OpenAI's new agent runs whole jobs on its own, and AI labs move in on consulting work

Give ChatGPT Work one instruction and it will spend the next few hours doing the job on its own, then hand back a finished document, spreadsheet, or set of slides. OpenAI released the agent on July 9. It reads the files and apps a team already uses and breaks a large task into steps, working without someone watching.

Give it customer research and a target market, and it can draft a campaign brief, build the creative, then translate and reformat all of it for ten regional markets. It runs in the background, so a task started on a phone can run overnight on a laptop and be checked from any device, and on the desktop it can open local files and click through software the way a person would.

A person still approves the decisions that need judgment: the agent pauses and asks before sensitive actions, and an administrator sets which data and tools it may reach. OpenAI says nearly every one of its own teams, from finance to sales, now uses these tools, and the release comes months after Anthropic shipped a similar agent, Claude Cowork.

Earlier versions of ChatGPT answered a question; this one does the job the question was about. The work it targets, gathering the data, drafting the report, formatting the deck, chasing updates across Slack and Teams, is the daily substance of a coordinator, an analyst, or a junior marketer. It runs on the new GPT-5.6 models and is available first to Pro, Enterprise, and education accounts, with Plus and Business users following in the coming days.

On June 18, Accenture had the worst day in its history as a public company. Its shares fell about 20% after it forecast weaker revenue and softer bookings, and Bloomberg Intelligence tied the slide to a single worry: that AI is cutting into demand for consulting itself. The stock is down more than 50% this year.

The work under threat is specific. A large share of consulting revenue comes from helping companies build software and put new systems in place, billed by the hour. When AI tools do more of that in less time, the hours fall.

Now the model makers want that work directly. OpenAI's deployment arm, seeded with $4 billion to buy companies, agreed this week to acquire Northslope, its second such purchase in two months. It adds hundreds of what the industry calls forward deployed engineers, people who sit inside a client and build AI systems around how the business runs. Palantir has used the method for years, and Northslope's founders came from there; Microsoft and Anthropic have built the same kind of arm.

Frontier models now perform much alike, so the edge is no longer the model but getting it working inside a specific, messy business. For a buyer, the company that once charged to install the system and the company that makes the model are becoming the same vendor.

The European Commission ended its antitrust case against SAP on July 9 after the company agreed to loosen the rules that keep customers tied to its support. SAP's annual maintenance runs about 22% of a licence's value, often the largest recurring line in an IT budget after staff, and third-party providers charge roughly half that.

Until now, leaving was costly and hard to reverse. Under the binding commitments, which apply worldwide for ten years, SAP will drop the fees charged to customers who return after using another provider, and let a company split its systems and support each part separately, or not at all.

That gives a finance or procurement chief a real option: keep the existing software on their own systems rather than move to SAP's cloud. The change still matters for AI, because the automation SAP is selling, the tools that trim work in finance and procurement, is offered mainly to customers who make that cloud migration. Mainstream support for the older ECC system ends in December 2027.

In mid-June, the US government ordered Anthropic to block foreign nationals from its two newest models. Because it could not screen in real time who was on which account, Anthropic switched both off for every customer worldwide the same evening, three days after launch. China is now weighing the same move in reverse, considering whether to cut foreign access to its most advanced models.

For any business outside the two countries, that is a plain continuity risk: a tool a team depends on can be repriced or withdrawn by a government it does not answer to, with no notice.

One answer now drawing money is to own the model instead of renting it. Prime Intellect, which sells the tools and computing power to train a company's own agent, raised $130 million this week. It is practical now because of a technique called reinforcement learning, which lets a company refine a smaller model on its own data and tasks instead of building one from scratch. The spending platform Ramp used it to build an agent that answers questions inside its own spreadsheets, the kind of narrow, reliable tool it wanted rather than waiting on a better general model. A model a company trains and holds itself cannot be switched off from Washington or Beijing.

An AI notetaker joins a video call as a participant, records everyone, and minutes later sends a summary and a task list. The legal exposure it creates falls on the person who set it up, not the vendor.

When the tool labels who said what, it builds a voiceprint, a biometric identifier that Illinois law requires written consent to collect; class actions against Fireflies.ai and Otter.ai are testing that now. A bot that sends a client call to an outside server can also waive the legal privilege a company assumed covered the conversation: in February a New York federal judge ordered a defendant to hand prosecutors documents he had shared with Anthropic's Claude, because sharing them with a third party broke the protection.

Consent from the host does not pass to the other participants, and more than a dozen states require everyone on a call to agree before recording.