Walk into any 30-person professional services firm and ask where the contracts live. The answer will be a SharePoint folder, a Dropbox tree, a network share, or some combination of all three. Now ask how the team finds the right contract on a given day. The answer is a person who remembers approximately where it is, plus a Ctrl-F search through filenames.
That is the operating model for thousands of SMBs. The corpus is enormous, the institutional knowledge to navigate it is in two or three people's heads, and the cost of a misfiled or unfindable document is absorbed silently into operations. Document intelligence changes the equation. Each PDF, scan, and Word doc becomes a structured record with extracted fields, classification tags, and references to other documents. The corpus becomes a database.
What Document Intelligence Extracts
The extraction layer pulls structured fields from unstructured documents. The fields depend on document type, but the pattern is consistent across verticals.
Identity fields. Names, addresses, entity types, signatories, witnesses, notaries. The basic "who" of any document.
Dates and deadlines. Effective dates, termination dates, renewal dates, statute of limitations, filing deadlines, project milestones. The basic "when."
Dollar amounts. Contract values, invoice amounts, retainer figures, deductibles, change orders, escalation clauses. The basic "how much."
Cross-references. "This contract supersedes Master Services Agreement dated March 4, 2024." "See Schedule B for pricing." "Subject to compliance with the rider attached." Cross-references are the connective tissue of document corpora, and they are the part that traditional OCR misses entirely.
The shift. A static PDF folder becomes a queryable record set. "Show me every active contract with auto-renewal in the next 60 days" becomes a 200ms query rather than a three-day audit project.
How We Build It
The stack is mature in 2026 and the pieces compose well. Five components do most of the work.
Cloudflare R2 for storage. Original documents live in R2 with retention policies, per-user access controls, and edge replication. R2 is cheap, fast, and integrates natively with the Cloudflare Workers running the extraction logic.
Anthropic Claude or GPT-4o for vision and extraction. The vision-capable models read scanned documents directly. We call them with structured output schemas matched to document type. A contract extraction returns one schema. An invoice extraction returns a different schema. A plan set extraction returns a third. The model handles OCR, layout interpretation, and field extraction in one pass.
Structured output schema per document type. Each document class has a defined schema. We do not ask the model to "extract all the data." We ask for the 12 fields we know we need, with type validation on each. The schema makes the output reliable and the downstream record predictable.
Routing to the matter, project, or account record. Every extracted document is associated with a parent record in the system. A new permit goes to the property record. A new contract goes to the matter. A new invoice goes to the project. The association is part of the extraction step.
Audit logging on every extraction. Each extraction is logged with the source file, the model used, the timestamp, the confidence score, and the user who triggered it. The audit trail is queryable and tied to the parent record.
What Changes Downstream
Once the corpus is structured, three things change permanently in how the team operates.
Search becomes structured. "Show me all contracts with David Lee as signatory between 2022 and 2024" replaces "go ask Karen, she remembers David Lee." The senior staff who used to be the institutional memory still know things, but they are no longer the only path to the answer.
Reports become possible. "What is our auto-renewal exposure for the next 90 days?" "How many of our active matters have minor children involved?" "Which projects have permits expiring this quarter?" Questions that used to take a week of audit work become real-time dashboard cards.
Compliance becomes continuous. Compliance audits used to be a once-a-year scramble through document folders. Continuous monitoring queries the structured corpus daily and flags exceptions. The shift from "audit when forced to" to "audit continuously" is the single most valuable downstream change.
Reference Workflows
Two of our flagship builds use document intelligence as a core layer.
For California's largest hillside structural engineering firm, we built a plan-set and permit indexer that ties every document to a property record. Senior engineers can now ask "show me every retaining wall job we have done within 500 feet of this address" and get a referenced answer with thumbnails of the historical plans. Read the full breakdown at the structural engineering case study.
For an Encino-based estate and family law firm serving high net worth families, we are building intake form extraction that pulls trust schedules, beneficiary lists, asset inventories, and family member identification into structured records. The attorney's first 20 minutes of every new matter used to be data entry. It is now review. Read the build details at the law firm employee dashboard case study.
Compliance and Audit
For regulated industries and high-trust client segments, document intelligence has to clear three compliance bars at once. Per-document retention, per-user access, and audit logging on every extraction.
Per-document retention means each document type has its own retention rule. A signed engagement letter retains for the life of the matter plus seven years. An expired draft retains for 90 days. The system enforces these automatically.
Per-user access means the structured record inherits the source document's access posture. A document the firm shared only with a specific partner is queryable only by that partner. Extraction does not leak data.
Audit logging means every extraction call is recorded. If a regulator asks "who saw this document and when," the answer is in the log. If a client asks "what data did your AI process about my matter," the answer is in the log. ISO 42001-aligned controls are built in from the start.
Document intelligence is one of the highest-leverage AI workflows we deploy, because the corpus already exists, the value is immediate, and the compliance posture can be tighter than the manual operating model. If your team is searching by filename in 2026, the upgrade is overdue.