General-purpose language models can summarize contracts reasonably well, but they struggle when asked to calculate what a company owes a supplier after accounting for three amendments, a revised pricing schedule, and internal procurement policies, according to a feature on Docusign's contract AI capabilities published by The New Stack. The report examines how Docusign is addressing this challenge through Docusign Iris, the contract AI engine at the center of its Intelligent Agreement Management platform, which launched in April 2024 and became available to U.S. customers that May. The system combines contract management expertise, agreement-specific data, and AI models to help teams extract insights and act across contract workflows.

Docusign CTO Sagnik Nandy identifies four categories of context that differentiate the platform from generic solutions: within-agreement, cross-agreement, organizational, and user context. The first involves following a term defined on one page and referenced again later in the same contract. The second becomes harder when that term appears across multiple contracts and amendments, because to a language model, an agreement that's gone through two or three amendments looks like three or four independent PDF files rather than a single contractual relationship. The third category encompasses company-specific policies, such as legal playbooks spelling out acceptable terms and procurement rules determining approval authority. The fourth includes someone's language preferences, previous questions, and feedback.

According to Nandy, general-purpose LLMs may handle much of the within-agreement context, but the other three categories prove considerably harder. He explains that answering a question about payment obligations to a particular supplier involves several steps: Iris needs to recognize the supplier as a company, locate its agreements, account for amendments, and establish how the obligations have changed. "While it seems like a straightforward question, to make it truly happen, you need rich search functionality. You need a repository," Nandy says, estimating that such a question could involve five or six tools drawing on Docusign's contract-specific data. The report notes that Iris combines intelligent extractions, semantic search, an agentic layer, the ability to build agents, and rich context rather than relying on a language model to do everything.

The platform addresses automation limits by combining agentic capabilities with business rules and approval processes, letting organizations decide which parts of a contract workflow to automate and where people remain in control, the report explains. Renewals below a certain spending threshold might follow standard procedures handled by an Iris-powered agent, while more expensive or sensitive agreements would go to a person for review. For developers building enterprise contract AI agents, the system needs reliable contract data, a way to connect related documents, an understanding of company rules, and clear limits on what it can do without approval. The report cites Experian's results connecting Docusign with its configure-price-quote solution: contract cycle times dropped from roughly 10 days to hours, though those improvements aren't necessarily attributable to autonomous agents. Reading a contract is relatively straightforward, but knowing what it means for the business is where the real engineering work begins. Organizations evaluating contract AI will need to weigh the upfront investment in configuring approval workflows and business rules against the potential for faster cycle times, while accepting that the most complex negotiations will still require human judgment at critical decision points.