Docusign cuts AI costs 90% with smaller, smarter models
By matching AI models to specific tasks instead of using expensive general-purpose ones, Docusign processes 1 million contracts daily at a fraction of the cost.
In brief
By matching AI models to specific tasks instead of using expensive general-purpose ones, Docusign processes 1 million contracts daily at a fraction of the cost. Task-specific AI models trained for narrow jobs beat expensive general-purpose models on both cost and accuracy, and the savings unlock new products…
Docusign's AI bill became unsustainable
Processing over 1 million contracts daily with large general-purpose AI models created massive costs and delays. The company needed a smarter approach to stay viable.
Small models handle routine work
Docusign trained smaller AI models on Microsoft Foundry to extract the same fields from millions of documents. These cheap models now handle routine extraction work that doesn't need advanced reasoning.
Frontier models reserved for judgment calls
Docusign now saves expensive frontier models only for complex reasoning tasks like untangling clauses or comparing terms across documents. This strategic allocation cuts costs dramatically.
“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.”
Filtering contracts before processing
Docusign added a filter that identifies relevant passages and sends only those to AI models, not entire contracts. This trimmed input actually produced better answers than full documents sent to larger models.
Results: 90% cost cut, 8x throughput
The small model approach cut AI processing costs by 90% while raising throughput eightfold. Accuracy stayed within 2 percentage points of the larger models, and per-document costs fell 50-fold overall.
Cheap extraction funds new agents
Docusign's May launch of Iris assistant and AI agents for sales, HR, procurement and legal became affordable because they run on data from cheap extraction models. Lower extraction costs mean lower agent costs.
Real-world payoff at SEI
Management consulting firm SEI uses the system to extract obligations from master service agreements and connect them to Workday and HubSpot, so new work orders start with legal terms already loaded.
Enterprise AI ROI still elusive
Despite 81% to 95% of large U.S. enterprises deploying AI, only 5% to 10% say investments have fully paid back. Most expect five to six years to break even, yet all plan to raise AI budgets next year.
The essence
Task-specific AI models trained for narrow jobs beat expensive general-purpose models on both cost and accuracy, and the savings unlock new products.
