What is ChatGPT Work and should your firm switch it on?

ChatGPT Work is OpenAI's new agent mode, rolled out to paid plans over the past fortnight, and it changes what ChatGPT actually is. Instead of answering one question at a time, it can now take on a whole task, break it into steps, work across your firm's connected apps (email, calendar, SharePoint, Teams, Google Drive and more) and hand back a finished document, spreadsheet or presentation. For Australian professional services firms, the short answer is: it is worth a controlled pilot this month, but only after you have set the admin controls and decided which of your firm's data it is allowed to touch.

What did OpenAI actually announce?

On 9 July 2026, OpenAI launched ChatGPT Work, a new agent built into ChatGPT on web, mobile and desktop. Alongside it came a new family of GPT-5.6 models in three tiers: Sol (the flagship for the hardest work), Terra (the balanced everyday option) and Luna (fast and cheap). The rollout started with Pro, Enterprise and Edu subscribers, with Plus and Business plans following over the days after launch, which is why many firms are only seeing it appear now.

The important shift is not the new models. It is the move from a chatbot that answers questions to an agent that completes work. ChatGPT Work can plan a multi-step task, keep going for an extended period, pull information from the services you connect to it, and produce documents, spreadsheets, presentations and simple web tools as the output.

What can it actually do inside a firm?

Think of ChatGPT Work as a junior team member who never gets tired but needs very clear instructions and full supervision. Connected to the right systems, the kinds of tasks firms can realistically hand it include:

• Drafting from your own material. Pull the relevant background from a matter folder, engagement file or project directory in SharePoint or Google Drive, then produce a first-draft report, letter of advice, engagement summary or client update in your firm's structure.
• Research and synthesis. Run a research task across multiple sources, compare what it finds, and return a structured briefing note rather than a wall of links.
• Document production. Turn meeting notes and source files into a formatted spreadsheet, board paper or client presentation, rather than you assembling it manually.
• Multi-step admin. Chase down the inputs for a recurring deliverable across email, calendar and project management tools, and assemble the pieces into a draft for review.

For an accounting practice that might mean a first-cut management report pack from the month's files. For a law firm, a chronology or brief summary drawn from a matter folder. For an architecture practice, a consolidated consultant-comments register. For a real estate agency, listing copy and vendor reports from file notes. For a medical or allied health practice, the strict rule stays the same as always: no patient-identifiable information in any general AI tool without proper agreements and approvals in place.

What about the new GPT-5.6 models?

The GPT-5.6 naming (Sol, Terra and Luna) is OpenAI simplifying a model line-up that had become confusing. For most firm users the practical guidance is simple: the default Sol model on paid plans is what you will use in chat, and you do not need to think about the other tiers unless your firm builds on the API. OpenAI says Sol sets new marks across coding, knowledge work and science, and the older GPT-5.4 models are being retired from 23 July, so any internal guides or prompt libraries that name a specific old model are worth updating this week.

Where are the risks for a professional services firm?

This is the section to read twice, because an agent that can act is a bigger governance question than a chatbot that can only talk.

• It can reach further than you think. ChatGPT Work connects to email, calendars, file storage, CRMs and (on the desktop app) local files and browsing. Every one of those connections is a decision about client-confidential data, not a convenience setting.
• Confidentiality and privilege obligations do not change. Whatever the tool can access, your professional obligations around client confidentiality, legal professional privilege and privacy law still apply. If a folder should not leave the firm, it should not be connected to an agent.
• Oversight features exist, but they need to be switched on and used. On Enterprise and Edu plans, administrators control which tools and data the agent can reach, an auto-review system checks sensitive actions before they happen, and a Compliance API gives visibility into what the agent has done. Independent validation of these protections is still thin, so treat them as a seatbelt, not a guarantee.
• Output still needs a professional's eye. Agent mode produces polished-looking work at speed. Polish is not accuracy. Everything it drafts is a first draft until a qualified person has checked it, exactly as you would treat work from a new graduate.

What this means for your firm

Our practical read: this is the clearest signal yet that AI in professional services is moving from "asking questions" to "delegating work", and firms that build the supervision habit early will get the compounding benefit. A sensible path for the next month looks like this:

1. Decide who pilots it. Two or three people, on a Business or Enterprise plan, not the whole firm at once.
2. Start with low-risk connections. A sandboxed folder of non-confidential material first. Prove the workflow before you connect anything that matters.
3. Set the admin controls before the first task. Connector permissions, auto-review and logging are the point of the exercise, not an afterthought.
4. Pick two recurring tasks. Choose deliverables your team produces every week, run them through ChatGPT Work with human review, and compare time and quality honestly.
5. Write down what you learn. A one-page internal note on what worked, what failed and what is not allowed becomes the start of your firm's AI policy in practice.

Want help getting this right?

Rolling out agentic AI across a firm is exactly the kind of change that goes better with structured training than with one enthusiastic early adopter. Jork Training runs live sector-specific sessions on putting AI to work safely in professional practices, and our team delivers in-house programs that take a whole firm from first prompts to supervised agent workflows. Browse the upcoming webinars for your sector, or find out more about in-house training and consulting if you want a program built around your firm's systems and risk profile.