How much autonomy are you willing to grant a large language model (LLM) in your digital life? Maximizing the utility of such models often requires relinquishing control, which can be daunting for those hesitant about AI or protective of their personal data. For Andrew Ambrosino, the lead engineer of OpenAI’s desktop application, embracing this level of access is crucial for innovation; his app has full access to his inbox, Slack, Notion, Figma, and more.
“If I'm requesting it to draft a document, there's always a chance it could reference a private message related to that document without recognizing that it shouldn't disclose specific details. Yes, that risk exists,” Ambrosino explained to TechCrunch. “I’m willing to face personal risks when necessary, but fortunately, I haven't encountered significant issues.”
Ambrosino is currently focused on OpenAI's ambitious ChatGPT Work initiative, launched last month, which can be accessed through the company’s entry-level $20 monthly subscription. Designed for white-collar professionals, it connects LLMs to the automated workflows of diverse fields such as finance, healthcare, and accounting.
OpenAI's promotional material encapsulates this vision: a future where artificial intelligence transcends mere question-answering, enabling users to transform their grandest ideas into tangible outcomes.
While the integration of AI tools has been steadily advancing in software development, wider adoption across other professional domains has been slower. ChatGPT Work, built upon the Codex coding model, aims to equip non-engineers with the powerful capabilities that software developers have accessed from agents — allowing AI to independently complete complex tasks rather than just respond to queries.
Thibault Sottiaux, who heads OpenAI's core product development, described the potential: “In this new paradigm, ChatGPT can autonomously tackle complex projects, providing an experience that is both enjoyable and secure. This aligns with OpenAI’s mission to democratize technology.”
From a commercial perspective, this transition is critical. Agents that perform tasks over extended periods generate more tokens, enhancing profitability on a per-user basis. Expanding into new job roles is essential, not just for OpenAI but for the entire AI industry. While AI-focused on programming has proven lucrative, it only represents a fraction of the broader professional landscape, necessitating advancements beyond traditionally lucrative coding applications. Competitors like Harvey (aimed at legal professionals) and Clay (centered on sales) are also vying for similar user segments, adopting a model-agnostic strategy to leverage the most effective AI solutions available.
Analysts highlight this as a significant hurdle for OpenAI and its counterparts. “If AI research labs fail to quickly acquire the necessary complementary assets for scaling in various markets, the value will shift elsewhere,” cautioned Christian Catalini in a16z’s “It’s time to build” blog.
To make AI applications user-friendly for those outside the programming realm, OpenAI’s non-technical teams began implementing Codex “during a period when it was rather unwelcoming to them—prompting queries about coding and displaying technical outputs like an empty diff for changes,” Ambrosino noted. “Thus, we initiated a shift toward a more general-purpose model from February until now.”
An internal study backed by OpenAI revealed that while 98% of its employees utilized Codex, only 17% of organizational subscribers and fewer than 1% of individual users engaged with the agent-based coding tool. This discrepancy between the almost universal internal adoption and minimal external usage illuminates both a challenge and an opportunity for the company.
“The more value we create for our users, the more they are inclined to pay for a portion of that value, and that’s fundamentally how we view ChatGPT,” Sottiaux stated. “When you use it, the compelling benefits make paying $20 a month feel justified.”
Understanding the disconnect requires insight into OpenAI’s engineering approaches. Each LLM necessitates what is termed a “harness”—the software interface surrounding a model that filters information, dictates tool usage, and formats responses for users.
If the ambition is to enable the model to carry out tasks autonomously, this harness must equip it with the necessary tools and directives for comprehensive projects. While developers find command-line interfaces (CLIs) effective for LLM-based coding, the average user prefers more intuitive, user-friendly designs.
“You need a solution that interacts seamlessly with your chaotic digital life and older tools,” Ambrosino told TechCrunch. The aim is to make AI capabilities more accessible through tools that simplify complex interactions.
Applications like Claude Code and Codex have initiated a “vibe coding” revolution by removing the complexity of writing actual code, letting users articulate their needs to the model. OpenAI aims to streamline the functionality found in advanced tools like OpenClaw, making them accessible via simple prompts.
“Without these solutions in front of the model, experts might achieve the desired outcomes, but mass adoption would remain elusive,” Ambrosino explained. The ongoing debate within OpenAI highlights a tension between what power users require and the needs of mainstream users.
As for ChatGPT Work, it offers additional capabilities for selecting projects and integrating plugins, all while adhering to the same user-friendly interface ethos as other OpenAI services. Ambrosino compared this evolution to skeuomorphism—when digital tools mimic physical counterparts. “This wasn’t merely design fluff; it helped users acclimatize to the transition,” he remarked.
While OpenAI has yet to disclose specific usage statistics for Work in comparison to Codex, the app is currently utilized by around 20 million users, a stark contrast to the more than a billion users engaging with ChatGPT online.
OpenAI has positioned ChatGPT Work as an optimal solution for routine, data-heavy coordination tasks. Its employees are using it for weekly metrics reports and transforming spreadsheets into planning documents.
I’ve interacted with venture capitalists using such agents for assembling analytical notes and investment documents, while operations teams are crafting tailored dashboards and visualizations. Even Sam Altman has been using it to streamline vacation planning. One OpenAI engineer described how he used the program to analyze a Slack conversation about a technical challenge and receive insightful data visualizations in return.
“There’s an overwhelming amount of information at our disposal as employees,” said Akshay Nathan, who leads product engineering at OpenAI. “The challenge lies in effectively sifting through this information and taking actionable steps based on it. Much of this data exists across various platforms, and that’s where ChatGPT shines, unlocking access to it all.”
This performance could represent the digital personal assistant many AI enthusiasts have envisioned. Like other agents such as Claude Cowork and Perplexity AI, ChatGPT Work integrates seamlessly into your existing digital workspace—email, browsers, and various SaaS solutions—and leverages context effectively.
In practice, its capabilities can be impressive: I tasked ChatGPT Work with retrieving an oddly formatted preschool calendar from my email and inserting it into my Google Calendar, saving me from mundane data entry. I’d rather not forget school events or miss out on arranging childcare during vacations.
Though I was initially hesitant to trust OpenAI with sensitive access to my emails, interviews, or drafts (AI critics can breathe easy), I found it could indeed be more beneficial with greater trust. I also requested it process financial data on publicly traded firms, leading to an auto-generating dashboard that tracked key metrics, something I used to do manually with Python scripts. It even sends me weekly academic updates on new AI research.
However, the challenge lies in making the tool intuitive to set up. Configuring permissions for agents to access cloud drives proved confusing, with multiple attempts required for something as simple as read-only access, often met with cryptic error messages. The model wasn’t particularly helpful, but eventually, a mobile app prompt clarified that only complete access would allow it to function properly.
Many essential features are only accessible via the web app, necessitating a simultaneous use of both platforms. Occasionally, ChatGPT Work's limitations can be perplexing—it can create events linked to your Google Calendar but can’t generate new calendars. And using it for low-effort tasks can yield frustrating results, akin to dealing with a less capable intern.
Such feedback from early adopters indicates a concern that overly complex setups may discourage new users. Joe Gershenson, the engineering lead for OpenAI’s harness, acknowledged that refining the effort settings for new users is essential, assuring, “We are continually working to improve how users find the right balance in reasoning.”
OpenAI faces additional hurdles in penetrating mainstream white-collar roles; many workflows lack the clear measurability inherent in code. Software either functions or doesn’t, highlighting an apparent gap in identifying the nuances of efficiency in business presentations, strategy, or sales tactics.
“Designing for a hugely versatile product like this comes with its unique challenges,” Ambrosino remarked. When I pressed for specifics on targeted workflows, engineers expressed hesitation, suggesting this inquiry would best suit OpenAI's research team.
Later, OpenAI clarified that they draw on their GDPval benchmark—which encompasses 44 occupations and various knowledge work assessments—and supplement this with user input. An unofficial insight is that feedback also comes directly from OpenAI employees. Ambrosino noted, “We must continually evaluate if we're developing workflows relevant to broader audiences or overly specialized based on our own experiences.”
As the user base grows, valuable insights derived from actual usage patterns could yield significant data, following in the footsteps of coding tool successes, assuming users opt to share their data for further training (a choice I did not make).
Regarding competitive dynamics, despite frequent discussions about interfaces, OpenAI's engineers appeared hesitant to comment on how their products stack up against Claude Cowork and similar offerings designed for mass audiences.
“The truth is, I don't focus on what others are developing,” Gershenson remarked. “I honestly do not think about our competitors at all.”
This skepticism is understandable considering the apparent similarities in user interfaces across platforms and the natural competitive intelligence intrinsic to business. Notably, the first action from ChatGPT Work was to prompt me to migrate my Claude Cowork data.
The intense competition could be traced partly to Claude Code's pioneering role in AI-driven programming, reshaping how software developers engage in their craft. This rivalry is heightened by the fact that OpenAI, which had conceptualized a similar tool first, did not execute it optimally.
In the initial rollout of Codex as a web application, engineers perhaps became overly confident about the model's capabilities to autonomously manage tasks with minimal user interaction. Conversely, Claude Code was designed for an interactive user experience, encouraging more thorough dialogue and feedback throughout the process.
Anthropic's strategy, although more labor-intensive for users, proved effective. “Our product was slightly ahead of its time, an ambitious approach too early for the model and harness we had available,” noted Ambrosino.
OpenAI has since adapted, integrating more opportunities for user interaction within the model. The current iteration of Codex, now available as desktop and mobile applications, appears to be gaining traction, showing increasing interest in April, with positive survey findings suggested for enterprise adoption.
Part of this momentum can be attributed to refining product-market fit and addressing safety concerns surrounding Anthropic’s models. OpenAI continues pushing toward more user-centric systems, which is exemplified in the development of ChatGPT Work. The engineers I spoke with emphasized their robust model and efficient pricing as the key differentiators.
“While it may be frustrating, often the core distinction lies in the model itself. We have prioritized leveraging our latest, most powerful, and cost-effective technologies within this application,” Ambrosino concluded.
This perspective aligns with the “bitter lesson” from AI research, underscoring that a superior, more generalized model is often more valuable than specialized experiences. For advocates of AI, the harness is perceived as a temporary support structure rather than a competitive barrier.
“Short-term gains can be had by including numerous features and conditions, but the reality is that the next wave of models will render them obsolete in a matter of months,” Gershenson remarked. His team aims for a straightforward design that empowers the model with the necessary context while avoiding unnecessary additions.
The pressing question remains whether the average user or model can navigate this complexity. Professor Ethan Mollick from the Wharton School of Business, who studies AI in workplace environments, views Claude as the more user-friendly choice, stating that “ChatGPT often aims for a seamless experience, attempting to handle everything for you, whereas Claude engages in a more interactive process with numerous comparison options.”
Sottiaux and the broader OpenAI team maintain a different perspective, asserting that their conversational interface is more effective than merely learning applications. “We are confident that the market is primed for this approach,” he remarked, vindicating the substantial user uptake.
Nevertheless, whether a model-specific harness is optimal for maximizing performance remains uncertain. Studies within organizations like Composio and Databricks indicate that various combinations of harnesses and models exhibit differing efficiencies in coding benchmarks. Evidence shows tools like Pi, developed by Earendil, outperforming Codex with the same underlying model.
Zechner, the creator of Pi, suggests that his minimalist harness successfully demonstrates the viability of an AGI-centric model, especially for coding tasks. He points out that while it lacks explicit features, its capacity to adapt and create interfaces compensates for the absence of traditional structures. He acknowledged the challenges OpenAI faces in appealing to a broader audience beyond technical professionals.
“All AI solutions are shaped by coding tasks because that is where training data is concentrated,” he contended. “For roles like management, decision impacts unfold over extended timelines, making it difficult to track progress through traditional feedback loops.”
Similar to other open-source avenues, he criticized large AI organizations for working to secure a user base through proprietary harness solutions, warning that reliance on their stack risks reducing them to mere model providers in competition with alternative sources.
Technical professionals have also noted that the insights gained from evaluating token usage and agent behavior in cutting-edge harnesses are significantly constrained. While this presents limited implications for non-technical users, successfully achieving widespread adoption will eventually necessitate deliberate discussions around operational costs.
For instance, casually exploring the tool via a $20 monthly subscription, I managed to utilize more than 80 million tokens in a span of just four days, which translated to over $65 in costs based on the model’s assessment (without a comprehensive usage dashboard within the app). This highlights a substantial subsidy far exceeding the subscription rate even during casual use.
“We are consistently striving to enhance efficiency,” Sottiaux said, referencing an upcoming 80% cost reduction for users leveraging OpenAI's Luna model. “If you check back six months from now, we anticipate you will be able to achieve the same functionalities at significantly lower costs.”
An additional pertinent inquiry involves whether the configuration and dependence on applications generates a lock-in effect through data retention or the meticulous nature of configuring access permissions across plugins.
During my visit to OpenAI's headquarters, marked by a calm yet slightly tense atmosphere, the urgency among the engineers was palpable; there’s considerable work to accomplish. As we spoke, Nathan, who oversees the product engineering team, highlighted that the objective remains the promise of a magical, all-in-one solution, while also acknowledging that the complexity still exceeds expectations. “I remain hopeful that we can resolve these challenges, ensuring a truly AI-native experience.”
