OpenAI is moving the AI proxy from a programmer tool to a broader office scene. TechCrunch reports that OpenAI wants ChatGPT Work to access mailboxes, Slack, mobile phones, and applications such as Notion, Figma to help users do their daily work across tools and steps, rather than just answering questions.
From writing code to processing workflow
This product was launched last month and is open to subscription levels of $20 per month. It is based on OpenAI ' s Codex tool, and aims to expand the capacity of the engineer ' s agency to more white-collar positions such as finance, communication, investment, medical, etc.
OpenAI Product Manager stated that the company wanted ChatGPT to be able to perform complex tasks on its own, not just in question and answer areas. For OpenAI, it's not just product direction, but business models. Longer and more step-by-step proxy assignments will consume more computing and token, generating higher single-user income.
Internal HF usage and external use remained limited
According to a study supported by OpenAI, 98 per cent of OpenAI staff were using Codex in June of this year, but only 17 per cent of organizational subscribers and less than 1 per cent of individual subscribers were using Codex.
This set of data reflects the reality that AI agents are already being used at high frequency within the company, but still face a clear threshold to move towards a wider market. OpenAI does not disclose a single user size for Work, but only indicates that there are approximately 20 million joint application users associated with Codex, while ChatGPT has more than 1 billion online users.
Hard to use in privileges, interfaces and habits
In the view of the OpenAI engineering team, the key to the real use of AI agents by non-engineered users is not just modelling capabilities, but also the design of the outer layer. If the model is to perform its tasks, it must be given more tool privileges and understand the software, website and work records used by users.
This also raises privacy and control issues. The report mentions the risk of extracting sensitive information when agents have access to private mail, mailboxes or team communications. OpenAI engineers admit that these risks exist objectively, but believe that their boundaries can only be tested if the product is actually put into the work stream.
To lower the use threshold, OpenAI is making more intuitive the technology-based interactive approach. Work added a much more understandable entry to project selection, plugins, etc. than command lines and code interfaces, in the hope that non-technical jobs would also be available.
I'll start with a lot of data.
At present, OpenAI continues to position such products in favour of conventional, data-intensive tasks, such as the organization of weekly indicators, the conversion of tables into planning tools, the consolidation of communication records, the generation of graphs and internal analysis materials.
It was also mentioned that some wind-in agencies were already using them to collate company information and generate investment memorandums, and that some operators were using it to build customized dashboards and visualization of data. OpenAI believes that employees of an enterprise are facing a growing amount of information on a daily basis, and that the value of AI agents is to help users to screen and act more quickly on the basis of the system privileges already in place.