Docusign Processes 1 Million Contracts a Day on Small AI Models

Even digital signatures and contracts have a need for artificial intelligence these days.

Case in point: Docusign. It ran large general-purpose AI models on every contract it processed, and at more than 1 million agreements a day the cost and delays became a problem. The company swapped them for smaller models trained for specific jobs and cut AI processing costs 90% while raising throughput as much as eightfold, Microsoft said Tuesday (Sept. 15) in a customer case study. Accuracy stayed within 2 percentage points of the larger models.

Docusign’s AI engine, Iris, reads each agreement and pulls out more than 50 facts a business needs to act on, including contract value, governing law and renewal dates. Large customers arrive with backlogs of tens of thousands of legacy contracts scattered across 20 or more systems.

Sending each full contract to a frontier model worked in a demo. At production volume the per-call cost became the dominant line in the infrastructure budget. Ramachandra Kota, Docusign’s senior director of applied science, said in the Microsoft case study: “At a million documents a day, the difference between sending a full 100-page contract and sending the 4,000 tokens that actually matter is the difference between a viable business and an economics problem.”

Per-token prices have fallen about 98% since 2022 while enterprise AI bills have risen an estimated 320% over the same period, PYMNTS reported in June, because usage has grown faster than unit prices have dropped.

Cheap Models for Volume, Big Models for Judgment

Docusign’s fix was to stop treating every task as if it needed the most capable model. Routine extraction, the same fields pulled the same way across millions of documents, now goes to small models the company trained itself on Microsoft Foundry. Frontier models are held back for work that calls for reasoning, such as untangling a complex clause or comparing terms across several documents.

The company also stopped sending whole contracts to the model. A filter finds the passages likely to hold the answer and sends only those. On several types of extraction, Docusign said, the trimmed input produced better answers than the full document sent to a bigger model, because the boilerplate that buries the relevant clause is gone.

The small models beat the frontier model they replaced in production by 7 percentage points of accuracy, and Docusign retired a cleanup step that had been patching the old model’s mistakes. Across the whole pipeline, per-document cost fell 50-fold versus the earlier system, the company said in a June blog post.

The Savings Fund the Agents

The cheap extraction layer is what makes Docusign’s next product line affordable. The company in May unveiled an Iris assistant, AI agents and an Agent Studio in May to automate contract work across sales, HR, procurement and legal. Every one of those agents acts on data the extraction models produce first. If extraction still cost pennies a document, the agents would inherit that bill.

SEI, a management consulting firm, uses the system to pull obligations from its master service agreements and connect them to Workday and HubSpot, so new work orders start with the client’s legal terms already in view, SEI CEO Bill Gallagher said in the Microsoft case study.

According to PYMNTS Intelligence, between 81% and 95% of large U.S. enterprises have broadly deployed or embedded new AI in data and technology functions, but only 5% to 10% say the investment has fully paid back. At least half in every industry surveyed put full payback at five to six years out. At least 8 in 10 plan to raise AI budgets next year regardless of this, and none plan a cut, according to the Enterprise AI Benchmark August report.