The Rexera Case: Automating a Document Workflow, Not Just Adding RAG

The case: automating a transaction, not just a document

Rexera is a US technology startup founded in 2020 to address repetitive operations in residential real-estate transactions. The job is larger than recognizing a PDF. A workflow may have to identify counterparties, request documents from homeowner associations and municipalities, follow up, extract conditions and fees, reconcile facts, flag risk, and return an evidence-backed result to a closing specialist.

In an AWS case study, the company reports processing five million pages and more than 5,000 transactions per month. Reported customer outcomes include a 99% reduction in manual document-review workload, an average of four hours saved per transaction, and a 25% reduction in operating costs. A separate RAG tool reportedly cut manual review time by 80%. Rexera also claims more than 95% accuracy on straightforward cases when multiple models are used.

These figures were supplied by project participants and are not an independent audit. “99% less manual workload” does not mean that 99% of documents can be accepted without review: the metric may combine automated discovery, sorting, extraction, communication, and preparation. The useful lesson for a manager is the decomposition of the workflow, not the promotional percentage.

What the company changed

Rexera started with a narrow, unpleasant, repeatable step: acquiring homeowner-association documents. Its website describes the original workflow as more than 11 manual steps, including finding the association, discovering ordering rules, paying a fee, sending a request, following up, receiving the package, checking completeness, and extracting data. The platform later extended the pattern to municipal liens, mortgage payoffs, condominium reviews, and other operations.

According to AWS, a two-week proof of concept was followed by a two-week migration to Amazon Bedrock, with production use beginning in the second quarter of 2024. The platform uses multiple language models, document RAG, data verification, and agents that can interact with communications and external systems. A separate AWS and Zilliz case describes more than 10,000 tasks per day and millions of pages per month. Participants report that moving to one managed vector database reduced overall costs by 50%, latency by 30%, and improved retrieval accuracy by 40% compared with previously tested options.

RealPage acquired Rexera in July 2025. An acquisition does not validate every performance claim, but it does confirm that this was an operating vertical product rather than a laboratory demo.

A practical architecture

Public materials do not disclose the complete design, so the following is an editorial reconstruction from described functions, not Rexera's internal documentation.

1. An intake orchestrator receives the property, workflow type, and deadline, then creates a stable case identifier.
2. Connectors discover counterparties, send requests, and record every event in an audit log.
3. Documents pass through malware scanning, OCR, segmentation, and classification.
4. An index stores passages with the property, version, page, access policy, and validity period.
5. RAG retrieves only relevant passages; a model returns structured fields and evidence references.
6. Deterministic code validates dates, amounts, mandatory fields, contradictions, and duplicates.
7. Low-risk steps continue automatically, while legal, financial, and ambiguous cases enter a specialist's exception queue.
8. After approval, a separate executor performs a write or sends a request with an idempotency key so a retry cannot create a duplicate payment or request.

The key lesson is that an LLM should not read a document, decide, pay, and change a system of record in one unbroken step. The model proposes an interpretation. Policies, permissions, validation, and execution remain separate components.

Why RAG is only one layer

RAG attracts attention in the case, but savings arise around it. If an employee must still discover the organization, request a file, verify completeness, and copy data into a CRM, a fast answer over a PDF optimizes only a small part of the job.

An end-to-end outcome needs:

• a case object with status, deadline, and owner;
• a registry of document sources and versions;
• connectors to email, CRM, telephony, storage, and accounting systems;
• completeness and consistency rules;
• an exception queue with a reason for human escalation;
• an action log and a way to reverse changes;
• metrics at the completed-operation level.

That is why implementation should begin with a process map rather than a vector database choice. Vector search can be replaced. An implicit permission, missed deadline, or incorrect payment is much more expensive to replace.

What transfers to a Russian small business

An American HOA transaction is not identical to Russian practice. The pattern still applies to real-estate agencies, law firms, mortgage brokers, property managers, insurance intermediaries, and outsourced accounting teams—anywhere a case is assembled from inconsistent documents and correspondence.

A local deployment is particularly appropriate when files contain passports, contracts, bank details, client data, or commercial terms. The managed models in the original case can be replaced by a local model at the extraction and explanation layer, but that does not remove the need for OCR, version lineage, access control, and safe integrations.

A minimum pilot dataset includes:

• 200–500 completed, sanitized cases;
• ground-truth fields and final outcomes;
• mandatory-document lists for each workflow type;
• real exceptions such as poor scans, new forms, duplicates, contradictions, and missed deadlines;
• employee time per stage;
• the cost of a miss and the cost of a false alarm.

If the archive cannot be sanitized, place the evaluation environment inside the same protected boundary as the source documents. Model access must inherit user permissions; the vector index must not become a common pool of passages visible to every department.

Calculating economics without a presentation percentage

A small company should calculate cost per accepted case. Consider an explicitly modeled example: 600 cases per month, 45 minutes of manual handling per case, and a fully loaded staff cost of RUB 900 per hour. The baseline is 450 hours, or RUB 405,000 per month.

A pilot automates preparation for 60% of cases, reducing those cases to 15 minutes, while the remaining 40% still require 45 minutes. The saving is 180 hours, or RUB 162,000. Infrastructure, support, exception review, and implementation amortization must then be subtracted. If total monthly costs are RUB 110,000, the net effect is RUB 52,000—not “60% lower costs.”

The model must be tested for sensitivity:

• how many cases finish without rework;
• whether review time grows because the model produces long answers;
• the financial exposure of one critical error;
• time spent updating templates and the index;
• idle time for a dedicated GPU outside peak periods;
• whether extraction can run as a nightly batch.

A local server is justified not by confidentiality alone, but by the combination of data requirements, steady demand, and lower total cost over the expected operating horizon.

A safer four-stage pilot

First, choose one case type and one measurable outcome—for example, package completeness and extraction of five critical fields with page references. Do not automate calls, payments, and the final decision at the same time.

Second, build a shadow pipeline. It reads copies, prepares a case card, and explains each alert, but cannot contact external parties or modify the CRM. A specialist compares the proposal with the actual outcome.

Third, automate only reversible actions: classification, draft creation, reminders to an owner, or a request for a missing file after confirmation. Keep explicit approval for money, signatures, account-detail changes, and legal conclusions.

Finally, release by risk class. Measure cases completed without rework, false acceptance, evidence completeness, cycle time, cost per case, and exception-queue load. If the queue grows faster than volume, automation has merely moved manual work.

The management takeaway

The Rexera case shows that practical AI begins with a governed workflow, not “chat with documents.” RAG can find evidence, multiple models can divide work, and an agent can connect steps. The business value, however, comes from audit logs, deterministic rules, an exception queue, and a person's authority to stop an operation.

The best first step for a small company is to take one case type and 200–500 historical examples, then run evidence-backed extraction in shadow mode for two weeks. Choose a local model, cloud service, and server only after measuring cost per accepted result and the rate of critical errors.