OpenAI has introduced "workspace agents" in ChatGPT. Teams can create shared assistants to tackle complex tasks and lengthy processes.
The virtual assistants are powered by Codex and can handle tasks typically performed by humans, such as preparing reports, writing code, and responding to messages.
These AI agents operate in the cloud, allowing them to execute assignments even when the human user is unavailable. They are designed for collaborative use within organizations. Teams can create an assistant once and then utilize it collectively in ChatGPT or Slack.
“AI has already helped individuals become more productive in their work. However, many key processes within a company rely on shared context, task delegation among employees, and inter-team decisions. Workspace agents are designed for this purpose: they can gather context from necessary systems, follow team processes, request approvals when needed, and coordinate work across different tools,” the announcement stated.
As an example, OpenAI mentioned a sales team agent that can compile information from call notes and client research, qualify new leads, and draft follow-up emails for the sales manager.
“This allows teams to spend less time on ‘stitching’ details together and more time engaging with clients,” the company noted.
The "workspace agents" are available through a research preview in the ChatGPT Business, Enterprise, and Edu plans, as well as in ChatGPT for teachers (K–12).
Privacy Protection by OpenAI
Separately, OpenAI has released a free tool called Privacy Filter, which sanitizes confidential information in conversations with ChatGPT before the chatbot sees it. This includes tax returns, medical records, work emails containing client names, or API keys.
The Privacy Filter is available on Hugging Face and GitHub and features 1.5 billion parameters. Anyone can download the software and use or modify it as needed.
When a user inputs text in a conversation with the chatbot, the program automatically replaces all confidential data with generic placeholders like [PRIVATE_PERSON] or [ACCOUNT_NUMBER].
The filter scans the text for eight categories of personal information: names, addresses, emails, phone numbers, URLs, dates, account numbers, and sensitive data like passwords and API keys.
OpenAI reported that the model is quite effective at recognizing important information, achieving a score of 96% on a standard test using the PII-Masking-300k dataset.
As a reminder, in April, Sam Altman's startup released the "thinking" image generator ChatGPT Images 2.0 — an advanced model capable of solving complex visual tasks and creating accurate, ready-to-use works.