In the past, when rewriting legacy software was not an option—or when external dependencies made it impossible—we turned to RPA (Robotic Process Automation) tools to automate manual processes. We used tools like UIPath whenever the work was deterministic and predictable.
With GenAI, we can now handle much more ambiguity in both input and output. For example, we can process natural language and make decisions that once required human intervention. This approach is often called generative process automation (GPA), or intelligent RPA.
Consider a simple routing example: "Does this customer complaint require a human agent?" Traditional RPA and classical NLP would escalate the ticket. A large language model, however, can now resolve it with just a few carefully crafted instructions (a system prompt). The development effort is still comparable to classic RPA. Both RPA and GPA can benefit from advancements in generative engineering practices though.
On the horizon, we see computer use agents that receive a high-level goal. These agents log in with your credentials and complete the task end-to-end. They leverage your desktop as an interface to the world, using the mouse, keyboard, and screen as input and output. The outcome is familiar—automated work—but you do not need to explicitly configure a workflow. Instead, the agent interprets your intent and acts accordingly, reducing the need for manual setup.
The idea gained traction when Anthropic launched computer use. OpenAI now exposes Operator to Pro subscribers.
Today, the experience is far from ready for prime time. It is slow and expensive. Reliability is also an issue. However, the period during which agents can operate autonomously is expanding quickly.
Enterprise security remains a major concern. Most computer use products run in isolated sandbox browsers, far removed from the systems used in most workplaces.
I expect the field to evolve rapidly, which will create new opportunities and challenges for organizations managing legacy systems.
The impact of these technologies can be summarized as follows:I do not expect one paradigm to dominate. Rather, I anticipate that we will do more of everything, continually reducing the need for manual effort.