The intern that never sleeps: ChatGPT Work and the week agents started finishing things

For three years, the ritual has been the same: you ask the AI, it answers, and then you do the actual work of turning that answer into the thing your job needed — the spreadsheet, the deck, the memo with the numbers in the right cells. ChatGPT Work, which OpenAI shipped alongside the GPT-5.6 launch this week, is aimed squarely at that last mile. You hand it a goal. It gathers what it needs across your connected apps and files, breaks the job into steps, executes them, and comes back with the finished artifact. Not a draft of an answer — a built sheet, a built deck, a built document.

What's genuinely new here

Agents that promise autonomy are not new; agents that ship finished work into the tools you already use are. The difference is integration depth. ChatGPT Work sits inside a workspace that unifies conversation, coding, and productivity — which means the agent isn't reaching across a fence into your files; it lives where they live. Whether that turns out to be delightful or unsettling in practice will depend on the boring things enterprise software always depends on: permissions, audit trails, and what happens when the agent confidently builds the wrong thing.

The same week, Accenture and Google Cloud announced a suite of pre-built agentic solutions aimed at mid-market companies — the $300 million to $3 billion revenue band that has historically been too small for custom AI consulting and too complex for off-the-shelf tools. That's the tell that this is a category now, not a feature: when the consultancies start packaging it for the middle of the market, the technology has left the demo phase.

The mid-market framing deserves a second look, because it's where the labor math gets real. A company with $500 million in revenue typically can't afford the seven-figure custom-agent consulting engagements the Fortune 100 has been buying, and until now that priced them out of serious automation. Pre-built agentic suites change the entry ticket: the pitch is a deployable agent for a defined back-office workflow — claims processing, order reconciliation, report assembly — at software prices rather than consulting prices. If that lands, the agent era arrives at thousands of mid-sized firms more or less simultaneously, in exactly the functions where a single automated workflow can displace meaningful headcount. The Fortune 100 adopted agents with pilot programs and change-management budgets. The mid-market will adopt them the way it adopted cloud software: quickly, cheaply, and with much less cushion for the people affected.

The question stopped being 'can AI do my tasks?' and became 'who checks the work when nobody watched it happen?'

The skill that just got scarce

A Korn Ferry talent survey published this week lands on exactly this point from the hiring side: 73 percent of talent leaders now rank critical thinking and problem-solving as the top skill they hire humans for. AI skills — prompt fluency, tool knowledge — ranked fifth. Read those two rankings together and the labor story of the agent era writes itself: as agents produce more of the first draft, the scarce human skill becomes judging the output. Catching the hallucinated number in the built spreadsheet. Noticing the deck's argument is subtly wrong. The first draft is becoming free; the judgment layer is becoming the job.

The honest caveat: launch-week agents always look better in launch videos than in month three. The pattern to date has been that autonomy works brilliantly on well-structured tasks and degrades fast on ambiguous ones, and nothing published this week suggests that curve has been repealed. What has changed is the default: the tools now assume they'll finish the job, and it's on you to catch them when they shouldn't have.