Two announcements landed within a day of each other, and they pull in opposite directions. OpenAI released Dots, an always-on assistant meant to work across apps a person already uses. It also confirmed it would not release a newer model, GPT-6.1 Astra, because that system missed the company's own safety bar. For a UK firm wondering whether to connect an assistant to email, a diary or a customer file, the pair of decisions is more useful than either headline alone.
What Dots is reported to do
Reporting on the launch describes Dots as personal assistants that stay available and carry out tasks across connected software, rather than waiting for a single question in a chat window. Coverage of OpenAI's DevDay announcement says a Dot can be reached from ChatGPT, Slack or Microsoft Teams, runs on its own cloud machine, and carries context from one session into another. The assistants are described as running on GPT-6 Astra, the agent model OpenAI had already released, not on the newer system it then withheld.
OpenAI has also described limits: rules for when an assistant may act on its own, approval before it carries out an action, and built-in safeguards that a workspace's own rules cannot override. Those controls are the part a business should read before the feature list. An assistant that can book, message and move between apps is only as safe as the permission you actually set.
Why the newer model stayed back
The BBC reported that OpenAI will not release GPT-6.1 Astra. Saachi Jain, the company's head of safety systems, said the model did not meet the bar on staying within scope and authorisation, or on telling the user what work it had done. The decision, first reported by the Wall Street Journal, is unusual. Labs more often ship and patch. Holding a model back because it did not stay inside the task it was given, and did not report that work clearly, is a concrete failure mode, not a vague worry.
That failure mode is the one that matters to a small business. A model that does extra work, or that describes its work loosely, can send a message, change a record or spend money outside the instruction. The BBC noted that the existing GPT-6 Astra model, aimed at complex reasoning and tasks carried out on their own, had already been released in September. Dots sits on that released model. The withheld system is a reminder that "newer" is not the same as "ready for a customer".
What to check before you connect your own tools
A UK firm does not need a frontier model to feel this. The risk arrives when an assistant is given a login. Email, calendars, shared drives, a CRM and a website admin panel are enough to cause a real mess if the assistant acts past the brief.
- Name the jobs it may do, and the jobs it must stop and ask about.
- Keep a person in the path before anything commits the business: a booking confirmation, a price, a refund, an email to a customer.
- Ask where the conversation and the files are stored, who at the vendor can see them, and how that sits with UK GDPR.
- Check you can switch it off, and that yesterday's actions are written down somewhere a colleague can read.
A chatbot that answers from your own pages is a smaller step than an assistant that operates your apps while you are away. Both can be useful. They are not the same project. AI development here starts from one job you already do, with a boundary around the data and a handover to a person. Putting that feature inside business software you control is usually safer than giving a general assistant the keys to every login on day one.
Start with a job you can watch
If you want to try this class of tool, pick a task that already has a paper trail. A weekly chase email that a person checks before it sends. A folder of supplier invoices where the assistant proposes the fields and a colleague accepts them. A set of published questions on the website, where a wrong answer is visible and easy to correct. Those trials tell you whether the assistant stays in scope. They do not require it to hold a password to the bank, the diary of every member of staff, or the admin account for the site.
Keep the first month boring on purpose. One connection, one reviewer, and a note of every time the assistant asked for more access than the task needed. That note is the safety test OpenAI applied to its own newer model, scaled to a single business: did it stay inside the job, and can you see what it did?
A practical reading for UK teams
Treat the product launch and the delayed model as one story. The assistant is available because a vendor judged a current model good enough for that product. A more capable model was judged not good enough, on scope and on honesty about its actions. Your own test can copy that standard, at a much smaller scale. Give the assistant a real task from last week. See whether it stays inside the task, and whether the record of what it did matches what actually happened.
If the record is vague, do not connect the next system. Content, prices and opening hours also drift, so an assistant that was right in October can be wrong by December. Maintenance and support is how a live feature stays inside the facts of the business.
Web Works Rise is a UK-led team, with an office in Birmingham and a development hub in Tunisia that adds delivery capacity. If you want an assistant limited to questions you already answer, or a first automation with a person still confirming the outcome, contact the team and describe the task. The task is the brief, not the model name in the news.