The Unanswered Message Problem: What OpenAI’s Dot Changes for People Who Drown in Their Own Inbox

OpenAI's dot is a persistent ChatGPT agent that keeps working between conversations. It reads connected apps like Gmail and Teams, sorts messages into what needs you, prepares draft replies and a daily brief, and returns results for your approval. It addresses missed messages by holding context across channels instead of leaving it fragmented.

1. A confession about my response times

If you have ever sent me an email, a Teams message, a WhatsApp or a LinkedIn note and heard nothing back, you were on the losing side of a choice I made every single day. I could either get something meaningful done, or I could read emails, do admin, work through the message queues and approve things. There were never enough hours for both, and every hour spent clearing an inbox was an hour not spent on what I think a CIO should be doing: architecture, unblocking teams that are stuck, coaching and teaching, and personally taking on the complex problems that nobody else has the time or context to crack.

The week dots launched, for example, I was deep in a problem with long running Claude Code sessions. Claude Code’s own compaction stops the session while it summarises, so I added asynchronous, pauseless context compaction to Claude Burst, my open source local gateway, along with session coordination, automatic handoffs and the ability to keep working with the laptop lid closed. In its first week it compacted roughly two thirds of the context that would otherwise have been resent (write up here). That kind of work needs long, uninterrupted stretches of thinking, and it does not happen in the gaps between approvals.

I chose doing over responding, and I never pretended otherwise. I did not need to announce that I was bad at replying, because people told me, often and quite directly. My standard answer was that if something was important, they should call me. The reply I heard more often than I would like to admit was a slightly weary “I tried to”, which is about as clear a verdict on a personal operating model as you will ever get. It also shaped my meetings in a way I am not proud of. A lot of them started with me apologising for something I had missed or had not done: the email I had not answered, the document I had not reviewed, the approval that had been sitting with me for a week. The first few minutes of a conversation that should have been about the work became a small act of contrition, and the other person had usually spent the days before it waiting on me. I owned it as a failure and still made the same call the next morning, because the alternative was a job made entirely of reading and approving.

To make this concrete, I took screenshots of my counters on the morning I wrote this. Outlook showed 1,869 unread emails. Gmail showed 26,657 unread. WhatsApp had 303 unread chats, 46 of them groups, and one engineering group had picked up ten new messages, including an @ mention of me, before eleven o’clock. Teams had 66 unread chats and 7 notifications. That is the steady state, not a bad week, and nobody reads their way out of those numbers while also doing the job they were hired to do.

The part I never made peace with was being unresponsive. I spent fifteen minutes at a time hunting through Google Drive for documents I had probably written myself. I missed deadlines sitting in plain sight in threads I had read once. Every task tracker I tried failed the same way, because a tracker only works if you feed it, and feeding it is exactly the admin I had chosen to sacrifice. I was never really in control; I had simply picked which part of my job I would let slip. For someone who spent a career building low latency systems where a missed packet is a failure, my personal communication stack had worse reliability than anything I would have allowed into production.

2. What a dot actually is

OpenAI announced dots at DevDay on 29 September 2026. A dot is a named, persistent agent in ChatGPT, powered by GPT-6 Astra, with its own cloud computer and browser (OpenAI docs). The difference from a normal chat is that a dot keeps working between conversations: it tracks progress, follows up as things change, and comes back to you when a decision needs your judgement. It uses ChatGPT’s plugin ecosystem, which OpenAI says covers more than 4,000 apps, including Gmail, Google Drive and GitHub (launch post).

There are three separate kinds of connection, and the distinction matters for everything that follows (getting started). A messaging channel is how you talk to the dot: today that is ChatGPT, voice calls, Slack and Teams, with texting listed as coming soon (messaging docs). A connected app or plugin is what lets it read and act on a service, within that plugin’s permissions. A connected computer gives it local files and apps while the machine is online. OpenAI is explicit that connecting a messaging channel does not grant access to your inbox, other apps or your computer.

When you are not actively working with it, a dot researches your connected sources to suggest next steps, and that research is read only; follow up actions need the applicable permissions.

3. Day one: what I connected and what it found

I set my dot up today, so this is a day one report rather than a verdict. Everything I wanted connected is connected except WhatsApp, which I come back to below.

The first useful thing it did was also the most uncomfortable. I am building a house, and a house build generates a steady stream of email from people who need a decision from me before they can carry on. When I asked it to look back through my existing threads for anything still waiting on me, it surfaced a batch of house related emails I had simply missed, buried under everything else. None were catastrophic yet, but every one of them was someone waiting on me, which is the “I tried to” problem in a different setting.

4. Email: the guilt turns into a queue

The first responsibility I gave it is email triage. I instructed mine to sort what arrives into what genuinely needs me, what can be answered from information I already have, and what is noise, to prepare drafts for the straightforward replies, and to ask before sending anything. That boundary is my instruction rather than a property of every dot. OpenAI’s own review layer decides, action by action, whether a dot can proceed, needs approval or must hand a step over to you, and you can add instructions such as showing you drafts first (OpenAI docs).

The shift I am hoping for is bigger than the time saved. For years my inbox was the thing I had consciously sacrificed, and every unread message represented someone paying the price for that choice. If something else holds the list of who is waiting and for how long, the guilt turns into a queue, and a queue I can clear in a few minutes no longer competes with the work that matters. Whether the drafts get close to how I actually write is something only the coming weeks will show.

5. Teams, WhatsApp and LinkedIn

Teams was the most costly of my channels, because it is where senior leaders tend to reach me. I find Teams genuinely difficult to use: chats, channels, mentions and activity live in slightly different places, and an important message from an executive looks the same as the fortieth reaction in a group chat. Some of the messages I most needed to answer were precisely the ones I was most likely to miss.

Here the connection model matters. Adding Teams as a contact method lets me talk to my dot inside Teams; it does not by itself let the dot read my Teams chats. That depends on a separate Microsoft plugin and its permissions. With that access in place, I can ask the dot what senior leaders are waiting on me for instead of fighting the Teams interface to find out.

WhatsApp is the gap I am working on now. There is no first party WhatsApp integration, so I am building a relay to bring WhatsApp messages somewhere the dot can see them, and I will write up how it works once it is stable. Given that 303 unread chats were sitting there this morning, it is the integration I most want to get right. LinkedIn is in a similar position: the LinkedIn plugins I have seen are on the advertising side rather than the inbox, so anything important there still has to be captured by hand.

6. Documents and commitments

Drive search works if you remember the title and fails if you remember the content. With Drive connected, the question changes from a filename to a description, such as “the document where we compared the two vendor proposals and someone raised the latency concern”, and because the dot also sees the email around it, it can connect a document to the conversation that produced it.

The same reading is what makes task tracking different. My failure mode with trackers was always capture rather than execution. A dot reading the email and message threads it has access to can pull out the commitments I made and the deadlines others gave me, which is exactly what it did with the house emails: several of them contained implicit deadlines, because a contractor waiting on a decision is a schedule slipping by the day. What this gives me is a reviewable record of commitments from the sources it can access, not a guarantee that nothing is ever missed; anything in a channel it cannot see, which today includes WhatsApp, is still outside that record.

7. Guardrails, because judgement still matters

I run technology for a bank, so connecting an always on agent to my email and documents is not a casual decision. The starting controls are reasonable: proactive research is read only, plugin permissions are managed through the existing ChatGPT app controls and are shared across dots, ChatGPT, Work and Codex, and the review layer described above sits in front of actions that affect your accounts. OpenAI also says plainly that a dot can make mistakes and that stopping a task does not undo actions already completed, so important results need checking.

A few sharp edges are worth knowing. Deleting a dot’s own saved memories means deleting the dot, so think before connecting something you might later want it to forget. Because permissions are shared across OpenAI’s products, review connections you made months ago for other purposes. Every bridge into an agent, including the WhatsApp relay I am building, is also a path for data to leave the place it lives. And everything here is my personal workflow; anything involving regulated data, client information or production systems belongs inside an organisation’s governance process, not an individual’s productivity experiment.

8. Where this goes next

Dots are rolling out to Pro users over 18 outside the European Economic Area, the UK and Switzerland, to Business Premium worldwide, and to Enterprise workspaces where an administrator enables them (access details). Conversations with your dot do not count toward ChatGPT usage limits, but tasks it starts or manages in Work or Codex count toward those products’ limits as usual, with an allowance for deeper work and extended limits in the first month after launch.

This is day one, and I am deliberately not drawing conclusions. Over the next few weeks I will finish the WhatsApp relay, correct the drafts until they sound like me, and then take the same four screenshots I took this morning so the comparison speaks for itself. My advice to anyone trying this is to give it one clear responsibility first, usually email triage, and to judge it on whether things stop falling through the cracks rather than on how clever the drafts are. For someone who spent years openly choosing doing over responding, and never liking the cost of that choice, the hope is that I will not have to make it at all.