Your AI assistant is becoming a work router. Build the queue before you trust the agent.

Something shifted this week, and it was not that OpenAI announced a better model.

GPT-5.6 is the headline. Fine. Stronger reasoning, longer tasks, a very handsome product page. We have all seen that film. We know how it ends. Everyone claps, nothing changes on Monday.

The more interesting part is ChatGPT Work.

OpenAI is now describing an agent that pulls from Slack, Teams, Drive, email, calendars, CRMs, project trackers and local files, then sits with a task for hours. Research becomes a campaign brief. The brief becomes assets. Context survives the whole way through. It ships with scheduled tasks and admin controls, because even OpenAI has worked out that letting an agent run unsupervised across a business is an expensive hobby.

That is not a chatbot story. That is a work-routing story.

And this is where I think a lot of businesses are about to trip over their own shoelaces.

The prompt box was never the point

For two years, most AI adoption has lived inside a text box. Someone opens a tool, asks for a draft, pastes the answer somewhere else, and then tells the board the company has been transformed. It has not. It has added a faster text box next to the existing mess.

The agent layer is a different animal. Once the system can read from your real tools and keep working while you are asleep, you are not managing a clever assistant. You have opened a new route for work to move through the business.

New routes need design. Otherwise they become new leaks.

Dan Shipper’s Codex Desktop walkthrough is useful here because it shows the shape of the thing without dressing it up in enterprise language. Persistent threads for different domains of work. Scheduled pulses that inspect email, Slack and meeting notes. Activity turned into cards with a summary, a proposed action and a review state. Work handed to the agent through a router instead of being dumped into a human inbox.

That pattern matters more than whether you personally care about Codex, ChatGPT Work, Claude Code, Gemini Enterprise or whatever gets renamed next Tuesday.

The durable version of the pattern is simple. AI work needs a queue.

Not a chat history. Not a Slack channel full of half-formed requests. Not “monitor my inbox and tell me what matters”, which is less an instruction and more a wish.

A proper queue.

What a queue actually contains

Every item in it should say where it came from, what type of work it is, who owns it, what evidence the agent used, what it is allowed to do, what needs approval, how much it can spend, what good output looks like, and the point at which a human has to step in.

Boring? Yes. That is usually where the money is.

Compare that to the thing most companies are about to point their agents at. The inbox mixes urgent client work, newsletters, supplier noise, internal politics, receipts, spam and half-remembered ideas in one pile. Ownership is vague. Dedupe is nonexistent. There is no clean line between read this, draft this, decide this, do this, and this is technically FYI but someone will lose their mind if it gets missed.

The inbox is a terrible operating system. It was never designed to be one. It is a pile with a search bar.

Now point a long-running agent at that pile and ask it to help.

It will help, for a while. Then it will invent a new species of confusion. Beautiful summaries of low-value noise. Well-drafted replies to things that needed no reply. Escalations with no commercial judgement behind them. Tasks that look complete right up until someone asks where the evidence came from and the room goes quiet.

This is why the queue beats the prompt.

A good queue forces the decisions up front. Is this item read-only, draft-only, or allowed to act? Can the agent send anything externally? Can it touch the CRM? Can it create a document, update a forecast, spend credits, assign work to a human? Who signs it off? What gets logged?

Most companies cannot answer those questions. They have tool subscriptions and optimism.

The industry is quietly saying the same thing

OpenAI’s own product page gives it away. The examples are not “write me a LinkedIn post”. They are month-end budget variance, lead review, launch checks, sales-meeting prep, account tracking, conference follow-up. Messy operational work with sources, owners and consequences attached.

OpenAI’s Codex research makes the same point from another angle. Agentic AI changes the unit of knowledge work from a single interaction to a delegated, longer-horizon task, and inside OpenAI that use has spread well past engineering into finance, legal, recruiting and support. Treat the task-duration numbers gently, since some of them are model-estimated rather than measured. The direction still holds.

Google’s agent-enterprise guidance opens with a set of deliberately awkward questions. Who is building the agent? Who is it for? Is it talking to humans or to other agents? What tools can it use? How does identity work? How do you govern it? Strip out the platform pitch and the questions are the entire product.

Anthropic’s Claude Code research arrives at the same door from a third direction. Humans still make most of the planning decisions. The agent makes more of the execution decisions. Domain expertise matters more, not less, because someone has to know what is worth doing, what counts as done, and when the agent has quietly gone sideways.

That is the bit a lot of AI commentary keeps skipping. Agents do not remove responsibility. They move responsibility earlier in the system.

If the agent executes, the human has to be clearer about intent, boundaries, evidence and review. If the agent works across tools, the business has to be clearer about permissions and ownership. If the agent runs on a schedule, someone has to define what gets surfaced, what gets ignored and what gets escalated. The thinking does not disappear. It gets front-loaded.

Where the commercial opportunity moved

The old agency model sold output. Campaigns, websites, content calendars, reports, decks. A lot of that output is getting cheaper, and some of it already has. Fighting that is a waste of a decade.

The value is moving to the operating system around the output.

What enters the queue. What the agent is allowed to read. What it must never touch. Which reports trigger a decision instead of dashboard theatre. Which client signals become tasks. Which proof has to be attached before anything gets published. Which actions need a human signature. Which workflows deserve automation, and which should stay manual because the edge cases are genuinely horrible.

None of that is glamorous. It is plumbing, governance and commercial judgement.

Good. That is where clients actually need help, and it is considerably harder to copy than a content calendar.

Start with one read-only queue

If you own a business, do not start with “which agent should we buy?”. Start with one queue.

Pick a boring workflow. Sales follow-up. Monthly reporting. Content research. Support triage. Proposal prep. Meeting-note actions. Something repetitive enough to matter, and risky enough that you would not let a bot freestyle it.

Then make the first version read-only.

Let the agent gather the source material, summarise the item, classify it, suggest the next action and attach the evidence. Do not let it send, update, publish or spend. Watch what it misses. Watch what it over-escalates. Watch where the source trail breaks and where a human still has to make the call.

Then, and only then, widen the permissions.

Slower than the demo? Yes. It is also how you avoid handing an invisible junior employee admin access and no manager.

The companies that win with agents will not be the ones with the longest prompt library. They will be the ones who know how work enters the system, where it waits, who owns it, what the machine may do, what the machine must prove, and when a human takes over.

The chat box was the first interface.

The queue is where this gets serious.

---

Pull quotes

• The next AI advantage is not the cleverest prompt. It is the cleanest route for work moving through the business.
• The inbox is a terrible operating system. It is a pile with a search bar.
• Agents do not remove responsibility. They move responsibility earlier in the system.
• Most companies cannot answer those questions. They have tool subscriptions and optimism.
• Build one read-only queue before you let an agent touch anything real.

Sources

• OpenAI, ChatGPT is now a partner for your most ambitious work, 2026-07-09
• OpenAI, How agents are transforming work, 2026-06-25
• Google Cloud, 20 questions for the Agentic Enterprise, 2026-07-07
• Anthropic, Agentic coding and persistent returns to expertise, 2026-06-16
• Dan Shipper / Startup Ideas, Codex Desktop as a work OS, 2026-07-09

Editorial notes

• OpenAI’s ChatGPT Work page is product marketing. Its examples are signals, not proof that every workflow performs as described.
• Codex work-duration figures include model-estimated human-time thresholds. Directional, not precise.
• The Dan Shipper reference is drawn from a machine transcript with known name and product errors. It supports the workflow pattern, not verbatim quotes.
• The argument is deliberately narrow. It does not claim agents are safe or universally ready. It claims that if agents are going to route work, the queue matters more than the prompt.