Contract intake agent

What it does

Customer contracts contain a variety of data, including customer details, billing contacts, pricing details, and billing terms. Instead of manually transcribing a contract into Sequence for every new customer, contract intake agent automatically extracts the key information and sets up the customer and billing schedule for review.

How it works

Today, we parse the document then match the extracted data against existing entities (e.g., customers) before generating the schedule. We manually manage and inject the relevant context. The output flows into Watchtower for human review before any changes are applied.

What we’re trying next

We're evaluating Gemini 3.5 Flash for higher-accuracy parsing; moving our context management to tool calls; and allowing customers to configure a system prompt so the agent can better handle bespoke terms (or remember this automatically based on the user’s changes during the review step).

What we tried

OCR preprocessing

We expected OCR would reduce hallucinations and lower latency & cost (if we only need to feed the OCR output into the LLM, instead of the PDF data). But it did not - the OCR-only output failed in some of our tests because the layout is crucial in some contracts. Although we primarily tested markdown output - a richer format might be more successful (HTML is in vogue right now).

Models

We ran evals across SOTA models and OCR + LLM hybrid pipelines. Several models had high accuracy, but Gemini won on the combination of accuracy, latency, and cost.

Per-field confidence UI

Our first attempt pre-populated the existing billing schedule editor and showed confidence indicators on every field. We scrapped it because there was too much custom UI to squeeze into the existing editor, a bit of information overload, and the confidence values weren’t particularly helpful - operators still wanted to review every field.