AI That Clicks for You: What to Set Up Before Computer Use AI Goes to Work

For two years the AI we got used to was "you ask, it answers." It handed you text and you did the work. In early September 2026 OpenAI released GPT-6 Astra with a clear focus on computer use AI: the model looks at the screen, moves the cursor, clicks buttons and fills in forms until the job is finished. For businesses, that means AI can finally reach the systems that have no API and have always depended on someone clicking through them. It also means an AI that acts on its own needs tighter management than a chatbot ever did.

What computer use AI is: from answering to doing

According to OpenAI and press coverage, GPT-6 Astra can fill in forms, update CRM records, manage calendars, research the web, install and troubleshoot software, and produce documents, spreadsheets and presentations that follow your templates. It is rolling out to ChatGPT Plus, Pro, Business and Enterprise plans and is available through the OpenAI API, Microsoft Azure and AWS Bedrock.

9to5Mac reports that it can run multistep work in the background on your Mac or desktop and is nearly twice as fast at computer use as earlier models. DataCamp's overview puts the time per task at roughly 47% less than its predecessor, GPT-5.6 Sol.

Which tasks are a good fit

Not every task suits it. The best first candidates meet three conditions:

• Repetitive with clear rules. Pasting weekly report figures into a back office, or creating customer records in a fixed format.
• Moving data between systems. Pulling details from email into an ERP, or updating a CRM from a form.
• No API available. Ageing internal systems or vendor portals that only people could operate before.

Anything that needs judgment on tone, involves money or goes out to the public should stay out of the AI's hands: payments, emails to customers, deleting data. And if your system can be given an API, that route is faster, more stable and easier to audit. We compare the options in making legacy systems ready for AI agents.

A 72.6% score means it still gets things wrong

Figures collected by DataCamp show GPT-6 Astra scoring 72.6% on the OSWorld 2.0 computer use benchmark, up from 65.7% for the previous model. That is real progress. Read it the other way, though: on standardised test tasks, more than a quarter were still not completed correctly.

The same overview notes two limitations worth knowing. Safeguards can sometimes pause or stop legitimate work, and monitoring of the model's reasoning regressed under adversarial prompting. Treat it as a new colleague whose work needs review. That is far more realistic than treating it as automation that cannot fail.

Four safeguards to set before going live

OpenAI itself is cautious here. Astra is off by default for enterprise workspaces and an admin has to enable it. TestingCatalog also notes that the model pauses when a consequential choice needs user input. On top of that, we recommend four safeguards:

1. A dedicated account with minimal permissions. Never let the AI log in as a manager or administrator. Give it its own account with only the functions it needs.
2. A defined scope. List the websites and applications it may operate. Everything else is off limits.
3. A person presses the irreversible buttons. Submitting payments, sending external email and deleting data always require human confirmation.
4. Logs and spot checks. Keep a record of what it did. Check daily at first, then weekly once it is stable.

The division of labour is clear: AI handles the repetitive clicking and copying; people set the scope, press the irreversible buttons and own the outcome.

Start with one workflow

Skip the company-wide rollout and pilot one workflow for two to four weeks. Choose a task that takes at least two hours a week and can be corrected if it goes wrong. Record how often the AI completes it, how often a person has to fix it, and how much time is saved. With numbers in hand you can decide whether to expand, switch to an API integration, or keep the task manual.

If more than one AI is already running in your company, build an inventory alongside the pilot; see AI agent management starts with an inventory. To discuss which workflows fit and how to design permissions and confirmation steps, get in touch with TitApeX or browse our services.

References

• OpenAI: GPT-6 Astra — The next generation in intelligence for work
• 9to5Mac: OpenAI releasing major upgrade to ChatGPT and Codex with GPT-6 Astra
• DataCamp: GPT-6 Astra — Features, Benchmarks, and Pricing
• DEV Community: OpenAI Launches GPT-6 Astra With Computer Use Tools and Broad Platform Rollout
• TestingCatalog: OpenAI launches GPT-6 Astra across ChatGPT and API