A verdict about delegation
Dots Are the
Winner (So Far)
Elapsed time can go up while your attention cost goes down.
For a delegation dump: Dots.
For chatty back-and-forth: Chat.
I am super impressed with OpenAI’s Dots. For the way I want to delegate work, this feels like a game changer. The interface is clean, beautiful, and remarkably uncomplicated. But the reason I am voting for Dots goes much deeper than the interface.
It changes how much of the AI’s work I have to manage.
That is what I want the advanced student to understand. We have spent years learning to be good operators of increasingly capable AI. We choose modes, models, reasoning settings, conversations, projects, folders, and tools. Then we supervise the handoffs between them. A Dot moves more of that coordination into the system. I give it responsibilities; it works out how to carry them forward.
My vote is for Dots over the Grok and Claude experiences I have been comparing. That is my present preference for delegation, not a controlled performance ranking. “So far” matters: products change, and a winner should keep earning the title.
Before the AI works, you work.
Consider the ordinary experience of doing serious work with AI on a phone or in a desktop app. Before you even describe the outcome, there can be a little internal checklist. Am I in Chat, Work, or Codex? Am I in the right conversation? Does this conversation have the history? Am I in the correct project? Is the right folder attached? Can it reach the source documents? Is the connected account the account I intended?
Then comes the model question. Did I select GPT-6.1 Sol with the reasoning effort I intended, or leave a lighter setting in place? OpenAI’s current model documentation describes both model selection and reasoning controls; the available labels vary by client and account. I am not teaching that High is required for everything. I am pointing out that choosing the configuration has become another task for the person.
Each choice can be perfectly reasonable in isolation. Together, they make the user responsible for assembling the conditions under which the AI can succeed. The user is doing a small orchestration job before doing the actual job.
A wrong choice can be consequential. An excellent model working from an old proposal may produce an excellent revision of the wrong proposal. A task started in the wrong project can miss the decisions that make the request intelligible. A system without access to the relevant folder may improvise around the absence unless it recognizes the gap. Fluency can make these mistakes harder to notice.
This is why a beautiful answer is not sufficient evidence of a well-configured task. The setup determines what the answer is an answer to. Advanced users often absorb that work so thoroughly that they stop counting it. They call the AI fast while quietly donating their own attention to make it fast.
One place to hand over the work.
With a Dot, the starting experience feels closer to the original simplicity of a chat box. I describe what needs to happen. I do not want to begin by navigating a collection of conversations and deciding where each part belongs. I want the system to find the relevant history, understand the responsibility, and coordinate the work.
OpenAI describes a Dot as an ongoing agent powered by GPT-6 Astra, with a cloud computer and browser. It can delegate work to Work or Codex and keep going between conversations. That is the product foundation for the orchestration I am responding to. Source: Meet dots.
The simplicity does not mean the machinery has vanished. Conversations still exist. Delegated tasks have their own context. OpenAI documents searching prior ChatGPT conversations and continuing existing local Codex work, with access limits. It does not promise universal access to every conversation in every app. Source: Tasks and memory.
That distinction makes the accomplishment more interesting. There is still a coordination problem, but more of it can be handled on my behalf. The Dot can research the context, decide what work to start, and bring results back together. Finding the right project, files, and execution environment becomes part of carrying out the responsibility, within the access it actually has. It is not a guarantee that every choice will be correct.
I am impressed by the shift in who carries the effort. A clean interface is valuable when it removes work from the person, rather than merely concealing that work behind a prettier surface. The test is whether I have to keep returning to route, explain, rescue, and reconnect everything.
The use case that makes the difference
The delegation dump.
Several outcomes. One spoken handoff. Then the phone goes away.
“Look at my unfulfilled orders and put together the delivery route in Google Maps. Call everyone with an appointment today, confirm they are coming, and answer their questions. Work out who will be here around noon and arrange lunch for that group. Find the email details for this Friday’s consulting project and write and publish the proposal in the agreed place. Read yesterday’s Plaud transcripts and do the appropriate follow-up for each person.”
An illustrative delegation in my working style. These are requested outcomes, not a report that this particular batch was executed or a promise that every account has the tools to execute it.
This is the center of my argument. I might speak that request or dictate it, press Enter, and put the phone in my pocket. I have handed over a bundle of responsibilities. I am not settling in for a long conversation about them.
The bundle has structure. The route depends on the orders and their addresses. Lunch depends on who will actually be there at noon. The proposal depends on finding the correct consulting details. Transcript follow-ups depend on identifying the people, commitments, and next steps in the source material. Some work can run alongside other work. Some must wait for an earlier result.
A useful orchestrator recognizes those dependencies. It does not simply split a paragraph into five unrelated prompts. If two appointment confirmations change the lunch headcount, that change needs to reach the lunch task. If a transcript contains a promise already fulfilled in email, the follow-up should reflect that fact. The value is in keeping the work coherent as information moves.
The request also needs usable boundaries. “Appropriate follow-up” works only when the relevant practice is established or can be determined. Otherwise, the system needs a decision. “Publish the proposal” needs a known destination and intended audience. An advanced delegator gives clear outcomes and authority, then lets the system discover the operational steps.
There are two clocks.
The first clock measures elapsed time. Start it when the delegation is submitted; stop it when the last required outcome is actually delivered. For a batch, completion means the whole agreed batch, not the first attractive draft.
The second clock measures the person’s attention cost. Count the minutes spent setting up the work, giving instructions, monitoring and redirecting it, and reviewing or repairing the results. Then account for the time needed to regain focus after an interruption.
Here, S, D, M, and V are distinct periods of active human involvement, measured once each. Each r is additional recovery time after an interruption; do not count the same minutes again inside monitoring. For a simple approximation, replace the sum with n × r: number of interruptions multiplied by an assumed average recovery time. This is a practical accounting model, not a scientific law of attention.
Now the apparently strange result becomes ordinary:
The Dot can take longer to finish while requiring less of me. It may spend extra time researching the request, finding context, creating tasks, checking results, and coordinating follow-through. Some of that overhead is exactly the work I would otherwise have been doing.
This is why “it took twenty minutes” is an incomplete evaluation. Twenty minutes of independent execution and twenty minutes of sustained human supervision are different experiences. The same wall clock can conceal very different amounts of human work.
A worked example · assumptions, not measurements
Longer run.
Less of you.
Assume the same five outcomes are delivered to the same standard. Direct operation finishes in 18 minutes; Dot delegation finishes in 30. The Dot takes 12 extra wall-clock minutes.
Elapsed time / E
67% more elapsed time
Human attention / A
63% less active attention71% less attention cost
The active-attention assumption is eight minutes for direct operation and three for the Dot. Both figures include initial setup and delegation, monitoring, and final review. Assume five interruptions in direct operation and one with the Dot, each requiring an additional 1.5 minutes to regain focus.
Adirect = 8 + (5 × 1.5) = 15.5 minutes
ADot = 3 + (1 × 1.5) = 4.5 minutes
Attention recovered = 15.5 − 4.5 = 11 minutes
The reduction is 11 ÷ 15.5, or about 71%. Turn off interruption recovery above and the advantage is still visible: active involvement falls from eight minutes to three. Neither comparison demonstrates what Dots will do in your business. The example demonstrates why slower delivery and a better delegation experience can coexist.
Do not subtract attention minutes from elapsed minutes and call the result productivity. They measure different things. Also, do not treat every unattended minute as productive human time. I may use the interval to think, walk, eat, or do nothing. The benefit is the freedom to choose.
The deadline still counts.
Less attention is valuable only if the outcome is useful. A route that arrives after the driver leaves is a failure. A lunch order that misses lunch is a failure. A proposal built from the wrong brief is a failure, however few times the system interrupted me.
So my decision rule has conditions:
All three matter. I am willing to trade some elapsed time for a reduction in attention when the work still lands on time and meets the standard. I am not giving a slow system credit merely for staying quiet.
One-time setup belongs in the accounting too. Suppose connecting tools and establishing the workflow requires H minutes. Across N similar delegations, an illustrative average attention cost is H ÷ N + ADot. Repetition can spread that initial investment across many runs, but only if the workflow stays reliable. Reconnection, new instructions, and repairs add attention back into the total.
The advanced student should measure the whole relationship: what happened before the prompt, during execution, and after the supposed completion. A low-attention interface that exports the cleanup to tomorrow has not reduced the work.
Choose a Dot when you intend to leave.
A Dot is not my first choice for chatty stuff. If I want to explore an idea, ask a quick question, debate a sentence, or think aloud with an AI, I want a responsive conversation. I am present because the exchange itself is useful. Each answer helps me decide what to ask next.
That changes the mathematics. In a tightly coupled conversation, waiting sits directly in the person’s path. If there are k turns and each reply adds a waiting interval w, the accumulated waiting is approximately k × w. Actual delays vary, but the structure is clear: the next human step depends on the next answer.
A delegation dump has a different structure. The person supplies a bundle of desired outcomes and can disengage while the system works. The work can branch, research, and reconnect without requiring the person to conduct every transition. Extra coordination time may have little attention cost because I am already doing something else.
A Dot can converse, and a conversation can be necessary to clarify a delegation. My recommendation is about choosing the tool for the job. If I plan to remain in the exchange, Chat is usually the better starting point for me. If I plan to hand over responsibilities and leave, the Dot’s orchestration becomes the feature I want.
This is also why a single quick question is a weak test of my verdict. It scarcely exercises ongoing coordination, context discovery, dependency management, or follow-through. A delegation dump does.
The outcome leaves the app.
The next thing I want to notice is the work showing up where it belongs. The route is usable in Maps. The calendar reflects appointment confirmations. Lunch is actually arranged. The proposal reaches its agreed destination. The right people receive the follow-ups.
A suggestion to create a route is not a saved, usable route. A draft email is not a delivered email. “I will handle lunch” is not an order confirmation. Writing and sending are separate completion states. The system’s account of its work must preserve those distinctions.
This is a major reason I prefer the Dot model. The app becomes the point of delegation, while the benefits appear across the tools and relationships in which the work matters. I do not need to admire the stream of activity to receive the value. I need the work to arrive.
But leaving the app does not mean abandoning verification. A useful completion record identifies the actual deliverable, its destination, and what remains unresolved. For the dump above, I would want the route link, appointment outcomes with unanswered calls identified, the lunch confirmation, the proposal’s publication or delivery record, and a record of transcript follow-ups. Those receipts let me review consequences without reconstructing the entire execution.
Review time is V in the equation. Keep counting it. If the Dot reduces review by supplying clear evidence, that is a real gain. If it offers a confident summary that forces me to investigate every claim, the attention savings shrink.
Give it a clear responsibility.
The full dump assumes suitable tools, connected accounts, and permission to act. A Dot cannot call customers merely because the request contains the word “call.” It needs an appropriate calling capability and authorization for that outreach. Ordering lunch needs an actual ordering path, a destination, dietary information, and a spending boundary. Publication needs a destination and authority to publish.
OpenAI’s computers and apps guide distinguishes cloud work from access to your own computer. Local work requires that computer to be online with the app open; cloud work can continue while personal devices are off. Connected tools have their own permissions and accounts. These are operating conditions, not details a clean interface can abolish.
OpenAI also documents action review and instructions that can authorize ongoing work within a defined scope. A request to draft does not authorize sending. A saved rule does not override built-in requirements. A Dot may still need a decision, connection, sign-in, or approval. Source: Control your dot.
My instruction would therefore finish with something like this: “Use our established permissions and workflow. Bring me missing decisions and blockers. Keep routine progress out of my way. Give me the actual completion records.” The goal is to make interruptions meaningful.
“Put the phone away” describes the freedom to stop monitoring the stream. It does not mean make yourself unreachable when a necessary decision arises. The Dot earns that freedom by continuing within its authority and returning when human judgment is actually required.
Why Dots win for me, so far
I want to delegate
the coordination, too.
The advance is not simply that AI can do more tasks. It is that I can hand over a bundle of work without first becoming the administrator of every conversation, configuration, and handoff.
That is a substantial change in the relationship. I still define the outcome. I still establish authority. I still make the decisions that require me and review the consequences. But the system carries more of the work of finding context, selecting an execution path, coordinating dependencies, and following through.
For chatty work, I want the conversation. For the delegation dump, I want to speak, submit, and move on with my day. Fifteen or twenty minutes of execution can be perfectly acceptable if the work meets its deadline and does not need me to sit beside it.
Elapsed time can go up while my attention cost goes down. The work can take longer and give me more of my life back. That is why my vote is for Dots.
John Rector So far.