ONLINEAGENT_OPS 2026.Q3 HOME HUBSMODELSBUILDINGCOMPANIESFILMMAKINGSECURITYLIVING WITH ITSKILLSALL HUBS, ONE TERMINAL → ARTICLES CRAFT RECORD BLOG TOOLS SKILLS SEARCH /
HOME›ARTICLES›AI IN WORK APPS: WHAT COPILOT, GEMINI AND SLACK
ARTICLES · GUIDEBEGINNER

AI in Work Apps: What Copilot, Gemini and Slack AI Can See

What the AI in Microsoft 365, Google Workspace, Slack, Zoom and Notion can see, what each vendor says on training, the main risks, and a six-point check.

READ11 min
WORDS2,149
SECTIONS7
TYPEGUIDE
CHECKED13 OCT 26

The AI buttons in work apps such as Microsoft 365, Google Workspace, Slack, Zoom and Notion can read whatever you can read: your files, email, chats and meetings. Each of these vendors says, on its own pages, that the AI shows only what you already have permission to see.

Each also says your work content is not used to train its AI models. So the main risks are not the vendor training on your data. They are files shared too widely, summaries that get things wrong, and hidden instructions inside shared documents. This guide explains each one and ends with a six-point check before you use one on client work.

TL;DR — THE SHORT VERSION
  • What it sees: everything your account can open. Nothing more, but also nothing less.
  • Training: all five vendors here say work content is not used to train their AI models, or not without permission. That covers the work versions your employer manages.
  • Risk 1, oversharing: a folder shared with “everyone” years ago is now one question away.
  • Risk 2, wrong summaries: a confident recap can drop a decision or name the wrong person.
  • Risk 3, prompt injection: a document can carry hidden instructions the AI may follow.
  • Admins hold the switches: whether it is on, who gets it, and how long prompts are kept.
◈ IN PLAIN TERMS

A work AI assistant is like a very fast colleague holding a copy of your office keys. It can only open doors you can open. But it opens them in seconds, including the ones somebody left unlocked by mistake years ago.

What can the AI in your work apps see?

The plain point: each assistant draws on the content your account can already reach in that app, plus your question. The table sums up what each vendor says, read on 13 October 2026.

AI in work apps: what each vendor says
AppWhat it can draw onTraining on your contentWho controls it
Microsoft Copilot (Microsoft 365)Draws on: emails, chats, documents, calendar, meetings and contacts you can access; the web, if allowedSays: not used to train the underlying AI modelsAdmins: data retention, which add-on agents run, web search
Gemini in Google WorkspaceDraws on: Workspace content you have permission to access, such as Gmail, Docs and DriveSays: not used for training outside your organisation without permissionAdmins: turning it on or off, how long chats are kept
Slack AIDraws on: channels and messages you can see, plus files and linked documentsSays: never used to train the AI models; separate opt-out for Slack’s other modelsOwners and admins: which AI features members get
Zoom AI featuresDraws on: meeting audio, transcript and chat, for summaries and questionsSays: no audio, video, chat or similar content used for trainingAdmins and hosts: features by account, group or user; hosts per meeting
Notion AIDraws on: workspace pages you can already accessSays: not used to train any models, by defaultWorkspace owners: web search and data-retaining AI settings

Microsoft. Microsoft 365 Copilot is now called Microsoft Copilot, Microsoft says, with no change to its privacy terms for organisations. It works with “emails, chats, and documents that you have permission to access”, and Microsoft says prompts, responses and that data “aren’t used to train foundation LLMs” (large language models, the AI behind it). Admins can view stored prompts and responses with Microsoft’s search and compliance tools and set how long they are kept.1

Google. Google says Gemini in Workspace will not “Access Workspace content that you don’t have permission to access”, and that your content is not human reviewed or used for AI training “outside your domain without permission”. Access to the Gemini app is on by default; admins can switch it off. Admins also decide how long prompts and responses are kept.2

Slack. Slack says its AI features “only use Slack data that members have access to at the time of request” and that customer data “is never used to train large language models”. One catch: Slack also runs other, non-generative models (for channel and emoji suggestions, for example), and to keep your data out of those your workspace has to opt out.34

Zoom. Zoom says it does not use “audio, video, chat, screen sharing, attachments, or other communications-like customer content” to train its own or third-party AI models. Admins turn its AI features on or off for the whole account or for groups, and users can sometimes choose for themselves.56

Notion. Notion says its AI “honors existing permissions” and that “By default, Notion and its AI Subprocessors do not use Customer Data to train any models.” Some features may need AI providers that keep data; workspace admins have to switch those on.7

These promises are for the work editions your employer pays for and manages. Google’s privacy hub, for example, applies only with a qualifying Workspace edition.8 A personal chatbot account follows consumer rules, which what happens to what you type explains.

TAKEAWAY

Read the vendor’s own privacy page for the app you use, and check you are on the work version. Training is rarely the problem; reach is.

Risk 1: oversharing through old permissions

The plain point: the AI respects your permissions, but many permissions are wider than anyone meant. The AI makes those mistakes easy to find.

Microsoft puts it plainly: Copilot “only surfaces organizational data to which individual users have at least view permissions”, so it is important that “the right users or groups have the right access to the right content”.9 Its own setup guide starts with “remediating oversharing”.10

Here is how it goes wrong, in this site’s reading. A salary spreadsheet sits in a folder shared with the whole company. Nobody ever stumbled on it. Then someone asks the assistant “what are people paid in my team?”, and it finds the file in seconds.Reasoning, 13 Oct 2026 — an illustration of the permission rule the vendors describe, not a reported case. Slack draws the same line: its AI “will never surface any results that Slack’s regular search would not”. The AI is no leakier than search. It is just much better at finding things.3

TAKEAWAY

If you find something you should not see, report it to IT. Do not use it, and do not ask follow-up questions about it.

Risk 2: summaries that sound right but are wrong

The plain point: a meeting recap or thread summary is a first draft. It can leave out the one decision that mattered, or credit the wrong person.

Microsoft says generative AI responses “aren’t guaranteed to be 100% factual” and asks users to review them “before sending them to others”.9 Slack adds citations so you can “jump to the original message and review it”.3 Use them. Errors cluster in the same places as anywhere else AI writes: names, numbers, dates and who agreed to what (why AI makes things up).

Meetings add a consent question. When Zoom’s meeting summary or meeting questions are on, attendees see a lit AI icon, and anyone can ask the host to turn the features off and delete what was generated so far.11 With outside guests, Zoom says shared content may be processed under the other side’s account settings (source above), so a client’s own AI may be summarising your call.

TAKEAWAY

Read the summary against your own memory or the source before you forward it, and say at the start of a call when AI is taking notes.

Risk 3: hidden instructions in shared documents

The plain point: the AI reads documents and emails as text, and it can mistake text inside them for instructions. This is called prompt injection.

The security group OWASP calls this “indirect” prompt injection: it happens “when an LLM accepts input from external sources, such as websites or files” that change how it behaves.12 A supplier’s document could hide a line telling the assistant to recommend that supplier, or to include a link in its summary. The UK’s NCSC says current AI models “simply do not enforce a security boundary between instructions and data”, and that prompt injection “cannot be fully mitigated with a product or appliance”.13

The vendors build defences. Microsoft uses classifiers to block “cross-prompt injection attack (XPIA)”, but notes they “may not be available in all Microsoft Copilot scenarios”.9 Google says Gemini “checks for threats before performing the action”.14 Defences lower the risk; they do not remove it. What is prompt injection? explains more.

TAKEAWAY

Be most careful with documents and emails from outside. Never let the AI send, share or delete something because a document told it to.

A six-point check before you use one on client work

The plain point: six questions, most answered once. Ask IT or your manager where you cannot answer yourself.

  1. Is it the approved work version? Use the edition your employer manages, signed in with your work account, never a personal one. Shadow AI explains why the account matters.
  2. What can it reach? Before switching it on for a team, check for folders, sites and channels shared with “everyone” or “anyone with the link”, and narrow them.
  3. Does the client contract allow it? Some contracts and confidentiality agreements limit which outside services may handle client material. Check before AI touches it.Reasoning, 13 Oct 2026 — this site’s advice; contracts differ, and the vendor’s privacy page does not change what you agreed with a client.
  4. What happens to prompts and recordings? Find the vendor’s privacy page (linked above) and ask IT how long your organisation keeps AI chats and meeting transcripts.
  5. Who checks the output? A person reads every summary, draft or answer before it goes to a client, especially names, figures, dates and decisions. Checking AI work sets out how closely.
  6. Does everyone know? Say when AI is taking notes in a meeting, and treat files and emails from outside as untrusted input.

Questions people ask

Does Copilot train on my company’s data?

Microsoft says prompts, responses and the data Copilot reaches in your organisation are not used to train the foundation models behind it (source above). That is a vendor statement about the work product; your company’s contract is what binds it.

Can my employer see what I ask the AI?

Often, yes. Microsoft says admins can view stored Copilot prompts and responses with Content search or Microsoft Purview, and Google lets admins set how long Gemini conversations are kept (sources above). Write prompts as you would a work email.

Can I turn these AI features off?

Usually your admin decides. Google turns the Gemini app on by default but lets admins disable it, Slack owners and admins choose which AI features members get, and Zoom admins can switch features off for the account or a group (sources above).

Where to go next

QUICK CHECK · 3 QUESTIONS

Pick an answer to see if you got it. Nothing is recorded.

1. What can the AI built into a work app, such as Copilot or Slack AI, see?

  • Only the document you have open
  • Everything your account can already open
  • Every file in the company, whatever the permissions
Show the answer

Answer: Everything your account can already open. The page says each vendor’s AI shows only what you already have permission to see, which is everything your account can reach.

2. Why might a work AI assistant show you a file you were never meant to see?

  • The file was shared more widely than anyone intended
  • The AI ignores permissions
  • The vendor trains its AI on your files
Show the answer

Answer: The file was shared more widely than anyone intended. The page calls this oversharing: the AI respects permissions, but old permissions are often wider than meant, and the AI finds such files fast.

3. What is prompt injection in a shared document?

  • A virus hidden in the file’s attachments
  • A summary that leaves out a decision
  • Hidden text the AI may treat as instructions
Show the answer

Answer: Hidden text the AI may treat as instructions. The page explains that the AI reads documents as text and can mistake text inside them for instructions.

SOURCES
  1. 1Microsoft Learn, Data, Privacy, and Security for Microsoft Copilot, updated 8 Oct 2026, read at source 13 Oct 2026“Microsoft 365 Copilot is now named Microsoft Copilot”; “There are no changes to security, compliance, and privacy for organizations.”; “Content in Microsoft Graph, such as emails, chats, and documents that you have permission to access.”; “It can generate responses anchored in your organizational data, such as user documents, emails, calendar, chats, meetings, and contacts.”; “Prompts, responses, and data accessed through Microsoft Graph aren’t used to train foundation LLMs, including those used by Microsoft Copilot.”; “To view and manage this stored data, admins can use Content search or Microsoft Purview. Admins can also use Microsoft Purview to set retention policies”; “Admins have full control to select which agents are allowed in their organization.”; “When you’re using web search, Microsoft Copilot parses the user’s prompt and identifies terms where web search would improve the quality of the response.”; “turning off optional connected experiences could affect the availability of web search.” ↑
  2. 2Google Workspace Help, Generative AI in Google Workspace Privacy Hub, page shows “Last updated: October 6, 2026”, read at source 27 Oct 2026. The address this page used on 13 Oct, support.google.com/a/answer/15706919, answered with a 301 (moved permanently) redirect to this one on 27 Oct 2026: “Access Workspace content that you don’t have permission to access” (listed under “Gemini DOES NOT”); “Your content is not human reviewed or otherwise used for Generative AI model training outside your domain without permission.”; “Access to the Gemini app as a core service is turned on by default. Admins manage access to the Gemini app and can choose to disable the service.”; Gemini in Workspace prompts and responses kept “90 days to indefinite, as determined by admins”↑
  3. 3Slack Help Center, Security for AI features in Slack, read at source 13 Oct 2026“Slack’s AI features only use Slack data that members have access to at the time of request and won’t display or use data from private channels or direct messages (DMs) they aren’t a part of.”; “Customer data never leaves Slack-controlled infrastructure and is never used to train large language models (LLMs).”; “Owners and admins can decide which AI features members of their workspace or Enterprise organization can use.” ↑a“AI searches will never surface any results that Slack’s regular search would not. Similarly, summaries and Slackbot responses will never contain content that you could not otherwise see while reading channels or DMs.” ↑b“Select a citation to jump to the original message and review it for more details or verification.” ↑c
  4. 4Slack, Privacy Principles: Search, Learning and Artificial Intelligence, read at source 13 Oct 2026“If you want to exclude your Customer Data from helping train Slack global models, you can opt out.”; “Slack will not use Customer Data to train generative AI models unless Customer provides affirmative opt-in consent.”; AI search “pulls from the following sources: Slack features (canvases, huddles canvas notes, clip transcripts and text snippets); files uploaded to Slack (PDFs, emails, docx, pptx, and Keynote); linked documents from Google Drive (Docs, Slides)” ↑
  5. 5Zoom, Zoom AI Whitepaper: Data Governance and Privacy, read at source 13 Oct 2026“Zoom does not use any customer audio, video, chat, screen sharing, attachments, or other communications-like customer content (such as poll results, whiteboard, and reactions) to train Zoom’s or its third-party artificial intelligence models.”; “When interacting with external participants, content (e.g., chats, meeting transcripts) exchanged between accounts may be processed according to the external account settings.” ↑
  6. 6Zoom, Zoom AI Whitepaper: Administrative Controls, read at source 13 Oct 2026“At the account level, administrators may enable or disable features for the entire account. At the group level, administrators can control which groups receive specific Zoom AI features”; “users may also be able to enable or disable that feature for themselves at the individual user level” ↑
  7. 7Notion Help Center, Notion AI security & privacy practices, read at source 13 Oct 2026“Yes, Notion AI honors existing permissions. The LLMs and AI Models used to generate AI responses for a user cannot see or use any information to which that user does not already have access.”; “By default, Notion and its AI Subprocessors do not use Customer Data to train any models.”; “there are certain AI-powered features that may require the use of data-retaining LLMs; in such cases, Notion will make available Workspace settings to enable administrators to turn on data-retaining LLMs. Otherwise, such data-retaining LLMs will remain off in your Workspace by default.” ↑
  8. 8Google Workspace Help, Generative AI in Google Workspace Privacy Hub, page shows “Last updated: October 6, 2026”, read at source 27 Oct 2026. The address this page used on 13 Oct, support.google.com/a/answer/15706919, answered with a 301 (moved permanently) redirect to this one on 27 Oct 2026: “For clarity this Privacy Hub only applies when you are using Gemini and Gemini Notebook with a qualifying edition of Google Workspace”↑
  9. 9Microsoft Learn, Data, Privacy, and Security for Microsoft Copilot, read at source 13 Oct 2026“Microsoft Copilot only surfaces organizational data to which individual users have at least view permissions. It’s important that you’re using the permission models available in Microsoft 365 services, such as SharePoint, to help ensure the right users or groups have the right access to the right content within your organization.” ↑a“The responses that generative AI produces aren’t guaranteed to be 100% factual. While we continue to improve responses, users should still use their judgment when reviewing the output before sending them to others.” ↑b“Microsoft Copilot helps mitigate these attacks by using proprietary techniques, such as jailbreak and cross-prompt injection attack (XPIA) classifiers.”; “These classifiers may not be available in all Microsoft Copilot scenarios.” ↑c
  10. 10Microsoft Learn, Secure & Governed Data Foundation for Microsoft Copilot, read at source 13 Oct 2026“This deployment blueprint outlines the essential steps for establishing a secure and governed foundation for Copilot by remediating oversharing, implementing reliable guardrails, and fulfilling AI-related regulatory obligations” ↑
  11. 11Zoom, Zoom AI Whitepaper: User Transparency and Notice, read at source 13 Oct 2026“When the in-meeting Zoom AI features Meeting questions and Meeting summary are activated for the meeting, attendees will see an illuminated AI Sparkle icon in the top-right corner of the meeting window.”; “participants can request the meeting host to disable these Zoom AI features for the meeting, along with the option to delete assets generated by Zoom AI up to that point.” ↑
  12. 12OWASP Gen AI Security Project, LLM01: Prompt Injection, read at source 13 Oct 2026“Indirect prompt injections occur when an LLM accepts input from external sources, such as websites or files. The content may have in the external content data that when interpreted by the model, alters the behavior of the model in unintended or unexpected ways.” ↑
  13. 13National Cyber Security Centre (UK), Prompt injection is not SQL injection (it may be worse), 8 Dec 2025, read at source 13 Oct 2026“Current large language models (LLMs) simply do not enforce a security boundary between instructions and data inside a prompt.”; “prompt injection attacks will remain a residual risk, and cannot be fully mitigated with a product or appliance etc.” ↑
  14. 14Google Workspace Help, Generative AI in Google Workspace Privacy Hub, page shows “Last updated: October 6, 2026”, read at source 27 Oct 2026. The address this page used on 13 Oct, support.google.com/a/answer/15706919, answered with a 301 (moved permanently) redirect to this one on 27 Oct 2026: “When users prompt Gemini to perform an action in Gemini in Workspace or the Gemini app, Gemini checks for threats before performing the action. If Gemini identifies malicious content or detects suspicious activity, it notifies the user and excludes the content, or blocks the action.”↑

Read at source 13 Oct 2026: Microsoft Learn, Data, Privacy, and Security for Microsoft Copilot and oversharing blueprint; Google Workspace Help, Generative AI in Google Workspace Privacy Hub (re-read 27 Oct 2026; the older support.google.com address redirects to it); Slack, Security for AI features and Privacy Principles; Zoom AI Whitepaper (data governance, admin controls, notice); Notion, Notion AI security & privacy practices; OWASP, LLM01: Prompt Injection; National Cyber Security Centre, Prompt injection is not SQL injection.

Feature names, plans and settings are each vendor’s own and change often; some features need a particular plan or an admin to switch them on. Check the vendor’s page. No vendor paid to appear, and the order is not a ranking. The office-keys comparison, the salary-file example and the six-point check are this site’s own, not claims from these sources.

Free printable AI cheat sheets, and AI news every week.

Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.

SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK
aipromptgeneer://$ about this siteAI PROMPT GENEERAI explained in plain English. Every claim links to its source, with the date it was checked.NEWSLETTERWhat changed in AI each week, in plain English.SUBSCRIBE FREE ↗AI PROMPT GENEER · MACHINE ASSISTANCE DISCLOSEDEVERY ARTICLE CARRIES ITS OWN CHECKED DATE