the minimum permissions required.
- 2. Test on known, non-sensitive data.
- 3. Require approval before sending or saving.
More detailScheduled tasks and monitoringAfter testing a connection, you can add a schedule. See task types and the Scheduled screen here.
Scheduled tasks and monitoring
When should the process run itself?
The most useful of the three is the quietest one: monitoring that says nothing until something actually changes.
A task at a specific time
Once, at a set time
You run the process at a specific moment because that is when the data is complete or when the result is needed. One-off or on a chosen day.
Watch out: if the data lands later, the task will compute an incomplete period.
A recurring task
Daily, weekly, monthly
The same process on a fixed rhythm. The value comes from comparability - every result has the same structure, so change over time is visible.
Watch out: a recurring task without QA repeats its mistake too - just regularly.
An event-triggered task
When something happens
No clock involved - the process starts in reaction to an event: a message arrives, a file appears, someone submits a form, a record changes. The reaction is immediate because you are not waiting for the next time slot.
Watch out: events arrive in bursts. Decide what happens when ten land at once, so you do not get ten separate messages.
Condition monitoring with change-only alerts
Checks often, speaks rarely
The process checks the state and stays silent while nothing changes. An alert arrives only when the condition is met - and it contains a before/after comparison.
Watch out: without a definition of "material change" you will get an alert for every typo on a page.
Five examples you can set up this week
Morning news review
One message each morning with at most five things that can affect your work.
Every day at 7:30 review the sources [LIST] from the last 24 hours. Pick at most 5 items relevant to [INDUSTRY]. For each: one sentence of fact, one sentence of "what this means for us", a link. Do not summarize material you could not open. 250 words total maximum. Show me the first 5 runs for approval before posting to the team channel.
Weekly report
A ready weekly summary on Friday afternoon with deviations and decisions to make.
Every Friday at 15:00 pull data from [SOURCE] for last week and the week before. Compute the metrics [LIST] and deviations vs. the prior week and the quarterly goal. List at most 3 decisions to make (what, who, by when) and 3 things to watch. Label every deviation above [THRESHOLD %] with a cause hypothesis, clearly marked as a hypothesis. QA: check that the parts sum to the whole and that the date range matches last week's. Send the report to me first - do not forward it without approval.
Price monitoring
Silence until a price changes. The alert contains the old price, the new price and a link.
Once a day at 7:00 check the public pricing pages [URL LIST]. On the first run record the baseline. On later runs compare with the previous state and send NOTHING if the price did not change - just append a history row. If a price moves by more than [THRESHOLD %], send an alert: old price, new price, date spotted, link. Do not bypass logins or paywalls. Before enabling automatic alerts, show me the baseline and the first comparison.
Competitor page monitoring
Alert only on a material change: packages, offer headline, a new feature, a new market.
Check the public pages [URL LIST] daily. Material changes are: pricing, packages, offer headline, a new feature, a new market, hiring changes. Ignore cosmetic changes (typos, photos, reordered sections). On a change describe: what it was, what it is, when spotted, the link and a hypothesis of what it means - clearly marked as a hypothesis. Check whether the change comes from an A/B test or a regional variant. Do not contact the competitor and do not create accounts.
New lead monitoring
A new lead matching your criteria reaches you with context and a draft reply - nothing sent. If the source can announce an event, use an event trigger and react instantly. If it cannot, fall back to a recurring check.
Run on every new submission from [SOURCE - form / mailbox / CRM view]. If the source cannot announce an event, check it every hour instead. If a new lead matching [ICP] appears, prepare: a one-sentence summary of the enquiry, at most 3 public signals about the company with links, a suggested priority and a draft reply (max 120 words). Save the draft as a draft - send NOTHING and change no CRM records. If there are no new leads, send no message at all.
Availability depends on your plan, region, workspace and administrator configuration.
How it fits together
Plugin, App, Skill, Scheduled Task
Four layers, four different questions. People mix them up constantly - and then wonder why the automation does nothing.
PLUGIN
What can the agent do at all?
A package of workflows and capabilities - a set of abilities added to the agent as one bundle.
e.g. A "Reporting" package bringing a complete set of reporting workflows.
APP
What can the agent reach?
The connection between ChatGPT and a system, its data and its actions - mail, drive, CRM, calendar, knowledge base.
e.g. A connected company drive the agent reads exports from.
SKILL
How do we do it here?
The instruction for how your company performs one specific process - your definitions, your step order, your rules.
e.g. "Weekly sales report" with your metric definitions.
SCHEDULED TASK
When or under what condition does it run?
The moment or condition that starts the process - a specific time, a recurring rhythm or a met condition.
e.g. Friday 15:00, or "the price moved by more than 5%".
Plugin gives capability. App gives access. Skill gives your way of working. Scheduled task gives the moment. Miss one layer and the automation either cannot act, acts wrongly, or never starts.
Codex C5 - Scheduled tasks
Scheduled - work that runs without you
In short ↳You set a task once and Codex repeats it: a Monday audit, a daily report, monitoring a page for changes. The result waits for you in a chat.
Where to find it: Codex → Scheduled
Select the screenshot to enlarge it.
What you see on screen
- 1
Create
A new recurring task: what to do, how often, where to leave the result.
- 2
Monitoring
The first entry shows Monitoring and a next-run time. Whether unchanged results generate notifications depends on the task configuration.
- 3
Weekly task
'Mondays at 8:00' - the classic: an audit, a sales summary, a list of things left hanging.
- 4
Daily task
A short daily briefing at a fixed hour. Better than a long weekly report nobody reads.
- 5
Active / Paused / Completed
Pause tasks during holidays and delete ones you stopped reading - otherwise they only burn your limit.
- 6
Next run
Shows the planned next execution. It does not confirm that the previous run succeeded; open the task and inspect its result or error.
- 7
All and search
All lists tasks across states. Search filters names; Active, Paused and Completed narrow the list by state.
- 8
Cloud symbol and schedules
The screenshot contains cloud marks and relative run times. Inspect task details for execution requirements, time zone and notifications; the symbol alone does not guarantee a run while your computer is off.
How to use it - in plain words
Start with one task that you currently do by hand every week. Run it for a month, then decide whether it is worth keeping.
Every scheduled task needs an owner - a person who reads the result. A report nobody opens is not automation, it is noise.
For advanced users: example prompt+
Scheduled task: Mondays at 8:00. 1. Check what changed in the project over the last 7 days. 2. List: what was finished, what is stuck, what is at risk. 3. Maximum one page, no jargon, ready to paste into Slack.
Try it on your account
Find Paused, Next run and Create. Check your task time and time zone before saving it.
More detailUse ChatGPT in WordOpen this if you use Word. The example covers the add-in, access and reviewing a document.
ChatGPT in Word
The document stays in Word, the assistant sits beside it
Before sharing, compare names, numbers and commitments with the notes. Save a copy before major edits. Accept only changes you have checked.
One add-in, three apps
The same add-in covers Word, Excel and PowerPoint. You install it from the Microsoft Marketplace and open it from the ribbon.
All plans, Free included
The add-in works on every ChatGPT plan. Usage is metered in tokens, and a long document costs more than editing a selected sentence.
Two gates in a company
A Microsoft 365 administrator has to allow the add-in, and a ChatGPT administrator decides whether Word may use the workspace. From 1 October 2026 access is on by default on the ChatGPT side, but an administrator can turn it off. A missing button is usually one of those two gates.
The open document and connected sources
The pane sees the document you have open plus sources connected in ChatGPT (for example Outlook, SharePoint, Google Drive) - it does not scan the other files on your drive.
Three steps, in this order

01
Check that the add-in is installed
Add-ins → My Add-ins. If ChatGPT is missing, check the installation and signed-in account, then ask your Microsoft 365 administrator to check add-in availability.

02
Open the document and the pane
Clicking ChatGPT in the ribbon opens the side pane. You work on the open file, so you do not copy content into a separate browser tab.

03
Order the work and read the result with the comments
The prompt reads like a brief for a person: what to fill in, from where, and what to flag when data is missing. The reply says what was done, and the remarks land in the document as comments.
Screenshots from our training: Word together with the ChatGPT add-in.
Availability depends on your plan, region, workspace and administrator configuration.
What the pane actually does well
A first draft from notes
You paste notes and source material and get a draft inside the document.
A summary and an executive summary
A short version of a long document without retyping it.
The same content for another reader
Rewriting a section for a client, a board or a technical team.
Tidy structure
Headings, numbering and section order fixed in one pass.
Consistent terminology
Spotting unclear sentences and names that mean different things across the document.
What the screenshot really shows
“Filled 11 NDA fields with assumed details” - and that word does the work
The pane filled eleven fields of the contract and kept the three-page layout. It also wrote down, in comments, what still needs a human: the entered details require confirmation, the counterparty's verified address is missing, and both signatures and signing dates are absent. Read that as a template for your own work - ask for the comments on purpose, because without them you get a document that looks finished and is not.
- Every assumed field needs a comment - otherwise the assumption poses as a fact.
- Missing data stays missing, not a value from the model's memory.
- Layout and clauses stay untouched until you decide otherwise.
The add-in or Copilot
Not either-or. Copilot is woven into the Microsoft ecosystem; the add-in gives you the assistant you already pay for, inside the file. What separates them is the account, the licence and the reach into company data.
What you sign in with
A ChatGPT account
A Microsoft 365 account with the Copilot entitlement
How it is billed
Your ChatGPT plan, usage metered in tokens
A Microsoft 365 licence
Integration with the company stack
The open document plus sources connected in ChatGPT
Deep in Outlook, Teams, SharePoint and OneDrive
Who controls access
The Microsoft 365 admin and the ChatGPT admin
The Microsoft 365 admin
Three prompts you can use today
A contract review that flags gaps
Fill the template and list what is missing from the source material.
Fill in this contract template with the details we agreed in [SOURCE, e.g. the email thread / a file]. Mark every field with a comment when it comes from an assumption rather than a source document. Separately list everything missing before signing, and do not put any remembered value in its place. Do not change the layout or remove any clause.
Rewriting a section for another reader
A different tone and level of detail without losing facts.
Rewrite the [SECTION] so that [AUDIENCE, e.g. a non-technical client] understands it. Keep numbers, names, dates and commitments unchanged - list them at the end as a checklist. Cut at most a third, add no new facts, and remove no conditions or exceptions.
A language check before sending
Inconsistent names, unclear sentences and promises without cover.
Review the document and list in a table: unclear sentences, the same term spelled in different ways, and sentences promising something the content does not deliver (a date, a scope, a result). Give the location and a suggested fix for each. Do not edit the text before I agree - list first, then decide.
Before the document goes out
Five checks that separate
a draft from a document
01
Fields filled from assumptions
Which numbers, dates and names lack a source you can compare them with?
02
Completeness
Is anything missing: an annex, a clause, a signature, a signing date?
03
People and scope
Do the parties, addresses and the scope of the obligation match?
04
Structure
After the edits, do headings, numbering and the table of contents still agree?
05
Copy and data
Save the source version before a big edit and check what may be sent to the tool.
The pane edits the file you have open, and mistakes are part of the deal - check facts, figures, citations and every field you did not supply yourself, and keep a copy of important work before a large edit. On our training we never open real contracts or client documents in the add-in, we work on a template and sample data.
Sources: chatgpt.com/apps/word and help.openai.com (17 Sep 2026)