Electronic signatures and intelligent agreement management
Docusign
Docusign makes software for creating, signing and managing business agreements. Its Intelligent Agreement Management platform builds on its eSignature business, extending from collecting signatures to automating contract processes and putting agreement data to work. eSignature lets teams send, sign and track documents; Contract Lifecycle Management automates the contract lifecycle; and Agreement Manager helps users find, analyze and act on agreements. These tools serve departments including sales, legal, procurement and HR.
Agreement intelligence depends on extracting the business terms inside documents. Contract tables make that difficult: merged cells, nested layouts, inconsistent formatting and tables spanning multiple pages can obscure the relationships between terms and values. Docusign’s contract table extraction work addresses practical questions such as which notification obligations apply after a service outage, what hourly rate a contractor agreed to, or how pricing provisions across exhibits affect a vendor renewal. Turning these tables into structured, queryable data helps teams retrieve terms that would otherwise require manual review.
Docusign integrates NVIDIA Nemotron Parse, an open vision-language model built for document understanding, into its layout detection and optical character recognition pipeline. The model uses document structure and semantics to interpret and reconstruct complex tables, and Docusign serves it through vLLM within its own environment, keeping sensitive agreement data there. The company tested the integration against real enterprise contracts to capture formatting variations, inconsistent table structures and mixed-language content. Its retained announcement describes table extraction in Agreement Manager as accepting beta customers, with general availability and a dedicated public API still forthcoming; it does not report a numerical accuracy result.
Docusign serves both small businesses and enterprises. Its company website reports 1.9 million-plus paying customers, use in more than 180 countries and over 1,000 pre-built integrations. Alongside those integrations, its existing APIs let developers incorporate Docusign technologies into other applications, connecting agreement workflows with the systems businesses already use.
Explore the recordings
The supplied Docusign archive contains one recording, co-presented with NVIDIA. It offers several paths into agreement data extraction: the business problem, a demonstrated workflow, and the engineering tradeoffs. Product capabilities and performance figures below describe what the speakers presented in the recording; they are not verification of today’s product status.
Start with the agreement-data problem and workflow
Watch Your Agreements Are a Database You Can't Query for Hiral Shah and Sean Sodha’s explanation of why agreement data—especially pricing tiers and tables—can be difficult to extract reliably. The Agreement Manager demo follows an order form through upload, metadata and pricing extraction, CSV export, and API access. This is the useful starting path for viewers interested in turning documents into queryable organizational data.
Follow the extraction architecture and its limits
Use the same recording to examine the presenters’ single-pass vision-language extraction approach, including layout, reading order, and preserved table structure. The Q&A adds an important qualification: Docusign described a hybrid pipeline that retained OCR for other fields and clause text. Their distinction between bulk preprocessing and low-latency document Q&A helps frame the throughput-versus-interactivity tradeoff.
Evaluate the reported results and future work
Return to the recording’s performance discussion for Shah’s report of roughly 20× greater table-extraction throughput than the other open-source models tested. Full comparison conditions were not supplied in the transcript, so treat this as a recorded result rather than a general benchmark. Retrieval, agents, quantization, and multi-token generation are discussed as future work in the recording; it does not establish their availability today.
1 talk
Newest first
1 speaker at AIE
Affiliations reflect their AIE appearances, not necessarily current employment.
