From Lease PDF to Audit-Ready in 15 Minutes
A controller at a major restaurant chain (8,800 locations, thousands of leases) watched our AI workspace generate a roll-forward analysis in real time during a live demo. He had used two of the largest lease accounting platforms on the market. His reaction:
"That workspace is one of the best things I've ever seen in a lease accounting software."
He was watching the system do in minutes what his team does manually every month. This post walks through the entire pipeline: from a raw lease PDF to an audit-ready deliverable.
The full pipeline
Seven stages, each building on the last. A lease enters as a PDF and exits as a reconciled, multi-tab Excel workbook ready for KPMG.
Every stage runs the same code paths as the REST API. The AI agent is not a chatbot wrapper over a separate system. It calls the same services, with the same RBAC enforcement, the same decimal.js precision, the same balance assertions. The accountant becomes the reviewer, not the execution engine.
The workspace
The design principle is simple: say what you need, get the outcome. No report templates. No parameter forms. No export-to-Excel-then-reshape workflow.
The agent runs an agentic loop: Claude with 40+ domain tools, up to 25 iterations per request. Prompt caching cuts input token cost ~90% on longer sessions. Context compaction preserves exact financial figures across turns. Every tool call is visible in real-time via SSE streaming.
The controller's team used to spend 2 hours with 2 people on every audit request. With the workspace, one person asks a question and downloads the answer in 15 minutes.
The math
Scale that across a quarter of audit requests and the numbers become dramatic.
Traditional implementation costs compound the problem further. At $75-$150 per lease for manual abstraction, enterprise implementations cost $200K-$500K and take 6-12 months. Our extraction pipeline processes a 500-lease portfolio in 2-4 weeks at a fraction of the cost.
The extraction pipeline uses tiered models (Haiku first, Sonnet fallback below 0.85 confidence), a 497-line cross-field validator, and confidence-based routing: 0.95 and above auto-accept (~85% of fields), 0.80–0.95 quick review, below 0.80 full review. Five always-review fields require human sign-off regardless of confidence.
Classification runs the ASC 842 five-test framework in under 10 seconds per lease. Schedule generation computes full amortization in under 3 seconds per lease. Journal entries are deterministic from schedule data with balance assertion, hard failing if debits and credits differ by more than $0.02. All entries created as DRAFT for maker-checker workflow.
"There's no software that does this, to my knowledge. My team generates this manually every month."
The pipeline turns that manual process into a conversation. The auditor gets the same deliverable, but it takes minutes instead of hours, and the math is penny-perfect.
See Document Intelligence and Intelligence to learn more about these capabilities.