--- title: "The next model will not empty your queue" subtitle: "GPT-6 Astra already drafts inbox, follow-ups, and ticket notes. Named people still know, select, and send." principle: "Work the overhang" date: "2026-09-19T14:22:05.094Z" author: "Mikko Atienza" canonical: "https://ciel.codes/showcase/blog/the-next-model-will-not-empty-your-queue" edition: "2026-09-19" source: "https://ciel.codes/showcase/pulse/2026-09-19" abstract: "GPT-6 Astra is already enough to change a week's work. Not in a demo. On the tickets, follow-ups, and inbox you will touch before Friday. The labs can freeze tomorrow. You would still have years of catching up inside the" cite: "Atienza, M. (2026). Work the overhang. Pulse." citations: - handle: "@emollick" url: "https://x.com/emollick/status/2101010622305493013" --- GPT-6 Astra is already enough to change a week's work. Not in a demo. On the tickets, follow-ups, and inbox you will touch before Friday. The labs can freeze tomorrow. You would still have years of catching up inside the account you already pay for. A shop of twenty usually treats the next release as the plan. The current model sits in a sidebar. The real queue stays in human hands because last quarter's trial was a blank chat and a cute prompt. That gap is the week. It is not a procurement. It is three live flows and a person who will still hit send. ## Hold the release. Work the gap. You already know which work repeats until someone gets tired. Inbox triage. The follow-up that should have gone yesterday. The ticket that needs a summary before the next person can move. Those jobs are not waiting on a smarter model. They are waiting on a trial that uses real threads, real names, and a named closer. Ethan Mollick (@emollick), a Wharton professor who watches how people actually absorb these tools at work, put the gap in one breath. > Even if AI development stopped today, we'd have years of catching up to do. The gap between what current models can do and what almost anyone is using them for is vast. > > Ethan Mollick (@emollick) He is not asking you to admire a trailer. He ran GPT-6 Astra on work that used to take a team weeks: reconstruct, iterate, drive other software, come back with a draft a person could reject. The interesting part for a shop this size is not the spectacle. It is that the same model will draft a refund note, a quiet follow-up, or a ticket summary in the time it takes your specialist to read it. What you still do by hand is not caution. It is unused capacity. Mollick's name for that unused capacity is the overhang. Close it on live work, or you will still be waiting when the next model ships. ## Put a name on each advantage A model that can draft the week still cannot own the week. Mollick's four advantages are not slogans. They are job seats. If nobody sits in them, GPT-6 Astra will fill the queue with fluent almost-right copy and you will spend Friday undoing it. - [ ] Name the specialist with deep knowledge of this work. They are the person who can glance at a draft and know the SKU is wrong, the tone is not yours, or the refund rule does not apply. - [ ] Name the person with wide knowledge. They know to ask for the pattern you actually use (a short status, a two-line apology, a handoff a tech can act on) instead of a blob of text. - [ ] Name who holds taste, and who holds the send. Taste keeps or kills. The send is a human name on a customer thread. Those can be the same person. They cannot be the model. - [ ] Name the owner who will try it on live work this week. That is agency. It is the difference between waiting for a memo and finding out on Tuesday whether Astra can triage your actual inbox. Scope and data access stay with the owner. Customer-facing sends stay with the closer. Internal notes can move faster, but they still get a human file. Nobody gets a pass because the draft looked confident. ```mermaid flowchart TD A["Owner names three live flows"] --> B{"Closer: will a customer see this?"} B -->|Yes| C["Named closer reads every draft"] B -->|No| D["Ops lead files the internal note"] C --> E["Specialist keeps or kills it"] E --> F["Closer sends under their name"] D --> G["Owner reads Friday's count"] F --> G style A fill:#1e1b4b,color:#ffffff style B fill:#4c1d95,color:#ffffff style C fill:#064e3b,color:#ffffff style D fill:#064e3b,color:#ffffff style E fill:#1e1b4b,color:#ffffff style F fill:#064e3b,color:#ffffff style G fill:#4c1d95,color:#ffffff ``` If a row has no named person, it is not a flow. It is a demo you will abandon. ## Tuesday is a live ticket, not a demo Give GPT-6 Astra the work object, not a toy. Paste the actual thread. Name the customer. State the rule the draft must not break. Then watch what it does. On inbox triage, it will sort, tag, and write a first pass. Your closer still decides what leaves. On follow-ups, it will write the reminder you have been carrying in your head. Your specialist still checks the promise, the date, and the dollar amount. On ticket summaries, it will compress the mess so the next person can start. Ops still files it where the team will find it. That is what the model actually does when you stop asking it for a poem. It compresses, drafts, and iterates when you reject the first pass. Mollick's point is that almost nobody is using that loop on real work. They are waiting for a version that will not need the loop. It will still need the loop. Taste is the scarce step once making is cheap. Do not expand the trial to your whole tool chest. Three flows. The ones that already hurt. If Astra invents a policy, the specialist kills the draft. If it wants data it should not have, the owner stops the paste. If it is good enough to send, a human sends it under their own name. ## What you can count without a study Nobody has a credible figure for how much of a twenty-person queue GPT-6 Astra could already draft. Do not fake one. Count the week you actually ran. | Measure | Count | |---|---| | Live flows run through GPT-6 Astra | | | Drafts a named specialist rejected | | | Customer sends with no named closer | | The first row should be two or three by Friday, not twelve. The second row is the point. Rejects mean deep knowledge showed up. A zero in that row usually means nobody read the drafts. The third row should stay at zero. A send without a closer is how fluent copy becomes a promise you did not make. > **Aside.** > Years of catching up is a claim about the world, not a tally for your shop. Friday's number is the flows you ran, the rejects you recorded, and the sends that still had a human name. If the reject pile is honest, you learned the shape of where Astra is useful on your work. If every draft sailed through, you did not staff the advantages. Staff them before you add a fourth flow. ## Run the live queue before Friday Pick the three flows on Monday. Put GPT-6 Astra on real threads by Tuesday. Keep the same closer through Friday so the count means something. - [ ] Owner writes the three flows on one page: inbox triage, follow-up drafts, ticket summaries. - [ ] Ops lead feeds Astra live items only, with the rule each draft must not break. - [ ] Specialist records every keep and every kill in the same place. - [ ] Closer sends, or holds. No draft leaves without their name. - [ ] Owner reads the three counts on Friday and decides which one flow survives next week. Stop waiting for a release to empty a queue it will not touch. You can run this yourselves on the account you already have.