What happened
OpenAI is working with legal tech company Ironclad to train AI agents on complex contract workflows. In a new post, OpenAI detailed how these agents can operate software interfaces to perform multi-step professional tasks. The demonstration shows an agent answering questions about a contract by navigating Ironclad's user interface, finding the right document, and extracting the relevant information. It also performs tasks like running a redline comparison between two documents. The stated goal is to build agents that can reliably handle specialised enterprise work.
How the room's reading it
The agent-building community sees this as a notable step up from simpler demos. For months, developers on X have been debating the readiness of agents for real enterprise work, with many pointing to a lack of reliability on complex, multi-step tasks. The Ironclad collaboration is being read as a serious proof point — a frontier lab tackling a messy, high-stakes workflow instead of a toy problem. Some practitioners remain sceptical, pointing out that reliability in legal tech needs to be near-perfect. The consensus, however, is that this provides a concrete blueprint for building agents that interact with existing enterprise software.
Sailfish's take
This isn't about the model's capabilities — it's about the training methodology. The hard part of building useful agents isn't the model, it's instrumenting real-world software and capturing high-quality data from complex human workflows. This is where most agentic products fail. We think the next wave of valuable AI companies will look more like workflow specialists than general-purpose agent platforms. This is where we'd focus our efforts — identifying and instrumenting the high-value, repetitive tasks inside enterprise SaaS. The model is a commodity; the workflow is the moat. If you're building a general agent, you're chasing the wrong problem.