# Your Agreements Are a Database You Can't Query

Hiral Shah, Docusign & Sean Sodha, NVIDIA | AI Engineer World's Fair 2026 | 18:22

Source: https://www.youtube.com/watch?v=_gvamfT8H-w
Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks.
Page: https://aietalks.com/talks/your-agreements-are-a-database-you-cant-query
Published: 2026-09-16
Tags: cost, latency, multimodal, rag

## TL;DR
- Docusign processes about a million agreements a day, but much of the information in them remains difficult to search and use.
- Pricing tables are difficult for generic document extraction because line-by-line reading loses merged cells, nested columns, and table structure.
- A purpose-built model with roughly 900 million parameters extracted tables about 20 times faster than the open-source models Docusign and NVIDIA tested.

## Summary
Hiral Shah and Sean Sodha describe agreement data as a large enterprise engineering problem. Docusign has 1.9 million paying customers, a billion users, and processes about a million agreements each day. Those agreements contain pricing tiers, SKUs, rate cards, service levels, and other terms that businesses often cannot query without manual review. The hardest material is usually in tables, where generic extraction systems lose merged cells and nested structure. Docusign and NVIDIA built a small vision-language model for this task. The model returns text, reading order, semantic formatting, layout, and preserved tables in one pass. In a demo, an order form becomes structured pricing data that can be downloaded as CSV or accessed through an API. The speakers argue that purpose-built models can be faster and cheaper at enterprise scale. They also explain why OCR and pre-processing still matter when organizations need to process large document collections before users ask questions.

## Key ideas
### Agreement data is hard to use at enterprise scale
[00:01](https://www.youtube.com/watch?v=_gvamfT8H-w&t=1s)
Hiral Shah describes agreements as a source of important business data that is stored in PDFs, PNGs, and other unstructured formats. Docusign has 1.9 million paying customers, a billion users, and processes about a million agreements each day. The work is to make that information structured, readable, and queryable. A business may want to know what it contracted for with a vendor, but the answer can be buried in one agreement or spread across agreements from 10 or 20 years of business history.

### Businesses leave negotiated value inside agreements
[02:00](https://www.youtube.com/watch?v=_gvamfT8H-w&t=120s)
Shah cites a Docusign study with Deloitte that estimated $2 trillion in negotiated value is captured in agreements but not recovered by organizations. He attributes this to human reading, human review, disconnected systems, and manual workflows. Docusign's intelligent agreement management approach covers the agreement lifecycle, from creation and negotiation through signing, storage, and analysis. The aim is to extract terms and send them into downstream systems where legal, procurement, sales, and business teams can use them.

### Tables break document extraction systems
[03:59](https://www.youtube.com/watch?v=_gvamfT8H-w&t=239s)
The terms people most need, including pricing tiers, SKUs, SLAs, and rate cards, often appear in tables. Shah says generic vision-language models and traditional document extraction tools read pages line by line, which damages table meaning. Merged cells, merged columns, merged rows, nested tables, and cells without clear boundaries create problems. The result is extra work for legal, procurement, and sales teams that may spend hours locating a basic term.

### The NeMo Retriever work targets document retrieval and extraction
[05:12](https://www.youtube.com/watch?v=_gvamfT8H-w&t=312s)
Sean Sodha explains that NVIDIA's NeMo Retriever initiative includes embedding models, reranking models, and document extraction models. Its retrieval workflow first finds the most relevant documents for a query, then finds the needed information inside those documents. The work with Docusign focused on NeMo Parse, a model built for complicated table extraction. Sodha describes the broader initiative as an open-source effort that publishes model weights, datasets, techniques, and blueprints.

### NeMo Parse is an extractor with one document-processing pass
[06:59](https://www.youtube.com/watch?v=_gvamfT8H-w&t=419s)
Sodha says NeMo Parse is a small vision-language model with about 850 to 900 million parameters. It is designed as an all-in-one model rather than a collection of separate models for tables, page elements, and other tasks. A document goes in, and the output includes semantic formatting, layout, text, reading order, and preserved table structure. He distinguishes it from a generator: the model is intended to extract information from documents.

### The demo turns an order form into usable pricing data
[08:32](https://www.youtube.com/watch?v=_gvamfT8H-w&t=512s)
In the Agreement Manager demo, Hiral Shah uploads an order form and shows the system processing it with AI. The application extracts metadata such as key terms and commercial details, then links those details back to the relevant document sections. The NVIDIA model extracts structured pricing data from complex tables. The result can be downloaded as a CSV for finance or procurement teams and is also available through an API.

### Purpose-built models reduce latency and cost at scale
[11:27](https://www.youtube.com/watch?v=_gvamfT8H-w&t=687s)
Shah says the partnership tested different models for different purposes and found that a purpose-built model helped them bring the work to market faster. Smaller context also reduced latency and cost for their scale. In their comparison, NeMo Parse extracted tables 20 times faster than the other open-source models they tested. The point is practical: a smaller model designed for one task can fit an enterprise pipeline better than a general model.

### Large document collections still need pre-processing and OCR
[13:44](https://www.youtube.com/watch?v=_gvamfT8H-w&t=824s)
In the Q&A, Shah and Sodha explain that OCR is not disappearing from every workflow. When a company has thousands or more documents, pre-processing and identifying document content up front can reduce later query latency. Procurement teams may also need extracted pricing data in systems such as Coupa, which requires batch processing rather than document-by-document agent calls. The speakers describe a hybrid architecture that uses layout-aware extraction for tables and OCR for other fields, metadata, and general text.

## Notable quotes
- Hiral Shah: "We process a million agreements a day that needs to now structure, make it readable, make it queryable, usable." (02:00)
- Sean Sodha: "It's not a generator. It's more of an extractor at the end of the day." (07:55)
- Hiral Shah: "For us, the lower context basically also meant lower latency, lower cost to deliver the scale that we are talking about." (11:48)
- Hiral Shah: "NeMo was 20x faster, which helps us when we're talking about the millions and billions of scale that we're serving for all of our customers." (12:08)

## Tools & references mentioned
- Docusign
- NVIDIA
- NeMo Retriever
- NeMo Parse
- Deloitte
- Coupa
- NVIDIA NIM
- OCR

## Who should watch
- You are building search, extraction, or downstream workflows over a large contract repository and need to decide what should be pre-processed.
- Your documents contain pricing tables, rate cards, SKUs, service levels, merged cells, or nested tables that generic parsers damage.
- You are comparing a general-purpose vision-language model with a smaller model built for one extraction task.

## Related talks

- [Building AI Agents that Actually Automate Knowledge Work](https://aietalks.com/talks/building-ai-agents-that-actually-automate-knowledge-work) (Jerry Liu, LlamaIndex, 17:57)
- [The Billable Hour is Dead; Long Live the Billable Hour](https://aietalks.com/talks/the-billable-hour-is-dead-long-live-the-billable-hour) (Kevin Madura & Mo Bhasin, AlixPartners, 17:04)
- [Building an Agentic Platform](https://aietalks.com/talks/building-an-agentic-platform) (Ben Kus, Box, 19:06)
- [Grep or Embeddings? Agentic Search Over Company Documents](https://aietalks.com/talks/grep-or-embeddings-agentic-search-over-company-documents) (George He, LlamaIndex, 23:22)
- [Design like Karpathy is watching](https://aietalks.com/talks/design-like-karpathy-is-watching) (Zeke Sikelianos, Replicate, 19:26)