OpenAI is trying to move AI agents beyond software development and into the daily routines of a much larger professional audience. Its central vehicle is ChatGPT Work, a $20-per-month product released last month on the company’s lowest subscription tier. The product connects language models to workplace systems so they can complete multistep projects instead of merely answering questions.
The ambition is broad, but TechCrunch’s examination finds a product caught between impressive autonomy and demanding setup. Greater usefulness depends on access to email, calendars, documents and other services, creating an immediate trade-off between convenience and control. Andrew Ambrosino, the lead engineer for OpenAI’s desktop app, gives the system extensive access to his own digital accounts, while acknowledging the possibility that it could mishandle information from a private message.
Coding success meets a wider adoption gap
ChatGPT Work is a modified version of Codex, OpenAI’s coding tool. Software agents already help developers complete extended tasks, but OpenAI now wants accountants, investors, doctors and other computer-based workers to gain similar capabilities without navigating developer-oriented interfaces. The commercial logic is also clear: longer agent tasks consume more tokens, and expanding into more professions could increase the value generated by each subscriber.
Internal use shows both the opportunity and the gap. An OpenAI-backed study found that 98% of the company’s employees used Codex in June. Among organizational subscribers, adoption was 17%, while fewer than 1% of individual subscribers used the agentic coding tool. OpenAI would not disclose separate usage for Work and Codex; TechCrunch reports that the combined app has 20 million users, compared with more than one billion people the company says use ChatGPT online.
OpenAI has adapted the experience after non-engineering employees encountered technical language and software-specific outputs. Work adds visible controls for projects and plugins while retaining a conversational interface. Ambrosino argues that such guidance is necessary during an early phase, even if some buttons eventually disappear as people learn to ask the model directly for what they need.
Useful automation requires extensive permission
OpenAI currently positions Work for routine, data-heavy coordination. Employees use it for weekly metrics and planning spreadsheets. TechCrunch’s reporter used it to move a poorly formatted preschool calendar from email into Google Calendar, build an updating financial dashboard, assemble a searchable database of space launches and send a weekly research email.
Those successes came with friction. Configuring cloud-drive permissions proved confusing, and the system ultimately required complete access rather than the read-only permission the reporter wanted. Important settings were split between web and mobile apps. Calendar access could create events but not new calendars, and lower effort settings produced weak results. These are direct product limitations, not proof that the broader agent approach cannot work, but they raise the cost of early adoption.
Evaluation is another obstacle. Code can often be tested for whether it runs, while the quality of a presentation, strategy or sales pitch is harder to measure. OpenAI told TechCrunch it uses GDPVal, a benchmark covering 44 occupations and hundreds of knowledge-work tests, alongside user feedback. Actual usage may produce valuable training traces when users do not opt out, yet it also makes trust and data access inseparable from product improvement.
Cost, competition and control shape the bet
OpenAI’s team says model quality is the main differentiator, even as competing agent products use remarkably similar interfaces. The company’s first Codex web app expected too much independent action from the model. It later added more opportunities for users to intervene, following an approach that had helped Anthropic’s Claude Code gain traction. Download and enterprise-use indicators cited by TechCrunch now suggest Codex has caught up or taken a slight lead.
Independent comparisons complicate any claim that one model-specific interface will always perform best. TechCrunch cites tests showing that model and harness combinations produce different results. In one Databricks comparison, the open-source Pi harness outperformed Codex while both used the same GPT-5.5 model.
Economics may be equally decisive. The reporter consumed more than 80 million tokens in four days on a $20 subscription; the model estimated those tokens would cost $65, although the app offered no usage dashboard. OpenAI says it is improving efficiency and points to an 80% price cut for its Luna model. That is a company claim about progress, not evidence that today’s consumer pricing is sustainable at heavy usage.
ChatGPT Work therefore demonstrates both the appeal and unfinished nature of general-purpose agents. It can remove tedious work, but its strongest results require broad permissions, careful configuration, high effort settings and significant computation. OpenAI’s challenge is not simply making agents capable. It is persuading ordinary professionals that the access, cost and learning involved are justified by reliable outcomes.
