OpenAI wants to bring the capabilities of coding agents to a much wider audience. According to Thibault Sottiaux, the company’s head of core products, ChatGPT Work represents an important step in that direction: a platform designed to help professionals delegate complex tasks to AI, without requiring technical expertise.

Sottiaux oversees a broad portfolio that includes OpenAI’s API, agent infrastructure, enterprise products, ChatGPT and Codex. In a recent interview, he discussed the thinking behind ChatGPT Work, the growing adoption of AI agents and the challenge of turning increasingly capable models into products that feel simple, natural and safe.

From coding agents to everyday professional work

Codex was originally built for a technically experienced audience. Developers could tolerate early limitations, experiment with new workflows and understand when human review was necessary.

ChatGPT Work targets a much broader group. Its purpose is to make similar agentic capabilities available to professionals who may need to analyse documents, conduct research, prepare reports or build presentations, but who do not necessarily know how AI systems work behind the interface.

For OpenAI, the challenge is therefore larger than improving model performance. The company must translate advanced capabilities into an experience that can be used intuitively on the web, on mobile devices and across different professional contexts.

Sottiaux described the objective as creating a product that is both simple and powerful. Instead of asking people to learn complicated software, OpenAI wants the technology to adapt to the way humans already communicate and work.

Voice interaction is one example of this direction. As conversations with AI become more natural, the interface itself may become less visible. Users will increasingly describe what they need, provide the necessary context and allow the system to determine how the work should be completed.

Are professionals ready to delegate entire tasks?

The transition from AI assistants to AI agents changes the nature of the interaction.

Traditional assistants help users complete individual steps. Agents can potentially manage an entire workflow, make intermediate decisions and produce a finished result with less direct supervision.

This creates an important product-design question: when should the AI take the lead, and when should the interface stop and ask the user to make a decision?

Some AI products expose more choices, checkpoints and alternative paths. OpenAI appears to favour a more seamless experience in which the model handles much of the complexity behind the scenes.

Sottiaux argues that adoption indicates that users are ready for this approach. He said ChatGPT Work had reached 20 million users, suggesting significant demand for tools that can execute professional tasks rather than simply answer questions.

However, adoption alone does not eliminate the need for control. In business environments, autonomy must remain proportional to the consequences of a mistake. Drafting an internal summary presents a very different risk from sending a client email, modifying company data or making a financial recommendation.

The most useful agentic systems will therefore need to balance convenience with transparency, permissions, review mechanisms and clear escalation points.

Product development as a process of discovery

Building a platform intended to support many types of professional work does not follow the traditional model of identifying one narrow problem and designing a fixed solution around it.

According to Sottiaux, developing these products also involves discovering what new models can do. As capabilities improve, OpenAI observes where models perform particularly well and then builds product experiences around those strengths.

He pointed to GPT-5.6 as an advance in general professional work, including processing large collections of documents, producing reports and presentations, and conducting in-depth research.

The company then uses feedback from real-world adoption to refine those experiences. This creates a continuous cycle: models unlock new capabilities, products expose them to users, usage reveals practical strengths and weaknesses, and the resulting insights guide further development.

This iterative approach can accelerate innovation, but it also means that organisations adopting AI must prepare for rapidly changing tools. Workflows designed around today’s limitations may need to be reconsidered as models become more capable.

The cost of intelligence is expected to fall

The economics of AI agents remain an important concern, particularly for organisations considering large-scale adoption.

Agentic systems may consume significant computational resources because they often perform multiple steps, process extensive context and use tools before producing a result. This can create a visible gap between the price paid by individual users and the underlying resources consumed.

Sottiaux expects efficiency improvements to reduce this pressure. OpenAI’s goal is to make a given level of capability progressively cheaper, allowing users to obtain more utility for the same amount of money.

For business leaders, however, the relevant calculation extends beyond token prices or subscriptions. The real question is whether an AI system reduces the total cost of completing a workflow while maintaining the required standards of accuracy, security and accountability.

A less expensive model does not automatically create a more efficient operation. Businesses must also consider implementation, integration, human review, governance and the cost of correcting unreliable outputs.

Trust remains the critical barrier

ChatGPT Work becomes considerably more useful when it can access the systems where professional activity already happens. That may include email, messages, documents, calendars and internal applications.

The same access also increases the potential impact of errors, manipulation or unauthorised actions.

Sottiaux emphasised OpenAI’s investment in model alignment, safety systems and transparent evaluation. These measures are essential, but organisations still need safeguards at the product and infrastructure levels.

Before giving an AI agent access to sensitive business systems, companies should define:

What information the agent can read
Which actions it can perform independently
Which decisions require human approval
How activity is logged and audited
What happens when the system is uncertain
How access can be suspended or revoked

Trust in workplace AI will depend not only on model intelligence, but also on how clearly organisations can understand and control its behaviour.

What this means for businesses

The development of ChatGPT Work reflects a broader transition in enterprise software. AI is moving from a feature that assists with isolated tasks to an operational layer capable of coordinating complete workflows.

For companies, the opportunity is substantial. AI agents may reduce repetitive work, accelerate research, support decision-making and allow teams to complete complex assignments faster.

The strategic challenge is choosing the right workflows and implementing the appropriate level of oversight. Organisations should start with tasks that are valuable, measurable and reversible, then expand autonomy as reliability becomes clearer.

The future of professional software may indeed feel simpler to the user. Behind that simplicity, however, businesses will still need thoughtful architecture, secure integrations, reliable data and clearly defined human responsibility.

AI agents can make work easier. Building systems that organisations can safely depend on remains the harder and more important task.

Source