AI contract review SaaS on Make.com: 1,000-page contracts read with citations, 94.9% accuracy

A property-law specialist in Australia wanted contract review that a lawyer can trust, which means every answer has to point at the clause it came from. No citation, no answer.

We built the product as a multi-tenant SaaS: a Next.js front end on Supabase, with Make.com as the entire backend. A firm uploads a property contract of 100 to 1,000 pages. Make.com splits it, drives the AI extraction and drafting, and returns a finished review document in the app and by email, every fact carrying a page reference. Each firm has its own quota, Stripe billing and an admin re-run tool.

The app itself is a proper product, not a form: a reviews table grouped by type with date filters and sort backed by the database, per-review run history with re-run, DOCX and PDF export, in-app or email delivery, a monthly review quota per firm and a file retention date on every submission. Front end on Vercel, database on Supabase, backend on Make.com.

The hard part was not the AI, it was long documents and rate limits. A 100 to 1,000 page contract cannot go through a model in one call, so the pipeline chunks, queues and resumes, honours Retry-After headers and uses idempotency keys so a retried chunk never writes twice. We built a golden-test set from real contracts and score every prompt version against it, with a five-category failure taxonomy.

Result: out-of-sample accuracy rose from a 70.2% mean to 79.5% between versions, with a best run of 94.9%, and a latent 1,000-page context bug was caught before it reached a client.

The full case study, with the product screens, is on my profile.

Prem Patel, Nex Automations. Make.com Level 5 certified, official Make partner, Zapier Certified Expert.