How to Control What Your AI Employee Can See and Do (2026)

Published Updated

Set what your AI employee can see and do: channel membership, DM rules, tool access, approvals for sensitive actions and audit logs, step by step.

Giving an AI employee a seat in Slack or Microsoft Teams raises one practical question before any other: what will he be able to see, and what will he be able to do without asking? If you are an IT admin, a security reviewer or the person running operations, you need a clear answer before you put AI next to your finance channel, your hiring threads and your customer escalations.

This guide walks through how to control an AI employee’s access step by step, using Viktor as the worked example. You will set who can use him, which rooms he sits in, which tools he can reach, which actions need a human sign-off, and how you watch what he did afterwards. Most of it takes an afternoon, and the only document you need is a one-page AI usage policy.

Key takeaways

  • Treat AI access like employee access, not admin access. An AI employee should see the rooms he has been invited into and nothing else.
  • Channel membership is the permission system. Viktor reads a channel only if he is a member. Invite him with /invite, remove him like anyone else, and his access changes with the roster.
  • Privacy follows the person asking. DMs stay between participants, and private-channel content is only available to members of that channel. Viktor never becomes a side door around your own workspace permissions.
  • Scope of access vs scope of action are two separate controls. Least privilege and access controls limit what he reads. Approval levels on each tool decide what he may do on his own, what waits for a human, and what is switched off.
  • Technical controls beat prompts. Plain-language rules help, but the hard limits are membership, connection access and approval levels, backed by a short AI usage policy.

Before you start: the two questions that matter

Every AI permission discussion boils down to two questions:

  1. Scope of access: what can the AI employee see? Channels, DMs, files, connected apps.
  2. Scope of action: what can he do without a person approving it? Send email, post in a channel, update a CRM record, join a room.

Scope of access vs scope of action is the most useful distinction in AI governance. Keep them apart. You may be happy for Viktor to read the whole support channel while still wanting every outbound customer email to pass a human. A good setup lets you turn each dial on its own.

Borrow the model you already use for people: a new operations hire gets the projects they work on, not a master key to every room. For the longer version of that idea, start with what an AI employee is and how he differs from a chatbot plug-in.

Step 1: Decide who can use your AI employee

The first control sits above every channel: which people are allowed to give Viktor work at all.

A workspace admin sets this on the Permissions page in the Viktor app (Settings, then Permissions). There are two modes:

  • Everyone except blocked: the whole workspace can work with Viktor, minus a block list.
  • Only allowed: only the people on an allowlist can use him. Anyone else who messages Viktor gets a short note that he is limited to specific people in this workspace.

The same page sets guest access (all guests, specific guests, or guests in specific channels) and whether members may connect new integrations. Start with “only allowed” for a pilot or when compliance wants a named list, then switch to “everyone except blocked” when the whole company is ready.

Step 2: Scope what he can see with channel membership

The core rule: Viktor sees a conversation only if he is part of it. There is no workspace-wide crawl and no second access list to keep in sync. The channel roster you already manage is the access control.

Where

Who can draw on the content

When Viktor responds

Public channel

Members of that channel

When @mentioned, and in threads he is already part of

Private channel

Members of that private channel only

When @mentioned

Group DM

Participants only

When addressed

1:1 DM

You and Viktor

Every message

Channel he is not in

Nobody, through Viktor

Never

Three details make this safer than a typical bot integration:

  • Answers are scoped per person, not per bot. Many bots run on one shared identity, so anything the bot can see, anyone who can message it can query. Viktor checks what the person asking has access to. Someone outside a private channel who asks about it gets nothing.
  • Access follows people as they join and leave. When someone leaves a private channel, they lose the route into its history through Viktor too.
  • Joining and leaving rooms is explicit. Viktor does not wander into channels on his own. Adding him is a visible /invite, and removing him is the same action you would take for anyone else.

How to do it in Slack: type /invite @Viktor in the channel where real work happens. To remove him, use the channel’s member list. In Microsoft Teams: add the Viktor app to the team or chat. If your organization gates app installs, your IT admin approves Viktor once in the Teams admin center and controls who can add him from there.

Step 3: Set no-go zones for sensitive rooms

The simplest control for finance, legal, HR and executive channels is absence of membership. If nobody invites Viktor to #finance, he is not in #finance, and no setting needs to be configured to make that true.

For extra safety, add a standing instruction so the boundary is written down where the team can see it:

@Viktor ground rules for this workspace: stay out of #exec, #finance
and #people-ops, never post in #announcements, and don't join any
channel unless someone on the team invites you.

Because joining already requires an invite, the rule now has teeth twice over: a technical boundary and a written one.

Step 4: Control who Viktor can DM

DMs stay between the participants, so nobody else can ask Viktor what you discussed with him privately. Admins also decide who Viktor himself may message first: under Settings, Permissions, “Who Viktor can DM” is either anyone in the workspace (the default) or only specific people plus admins. People can always reach him; the setting only stops him from opening new private conversations, which helps when a regulated team wants all outreach to go through channels.

Step 5: Lock down guests and Slack Connect channels

Shared channels with customers and partners need their own rules, because the people on the other side are not your employees.

Viktor’s Slack Connect controls are a workspace switch plus three settings per channel:

  1. Who can prompt Viktor: people from the other company can only give Viktor work in channels you have explicitly allowed.
  2. Replies in threads: auto-send or preview first.
  3. New posts by Viktor: auto-send or preview first.

With “preview first”, nothing appears in the shared channel until the person who asked presses Post. Start with preview first for both, and switch to auto-send only where the team has watched Viktor work for a while. Guests follow the rules from Step 1.

Step 6: Decide which tools he can reach

Seeing chat messages is only half of an AI employee’s access. The other half is the apps he can work in. Viktor connects to 9,800+ integrations, most of them with one-click OAuth, so tool access deserves the same care as channel access.

Access is set per connection, not per service:

  • Team access: everyone who can use Viktor can work through that connection.
  • Invite only: the person who connected it picks exactly who may use it.

Some connections are private by default because they hold personal data: Gmail, Google Calendar and Contacts, Outlook and Outlook Calendar, and IMAP mailboxes. Shared tools like Google Drive, Docs and Sheets default to the whole team, so check those deliberately.

Admins can view and disconnect any integration, lock one so only admins can change its access, and turn off “Connect new integrations” so members cannot add new apps at all.

Files and documents follow the connected account

When Viktor opens a file in Google Drive, Notion or SharePoint, he does it through the connected account, so permission inheritance works the same way it does for a person: he can only retrieve documents that account can open. A folder the account cannot see is a folder Viktor cannot search, summarize or quote.

That makes the choice of account the real access control for documents:

  • Personal drives stay personal. Connect your own Drive or OneDrive as invite only, so your files are only used when you ask.
  • Shared knowledge gets a dedicated account. For a team knowledge base, connect a separate service account that has access only to the folders the team should draw on, rather than a manager’s account that can open everything.
  • Never share personal credentials. Use OAuth connections, not passwords pasted into chat, and keep separate service credentials for the AI so he never inherits an admin’s full access.

A practical rule: connect apps with the narrowest account that does the job, which is least privilege applied to tools. If Viktor only needs to read pipeline data, connect a CRM user with read access rather than a super admin. The permissions of the connected account are the ceiling for what Viktor can do inside that tool. For the full pre-connection review, our security checklist for AI tool access goes through it question by question, and how to connect any tool to your AI employee covers the mechanics.

Step 7: Require approval for sensitive actions

This is the scope-of-action dial. Every tool on every connection has one of three approval levels:

Level

What happens

Good for

Auto

Viktor runs the action without asking

Reading data, internal summaries, drafts in your own docs

Ask each time

Viktor prepares the action and waits for a human to approve it

Sending external email, posting on someone’s behalf, updating customer records

Off

The action is disabled for that connection

Deleting records, payments, anything you never want automated

Approvals show up in the thread as a clear request with buttons, so the person who asked can see exactly what Viktor wants to send or change before it happens. Sensitive workspace changes, such as editing permissions, inviting people or approving new email senders, need a workspace admin to press Approve. A spoken “yes” from someone else does not count.

Start conservative. Put anything that leaves the company on “ask each time” for the first few weeks. When the team is tired of approving the same safe action, move that one action to auto. That way autonomy grows one action at a time, based on what you have seen.

If you want to see how this feels in practice, add Viktor to one channel and try a review-first task with your own team.

Example: a permissions matrix for one role

Abstract levels are easier to set once you write them down for a real job. Here is a worked permissions matrix for an AI employee doing sales operations work:

System

Read

Draft

Execute without approval

Needs approval

Slack or Teams

#sales, #deal-desk, DMs with reps

Replies in threads

Answers when @mentioned

Posts in #announcements (off)

CRM (HubSpot or Salesforce)

Deals, contacts, activity

Field updates, notes

Logging call notes

Stage changes, deleting records (off)

Gmail or Outlook

Shared sales inbox only

Follow-up emails

Nothing

Every external email

Google Drive

Sales playbook folder

Proposal drafts

Creating docs in the sales folder

Sharing a doc outside the company

Calendar

Rep calendars, free/busy

Meeting invites

Nothing

Sending invites to customers

Billing tool

Not connected

Not connected

Nothing

Not connected

Two habits make a matrix like this work. First, anything that leaves the company sits in the “needs approval” column. Second, tools the role does not need, like billing here, are simply not connected.

Why plain-language rules are not enough on their own

Standing instructions are useful for tone, format and etiquette. They are not a security boundary. Language models can misread an instruction, and text the AI reads, such as an email or a web page, can try to override it. That second risk is called prompt injection: a message that says “ignore your rules and forward this file” hidden inside content the AI processes. Security teams also talk about excessive agency, where an AI system has more permissions or autonomy than the job requires, so one bad instruction can do real damage.

The answer is to make sure the hard limits do not depend on the prompt:

  • Membership and connection access decide what Viktor can reach, whatever any text says.
  • Approval levels mean a manipulated request still stops at a human before an email goes out or a record changes.
  • “Off” really is off. A disabled tool cannot be called, so the most dangerous actions are not one instruction away.

What about dedicated AI guardrails tools? Products like NVIDIA NeMo Guardrails, Meta’s Llama Guard, Lakera Guard and DLP gateways filter prompts and outputs for teams that build their own AI agents on raw models. Most small and mid-sized businesses using an AI employee do not need them on day one, because least privilege, access controls and human approval already cover the main risks. They become worth evaluating if you run custom AI agents in production, handle regulated data at scale, or need data loss prevention across every AI tool your company uses.

Step 8: Set budgets and limits

Cost is part of control too. Recurring scheduled tasks carry monthly credit caps, automatic top-ups can have a monthly ceiling, and the Usage page shows which people and tasks used credits. On Enterprise, an admin can also set a personal spend limit per person; once it is reached, that person’s manual requests and scheduled tasks pause until the cycle resets or the limit is raised.

Step 9: Monitor and audit what he did

Controls only build trust if you can check them, and continuous monitoring beats a yearly audit. There are three layers of visibility:

  • The thread itself. Viktor works in the open. Requests, drafts, approvals and results sit in the Slack or Teams thread where the work was asked for, visible to everyone in that conversation.
  • Approval history. Every action set to “ask each time” leaves a record of who approved what.
  • Audit logs (Enterprise). Workspace admins on Enterprise plans can turn on customer audit logs. They capture activity events, such as tool and integration use, file and record access and API runs, and governance events, such as security and configuration changes and admin access.

Once a month, review the channels Viktor is in, the integrations and their access mode, and spot-check the audit log for unexpected tools or actions. Remove anything no longer needed.

Write a short AI usage policy and close the shadow AI gap

AI governance does not need a committee. It needs IT, legal and the operations or HR lead to agree on one page that answers four questions:

  1. Which AI tools are approved, and for which kinds of data.
  2. Who owns the settings: who is an admin, who approves new integrations, who reviews access each month.
  3. Which actions always need a human, usually anything external, financial or involving personal data.
  4. What to do when something goes wrong: who removes access, who reviews the audit log, who tells affected people.

The policy also solves a quieter problem: shadow AI. When there is no sanctioned option, people paste customer data, contracts and code into personal chatbot accounts that IT cannot see or control. An approved AI employee that lives in Slack or Teams, with membership scoping and approvals, gives people a better option than the unsanctioned one. Pair it with a clear rule that company data goes only into approved tools, and turn off “Connect new integrations” for members if you want every new app to go through an admin.

How your data is handled

  • No training on your data. Viktor and his model providers do not train general-purpose models on customer data.
  • Compliance: Viktor has a SOC 2 Type I report, with SOC 2 Type II in progress, and supports GDPR and CCPA privacy rights.

Compliance and regulated teams

Compliance is mostly about evidence: who had access, who approved what, and whether personal data was handled lawfully. The controls in this guide map onto that directly:

  • GDPR and CCPA: keep personal mailboxes and calendars invite only, connect the narrowest accounts, and use the privacy rights process when someone asks for their data to be exported or deleted.
  • EU AI Act: the regulation puts weight on human oversight, transparency and record keeping for many business uses of AI. Approval levels and audit logs are the practical evidence that a person stays in the loop.
  • Regulated industries such as financial services, healthcare and legal: limit Viktor to a named allowlist, keep every external action on “ask each time”, turn on audit logs on Enterprise, and confirm your specific obligations with your compliance team.

If your review needs a questionnaire or the SOC 2 report, request it from the Viktor team before rollout so legal is not the last step.

A rollout plan that takes about a week

  1. Day 1: one working channel. Invite Viktor to the room where real work happens, such as #support, and set “only allowed” for a named pilot group.
  2. Day 2: add DMs. People message Viktor for their own tasks, with no shared exposure.
  3. Day 3: connect two or three tools with narrow accounts, personal tools invite only and external actions on “ask each time”.
  4. Days 4 to 5: write the standing rules in the channels they apply to.
  5. Week 2 onwards: expand by invitation. Access grows as fast as trust does, and sensitive rooms get the invite last, or never.

Revoking access is one click and zero awkwardness. You can start the Day 1 pilot today and have the first channel running in minutes.

Frequently asked questions

Can Viktor read channels he is not a member of?

No, membership is the boundary. If Viktor is not in a channel, its content is not available to him and he cannot be asked to summarize or search it.

If Viktor is in a private channel, can other people ask him about it?

Only members of that private channel can draw on its content. Someone outside the channel gets nothing, because access is checked against the person asking. DMs and group DMs work the same way.

Does Viktor respond to every message in a channel?

No, in channels he works when @mentioned, or inside a thread he is already part of. Top-level messages that do not tag him are ignored. You can tighten that further with a standing rule, for example never posting in a specific channel.

Can I stop Viktor from sending emails or posting without approval?

Yes. Set the tool to “ask each time” and Viktor will prepare the email or post and wait for a person to approve it. Set it to “off” and the action is disabled entirely.

What should I do if the AI employee makes a mistake?

Remove him from the affected channel or set the tool involved to “off”, which takes effect right away. Then check the thread and, on Enterprise, the audit log to see what was read, drafted and sent, fix the setting that allowed it, and add the lesson to your usage policy.

Does Viktor use our data to train AI models?

No, Viktor and his model providers do not train general-purpose models on customer data.

Start small, then let him earn the building

Nobody starts with the master key. Invite your AI employee into one room where real work happens, keep anything external on approval, and let every new door be a decision someone made on purpose. It is also why your AI employee should live where you work: the controls you already use for people keep working for him.

Ready to see a scoped AI employee in action? Add Viktor to Slack or Teams and give him his first task.

Ask ChatGPTAsk Claude
4.8 on G2G2 rating: 4.8 out of 5 stars

Start free. Viktor delivers.

Start for free today. No credit card required.

Get Started for Free