ChatGPT Dots for Business: An Implementation Playbook for Always-On Agents

Most teams use ChatGPT as a chat box. You ask a question, it answers, the conversation ends. ChatGPT dots are built differently. OpenAI announced them on September 29, 2026, and they represent a shift in how AI fits into business operations. A dot is an always-on agent that takes responsibility for ongoing work and keeps making progress between your conversations. It has its own cloud computer, connects to your business apps, and learns from your feedback over time.

The name is styled in lowercase by OpenAI: "dots", "your dot", "a dot". The technology is powered by GPT-6 Astra, and the design is purpose-built for delegation, not just dialogue.

The opportunity is real. The risk is also real. Businesses get value, and avoid damage, when they implement a dot like a new hire with a job description, not a toy. This playbook walks through the implementation process: picking a responsibility with a clear result, connecting only what that responsibility needs, writing approval boundaries into instructions and rules, setting a cadence and review routine, and running admin governance. No hype, no invented case studies, no motivational filler. Just the operator practices that let you put a dot to work without losing control of accounts, customers, or data.

What a Dot Actually Is

A dot is an always-on agent inside ChatGPT. It runs on its own cloud computer and browser, so it can keep working when your laptop is off. You give it a responsibility, connect the apps and tools it needs, and it keeps making progress between conversations. When it needs your input or approval, it asks. When it finishes a task, it reports back.

The underlying model is GPT-6 Astra. The agent has access to relevant ChatGPT memory plus its own saved notes about your preferences, decisions, and ongoing work. It learns from your feedback over time. It can research, analyze data, prepare documents, and build software. It can create cloud coding tasks in a Codex cloud environment you have already set up, and can create Work or Codex tasks on a computer you have connected.

Dots connect to over 4,000 apps through the ChatGPT plugins ecosystem, according to OpenAI. Plugin permissions are shared across dots, ChatGPT, ChatGPT Work, and Codex, so what you connect in one place is available in the others. The dot can also use background agents that run in parallel and start new cloud threads as visible conversations.

You talk to your dot in ChatGPT on desktop, web, or the mobile app. You can also message it in Slack. Microsoft Teams is available as an invite-only alpha for workspaces. The same dot works across all channels. Messages stay in their channel, but context carries across. Before sharing information from a private conversation with others, the dot checks that you have allowed it.

Work happens in three ways. First, the dot works between conversations. It tracks progress, decides when to pause and when to wake up to continue. Second, you can schedule recurring tasks. You specify what to do, when to do it including time zone and end date, which changes deserve a notification, and where to deliver the result. The dot should confirm the schedule. Third, the dot can monitor events when a connected service supports it, such as watching a Slack channel for bug reports. Connecting a source alone does not create monitoring. You have to tell it which channels or events to watch.

You open your dot's profile in the desktop app to review activity under In progress, Scheduled, and Completed. You can open its computer from the profile, select Take over, and later select Return control. A completed run does not by itself confirm the result was achieved. Review the output.

Website sign-ins use a private sign-in form that sends credentials to the browser without exposing them to the model. The cloud browser has its own sessions, separate from your computer. Some sites block cloud browsers. Local computer access is optional and starts off. You can connect one personal computer through the ChatGPT desktop app. The computer must be online with the app open.

Dots are rolling out gradually starting September 29, 2026. Availability is for Pro users outside the EEA, UK, and Switzerland, Business Premium workspaces in all supported regions, and Enterprise, Edu, and Healthcare workspaces as a beta that is off until a workspace admin turns it on. Users must be 18 or older. You create your dot in the ChatGPT desktop app or ChatGPT in a desktop browser, not on mobile or mobile web. Once created, you can talk to it in the mobile app as well.

Your first dot is included in Pro or Business Premium at no extra cost. The plan includes an allowance for deeper work with extended limits in the first month after launch. Conversations with your dot do not count toward ChatGPT usage limits, but Codex or ChatGPT Work tasks your dot starts count toward those limits as usual. OpenAI says it will share usage terms later. Confirm current availability for your plan and region, and check current terms before rollout.

For a full introduction, see OpenAI's guide to dots.

Start with One Responsibility and a Real Brief

The first implementation mistake is vague delegation. A dot is not a general assistant. It is a specialist. Pick one responsibility with a clear result and a reviewable output. The responsibility should be narrow enough that you can write down what success looks like, which apps it needs, and where the approval boundaries are.

OpenAI's own documented patterns are a good starting point. Keep a sales proposal current as requirements change. Prepare content from each interview for review. Keep launch materials aligned with product changes. Update an analysis as new data arrives. Follow up on software feedback and prepare fixes for review. Review your calendar each morning. Track decisions, deadlines, and open items for a project.

The common thread is read-and-prepare work that ends in a draft or report for review. Start there. Do not start with customer-facing messages or actions that change shared data until you have run a few cycles and tightened the brief.

A responsibility brief is the job description. It should include the responsibility, the result you want, source material to share, which apps it may use, what to keep current, which changes deserve your attention, where to send routine updates versus decisions, cadence with time zone and end date, approval boundary, and what it must never do. This is based on OpenAI's guidance: describe the responsibility, share sources, review the first result and correct it before expanding.

Here is an example brief for a small service business:

Responsibility: Keep the service delivery checklist current as we complete client projects and learn what works.

Result: An updated checklist document in Google Drive that reflects lessons from recent projects, ready for me to review each Friday.

Source material: The current checklist is in Google Drive under "Operations / Service Delivery Checklist". Project notes are in the "Client Projects" folder, organized by client name. Slack channel "team-delivery" has real-time updates and issues.

Apps you may use: Use Google Drive only for the Operations folder and the Client Projects folder, and read Slack only in the team-delivery channel.

What to keep current: When a project closes, review the project notes and the Slack thread for that client. If there are steps we added, mistakes we caught, or feedback we received, draft an update to the checklist. If a step is mentioned multiple times as unclear or skipped, flag it for revision.

Which changes deserve my attention: Flag any step that was skipped on more than one project, or any client feedback that suggests we are missing something important. Routine additions and clarifications can go in the weekly update.

Where to send updates: Post a summary in Slack DM every Friday at 4:00 PM Pacific. Include what you reviewed, what you changed, and what you flagged. Do not post in the team channel.

Cadence: Review completed projects every Friday at 3:00 PM Pacific. End date is December 31, 2026. Confirm this schedule.

Approval boundary: Draft updates to the checklist document, but do not publish or share the updated version. I will review and publish it myself. Do not send messages to clients or the team channel. Do not delete any files.

What you must never do: Do not send any message to a client. Do not delete or move any file in Google Drive. Do not post in any Slack channel other than our DM. Do not change any document in the Client Projects folder.

This brief is specific. It names the folders, the channel, the result, the boundary. The dot can do useful work, and you can review the work before it affects anyone else. After you give the brief, ask the dot to confirm the schedule and the approval boundary. Review the first result. Correct anything that missed the mark. Then let it run for a few cycles before you expand the scope or add another responsibility.

Connect Only What the Responsibility Needs

There are three kinds of connections, and they serve different purposes. First, messaging channels let you talk to the dot and let it send you updates. Second, connected apps or plugins let it read data, write documents, send messages, or take actions in those services. Third, local computer access lets it work on your personal computer when the ChatGPT desktop app is open.

Plugin permissions are shared across dots, ChatGPT, ChatGPT Work, and Codex. If you have already connected your Google Drive or Slack in ChatGPT, those permissions are available to your dot. Audit the Plugins tab in ChatGPT before you start. Disconnect anything the dot does not need for its responsibility. The principle is least privilege. Connect only what the responsibility requires.

Local computer access is optional and off by default. Connect it only when the responsibility requires access to files or software on your personal machine. The computer must be online with the ChatGPT desktop app open for the dot to use it. Most responsibilities do not need this. Cloud apps and the dot's own cloud computer are enough.

When the dot signs into a website, use the private sign-in flow. It sends credentials to the browser without exposing them to the model. The cloud browser has its own sessions, separate from your devices. Never paste passwords into chat or documents. Some sites block cloud browsers. If a site does not work, that is a signal to use a different integration or escalate the task to you.

Disconnecting an app stops new access, but it does not delete what the dot already learned from it. The dot retains context for as long as you keep it. You currently cannot view, edit, or delete individual dot memories. Deleting the dot deletes its context. If you connect a sensitive source, understand that the dot can review connected information and form memories from it even without a specific question. This is called proactive research.

Proactive research happens when the dot is not actively working with you. It reads connected apps with read-only tools. Those tools cannot send messages, change app content, or control a browser or computer. The dot keeps private notes. Any follow-up action still requires normal permissions and approvals. This matters when you connect sensitive sources. The dot can learn from them in the background. If that is not acceptable for a particular data source, do not connect it.

When you revoke access to an app, the dot stops reading new data and cannot take new actions. But it does not forget what it already learned, and it does not undo changes it already made. Revoking access is not the same as auditing what happened. Use the Activity view and your app's own logs to review what the dot did before you revoked access.

The Control Layer: Instructions, Custom Rules, and Auto-Review

Dots ship with built-in safeguards, existing app permissions, and automatic action review. Auto-review checks actions that could affect accounts or share information against your instructions, Custom Rules, and safety requirements. It decides whether to proceed, ask for approval, or hand the step to you. This layer is active from the start. You build on top of it with instructions and Custom Rules.

Custom Rules let you set four behaviors for different types of actions. Take action without asking. Take action if pre-approved. Ask before taking action. Hand off to you. You choose the behavior that matches the risk and consequence of the action. OpenAI's Help Center calls the second option "Take action if pre-approved" while the ChatGPT guide labels it "Take action when you say so", and pre-approved means you explicitly requested the action.

The most sensitive actions, like changing a password or transferring money, require you to take over so you can complete them yourself. Permanently deleting data or installing software may require approval each time. In some cases, such as sending recurring messages, you can give approval in advance instead of each time. Approving one message is not ongoing permission. Drafting a message does not authorize sending it. When you approve a future message in advance, be specific about who it goes to, what it says, and when or under what conditions it should be sent. Any advance approval stays limited to what you authorized.

Purchases with cards saved on merchant sites require approval. Advance approval only counts when it specifically covers the purchase.

Custom Rules are instructions the dot tries to follow, and it can make mistakes. They cannot override built-in safety requirements. Write scoped instructions that specify who, what, and when. Do not write vague rules like "be careful with customer data." Write specific rules like "Ask before sending any message to a customer" or "Hand off to me before deleting any file in the shared project folder."

Here is a starter rule set for a dot that manages project documentation and sends internal updates:

• Ask before taking action: Sending any message to a customer or external partner.
• Ask before taking action: Posting in any Slack channel other than our DM or the "project-updates" channel.
• Hand off to you: Deleting any file in a shared folder.
• Hand off to you: Changing permissions on any document or folder.
• Take action if pre-approved: Posting a weekly summary in the "project-updates" channel every Friday at 5:00 PM, as long as the summary follows the template we reviewed.

In addition to Custom Rules, write a "never" list in the dot's instructions. For example: "Never send a message to a client without my approval. Never delete any file. Never change any budget number in the financial model. Never share any document outside the team folder." The "never" list is part of the brief. Neither instructions nor Custom Rules override built-in safety requirements, and the dot can still make mistakes following them.

Prompt injection protections are built in. The dot is designed to distinguish your instructions from content it encounters in documents, emails, or web pages. The protections reduce risk but do not eliminate it. If OpenAI's safety monitoring detects a safety concern, it can pause or stop the dot's work. Review consequential work. Dots can still make mistakes.

When the dot asks for approval, read what it plans to do. Check the recipient, the content, the timing. If it is wrong, correct it. If it is right, approve it. If you are not sure, open the dot's computer from its profile and select Take over, finish the task in the browser or app, then select Return control.

Cadence, Channels, and Updates

Recurring tasks need a saved schedule. Specify what to do, when to do it including time zone and end date, which changes deserve a notification, and where results go. Ask the dot to confirm the schedule. If the dot does not confirm, it may not have saved the schedule correctly. Check the Scheduled view in the desktop app.

Event monitoring requires a supported service and an explicit instruction. Connecting a Slack channel or a project management tool does not start monitoring. You have to tell the dot which channels or events to watch and what to do when it sees them. For example, "Watch the "bug-reports" channel in Slack. When someone posts a new bug, read the description, check if we have seen it before in the project tracker, and draft a reply with the status or next steps. Send me the draft in our DM. Do not post the reply in the channel until I approve it."

In Slack, you can DM the dot or mention it in a channel you have added it to. By default, the dot responds only to its owner. Only the owner can direct it. Adding it to a channel does not start monitoring. You have to tell it to watch that channel. Context carries across channels, but sharing private details in a team channel still requires permission. The dot will check before it shares something from a private conversation in a public channel.

Route routine updates and decisions to different places. Routine updates can go to a Slack DM, a specific channel, or a summary document. Decisions, exceptions, and anything that needs your judgment should come to you directly. Do not let routine updates drown out the signals that need your attention. In the responsibility brief, write where each type of update goes.

For example, a dot that tracks project deadlines might post a daily summary of on-track items in a "project-updates" channel, but send you a DM immediately when a deadline is at risk. A dot that prepares interview content might save drafts in a folder for you to review on your own schedule, but send you a message if it cannot find a key file or if an interview recording is missing.

Time zones matter. When you set a schedule, include the time zone. "Every Friday at 4:00 PM" is ambiguous if your team spans multiple zones. "Every Friday at 4:00 PM Pacific" is clear. The dot should confirm the schedule in your time zone.

An end date keeps a schedule from outliving the project it was built for. Set an end date that matches the project, the quarter, or the business cycle. Review the schedule before the end date and decide whether to extend it, change it, or stop it.

Review the Work Like a Manager

The Activity view shows in-progress, scheduled, and completed tasks. Open it in the desktop app. This is your primary review tool. A completed run does not by itself confirm the result was achieved. Open the task. Read what the dot did. Check the output.

When you need to see exactly what the dot is doing, open the dot's computer from its profile and select Take over. You see the browser, the apps, the files. You can complete the task yourself or make a correction, then select Return control. Taking over is not a failure. It is part of the workflow. Some tasks require human judgment. Some tasks hit an edge case the dot was not trained on. Take over, finish it, and update the instructions so the dot handles it better next time.

Use an evidence standard. Ask the dot for sources, facts versus assumptions, and numbers checked. If the dot says it updated a document, open the document. If it says it sent a message, check the sent folder or the channel. If it says it analyzed data, ask how it handled missing values, outliers, or conflicting sources. If it cannot answer, the work is not done.

Review consequential work before you use it. Consequential means it affects a customer, a financial decision, a compliance obligation, or a shared system. A draft proposal is not consequential until you send it. A budget analysis is not consequential until you make a decision based on it. A code change is not consequential until you deploy it. The review happens before the consequence.

Check the Scheduled view regularly. This shows upcoming and recurring tasks. If a task is running more often than you expected, or if the dot added a task you do not recognize, investigate. If you see a recurring task you do not recognize, ask the dot to list its scheduled tasks and explain where each came from, then change or cancel what you did not intend.

Prompt injection protections reduce but do not eliminate risk. If the dot reads a document or email that contains instructions, it is designed to distinguish those from your instructions. But mistakes are possible. If the dot takes an action you did not expect, check what it read right before the action. Look for conflicting instructions in the content. Update your "never" list and Custom Rules to prevent it next time.

Stopping, Pausing, and Offboarding

Pause stops the current main task. It does not stop delegated tasks or cancel scheduled runs. If the dot started a background task or a cloud thread, pausing the main task does not stop the background work. Go to Activity and stop delegated tasks separately. Go to Scheduled and cancel recurring tasks separately. Pause is not a global stop button.

Stopping a task does not undo completed actions. If the dot sent a message, posted a document, or changed a record, stopping the task does not unsend the message or revert the change. Whether a completed action can be reversed depends on the action and the app. You can ask the dot to help correct a mistake, and OpenAI says it may be able to reverse unintended edits to a document or recall an email, but some actions cannot be undone.

Deleting the dot deletes the dot's own context, conversations, saved memories, and scheduled tasks. Files, Codex threads, and ChatGPT conversations the dot created are stored separately and are not deleted. Memories in ChatGPT memory are managed separately. Deletion does not undo changes in connected apps or recall delivered messages. It cannot be undone.

Before you delete a dot, review what it has access to. Go to the Plugins tab and see which apps are connected. Go to Activity and see which tasks are in progress or scheduled. Go to Scheduled and cancel recurring tasks. Plugin connections live in ChatGPT's Plugins tab and are shared across dots, ChatGPT, ChatGPT Work, and Codex, so review and disconnect them there as a separate step.

If you are offboarding a dot because the responsibility is ending, document what it did. Export key outputs. Save the final version of documents it maintained. Update your runbook or process documentation with what you learned. The dot's memory is not a substitute for institutional knowledge. Write down what worked and what did not.

If you are offboarding a dot because it made a mistake, investigate before you delete it. What instruction did it follow? What did it misunderstand? What content did it read that might have confused it? Update your Custom Rules and instructions so the next dot does not repeat the mistake. Deleting the dot does not fix the process.

Workspace Admin Setup and Governance

Dots are available to Business Premium workspaces, and in Enterprise workspaces (including Edu and Healthcare) dots are a beta that is off by default until a workspace admin enables it. OpenAI's dots admin guide covers configuration for Enterprise workspaces. Business Premium admins should check which of these controls appear in their workspace settings.

Workspace owners manage these through workspace defaults and custom roles assigned to groups, under Workspace settings, Permissions & roles. Permissions and roles include Use dots, Add dots to Slack, Allow local computer access, and Use custom rules for dots. Members still complete their own Slack and local computer setup, while supported existing app connections may already be available.

Cloud computer capabilities include Cloud browser use, Cloud network access, and Cloud computer use. There is also a Use password manager setting. OpenAI advises reviewing how cloud computer controls and Use password manager apply across both dots and Work Cloud before changing them. Enable these capabilities only where your policies allow browsing websites, accessing online services, and interacting with apps.

Dots can use supported existing ChatGPT app connections, and Plugin controls are where you set app access and action restrictions. Lock it down to the apps your team actually needs.

For Slack, the ChatGPT app must be installed in your Slack workspace by the Slack owner or app manager. After the app is installed, each member can connect their dot to Slack. The admin permission "Add dots to Slack" controls whether users can make that connection.

OpenAI's admin guide says removing dots access does not replace disconnecting an app or signing out of a website, so review app accounts and website sessions separately.

Enterprise model controls and defaults do not apply to dots, so review dots access and capability permissions directly in workspace settings.

Use the Analytics API for the adoption metrics OpenAI makes available, and review current usage terms before rollout.

Supported Compliance API records can be used to investigate user messages and dots' replies. Confirm record coverage before relying on it for an audit.

If any cloud policy has enforce_residency enabled, local computer access is unavailable for dots. This is a hard constraint. If your data residency policy requires it, users cannot connect their personal computers to their dots.

Data training defaults: Business, Enterprise, and Edu workspace content is not used to train models by default. Personal plans are controlled by the "Improve the model for everyone" setting. OpenAI does not train directly on proactive research or the dot's private notes. If your compliance framework requires confirmation of training exclusions, verify the current terms with OpenAI before rollout.

Run a permissions audit before you enable dots. List every user who will have access. List every plugin that is connected. List every Custom Rule you plan to deploy. List every compliance or logging requirement you need to meet. Then configure the workspace to match. Do not enable dots and figure it out later. The risk is too high.

Set a weekly admin review cadence. Review who has Use dots access through roles, review Plugin controls and allowed actions, review Slack participation and local computer access settings, check available adoption metrics in the Analytics API, and confirm Compliance API coverage for your audit needs. Members review their own Activity and Scheduled. Governance is not a one-time setup. It is an ongoing practice.

Where a Dot Fits in Your Stack

A dot works for its owner. It is a personal agent. It has access to the apps and data its owner connects. It follows the instructions its owner writes. It reports to its owner. This is the right design for internal responsibilities: keeping a proposal current, preparing content for review, tracking project decisions, updating an analysis.

Customer-facing automation is a different job. Answering leads, booking appointments, triaging support requests, and handing off to humans requires a different architecture. That job belongs in your CRM, your customer-facing agent stack, or your support platform. It needs its own identity, its own escalation rules, and its own compliance controls. It should not run on a personal agent that has access to one employee's connected apps.

If you are building customer-facing automation, design for AI agent handoffs from the start. The agent needs to know when to escalate, where to route the escalation, and what context to pass to the human. A dot is not built for that. A dot is built to take responsibility for work its owner would otherwise do manually.

The line is clear. Internal work that ends in a draft, a report, or a decision for the owner: dot. Customer-facing work that represents the business and requires escalation design: customer-facing agent. Do not blur the line. Do not use a dot to answer customer emails or book sales calls unless you have designed the approval and escalation workflow, tested it, and documented it. Even then, a purpose-built customer-facing agent is the better choice.

Common Mistakes

Connecting every plugin on day one. Plugin permissions are shared across dots, ChatGPT, ChatGPT Work, and Codex. If you connect everything, the dot has access to everything. Start with the minimum set of apps the responsibility requires. Add more only when the responsibility expands.

Vague responsibilities with no result or review point. "Help me stay organized" is not a responsibility. "Keep the project tracker updated with decisions from the weekly meeting and send me a summary every Friday" is a responsibility. If you cannot write down what success looks like, the dot cannot deliver it.

Assuming a draft request lets it send. Asking the dot to draft a message does not authorize sending it. Drafting and sending are separate actions. Approve sending separately, and if you approve in advance, say who it goes to, what it says, and when it should be sent.

Assuming Pause stops everything. Pause stops the current main task. It does not stop delegated tasks or cancel scheduled runs. If you need to stop all work, go to Activity and Scheduled and stop each task separately.

Adding it to Slack and expecting it to monitor. Adding the dot to a Slack channel does not start monitoring. You have to tell it which channels to watch and what to do when it sees something. Connecting a source is not the same as instructing the dot to act on it.

Pasting passwords into chat. Use the private sign-in flow. Never paste passwords into chat or documents. The private sign-in form sends credentials to the browser without exposing them to the model.

Using outputs without checking numbers and sources. A completed run does not confirm the result. Ask for sources. Check the numbers. Open the document. If the dot says it analyzed data, ask how it handled edge cases. If it cannot answer, the work is not done.

Forgetting that memory cannot be pruned item by item. You currently cannot view, edit, or delete individual dot memories. If the dot learns something you do not want it to remember, you have to delete the dot and start over. Disconnecting an app does not delete what the dot already learned from it.

Skipping the admin permission review. If you are in an Enterprise workspace (and Business Premium workspaces where these controls apply), the admin controls permissions, cloud capabilities, and plugin access. Do not assume the defaults are correct. Audit the settings before you enable dots for your team.

A 30-Day Rollout Plan

This is a recommended approach for putting a dot to work in a business environment without losing control. It is not a promise of results. It is a process that builds trust in stages and gives you time to tighten the brief, the rules, and the review routine before you expand.

Week 1: One owner, one responsibility, minimal plugins, local computer off, drafts only.

• Pick one person to own the first dot. This should be someone who understands the responsibility, can write a clear brief, and has time to review the work daily.
• Pick one responsibility that ends in a draft or report for review. Use the patterns from OpenAI's documentation: keep a document current, prepare content from source material, track decisions and deadlines.
• Write the responsibility brief. Include the result, the source material, the apps it may use, the approval boundary, and the "never" list.
• Connect only the plugins the responsibility requires. Audit the Plugins tab first. Disconnect anything the dot does not need.
• Leave local computer access off unless the responsibility requires it.
• Set Custom Rules to "Ask before taking action" for any message or change. Set "Hand off to you" for deleting files or changing permissions.
• Run the first cycle. Review the result. Correct anything that missed the mark. Do not expand the scope yet.

Week 2: Tighten the brief and evidence standard, add one recurring scheduled check with a confirmed schedule.

• Review the work from week 1. What did the dot do well? What did it misunderstand? What instruction was ambiguous?
• Rewrite the brief to fix the ambiguities. Add examples if the dot missed the format or tone you wanted.
• Add one recurring scheduled task. Specify what, when including time zone and end date, which changes deserve a notification, and where results go. Ask the dot to confirm the schedule. Check the Scheduled view to verify it saved correctly.
• Set an evidence standard. When the dot completes a task, ask for sources, facts versus assumptions, and numbers checked. Open the output and verify it.
• Do not add new plugins or expand the scope yet. Focus on making the current responsibility reliable.

Week 3: Add a narrow pre-approved action via Custom Rules or a scoped instruction, keep customer-facing messages behind approval.

• Review the work from week 2. If the dot is reliably delivering drafts or reports that need minimal correction, add one narrow pre-approved action.
• Example: "Take action if pre-approved: Posting a weekly summary in the "project-updates" Slack channel every Friday at 5:00 PM Pacific, as long as the summary follows the template we reviewed."
• Write the pre-approval in Custom Rules or as a scoped instruction in the brief. Be specific about what is pre-approved and what still requires approval.
• Keep customer-facing messages behind approval. Do not pre-approve sending to customers, external partners, or public channels.
• Monitor the first few pre-approved actions closely. If the dot gets it wrong, revoke the pre-approval and go back to "Ask before taking action."

Week 4: Weekly review of Activity, Scheduled, Plugins, Custom Rules, and admin settings, then decide what earns broader trust and whether to expand to more team members.

• Run a weekly review. Open Activity and check in-progress, scheduled, and completed tasks. Open Scheduled and verify recurring tasks are running as expected. Open the Plugins tab and verify only the necessary apps are connected. Review Custom Rules and verify they match the current responsibility.
• If you are in a workspace, review admin settings. Review who has Use dots access through roles, review Plugin controls and allowed actions, review Slack participation and local computer access settings, check available adoption metrics in the Analytics API, and confirm Compliance API coverage for your audit needs. Members review their own Activity and Scheduled.
• Decide what earns broader trust. If the dot has reliably delivered the same type of output for two weeks with minimal correction, consider pre-approving that action or expanding the scope slightly.
• Decide whether to expand to more team members. If the first dot is working well, pick a second owner and a second responsibility. Use the same process: one responsibility, minimal plugins, drafts only, tighten the brief, add a schedule, add a narrow pre-approval, weekly review.
• Do not skip the review cadence. Weekly review is not optional. It is how you catch mistakes before they compound.

After 30 days, you will have a working dot, a tested brief, a set of Custom Rules that match the responsibility, and a review routine. You will also have a clear sense of what the dot does well and where it still needs your judgment. Use that knowledge to decide whether to expand the dot's scope, give another team member a dot of their own for a different responsibility, or keep the current setup and optimize it further.

Where SumoGrowth Helps

SumoGrowth is an AI implementation and consulting agency. We set up AI systems for businesses. We do not sell hype. We do not promise ROI numbers we cannot source. We help you write responsibility briefs, scope connections, configure Custom Rules and approval policies, set up admin governance, and build review routines. We teach your team how to run the weekly review, how to tighten instructions when the dot misses the mark, and how to decide what earns broader trust. We help you put ChatGPT dots to work on real responsibilities without losing control of accounts, customers, or data.

by Darius Gaynor