Mid-Market Services Firm Saves 1,200 Hours/Year with AI Document Automation

The client

A mid-market professional services firm (legal-adjacent vertical) with offices in Miami and Dallas, 180 staff, and a high-volume document-review workflow. Roughly 30 staff spent 6+ hours per week each on intake document review — extracting key facts, flagging risks, and producing summary memos. Names and exact verticals are anonymized at client request.

The situation

• Average client document package: 80–250 pages of PDFs.
• Manual review cycle: ~4 business days from intake to summary memo.
• Some staff had begun pasting client documents into consumer ChatGPT — a clear shadow-AI risk.
• Leadership wanted AI deployed, but only with auditability, data privacy, and a real ROI case.

The constraints

• Data privacy. Client documents could not be used to train any vendor model. Tenant isolation and BAA-level controls were required.
• Auditability. Every AI-generated output had to cite the source pages and be reviewable.
• Existing M365 stack. Solution had to work with SharePoint and Entra ID, not require a parallel platform.
• Realistic budget. No "boil the ocean" enterprise AI program — the firm wanted one workflow productionized first.

BASG's approach

1. AI governance first. Before any tooling, BASG drafted an AI acceptable-use and data-handling policy, identified shadow-AI risks (consumer ChatGPT use), and trained staff on what could and could not be sent to which tools.
2. Vendor evaluation. Compared three approaches — Microsoft 365 Copilot, a fine-tuned model, and a custom retrieval-augmented generation (RAG) system. Selected RAG on Azure OpenAI for tenant isolation, source-citation grounding, and cost predictability.
3. RAG architecture. Built on Azure OpenAI (no training on customer data), with a private vector index of the firm's playbook documents and risk taxonomies, document ingestion via SharePoint, and chunked retrieval with explicit source page citations on every output.
4. Workflow integration. Wrapped the RAG system in an internal web app behind Entra ID SSO. Staff submit a document package, receive a draft summary memo within ~10 minutes, then review and edit before client delivery. Every output logs the source pages cited.
5. Audit + governance. Full audit logging of who submitted what document, what the model returned, and what the human reviewer changed. Quarterly model evaluation against a held-out test set to detect drift.

Outcomes

• ~1,200 staff hours/year saved across the 30 reviewers — a hard-dollar reduction in cost-to-serve.
• Average turnaround dropped from ~4 business days to under 24 hours.
• Source-cited outputs. Every AI-generated memo lists the specific source pages it drew from, enabling trustworthy human review.
• Shadow AI eliminated. With a sanctioned, faster, auditable internal tool available, paste-into-ChatGPT behavior dropped to zero based on M365 DLP telemetry.
• Documented AI governance policy in place, integrated with the firm's existing security and compliance program.
• Quarterly model evaluation running against a held-out test set; first quarter showed a 4-point F1 improvement after a prompt revision.

Anonymized client testimonial

"We talked to two AI consultancies before BASG. The other two pitched us a transformation. BASG asked us where document review was costing us the most hours, built one workflow that solved that, and shipped it in eight weeks. We trust it because every answer cites the source pages, and we own the architecture."

— Managing Partner, mid-market professional services firm, Miami / Dallas

Could BASG do this for your firm?

If your team is hand-reviewing documents that an AI could draft in minutes — but you can't risk sending client data to consumer AI tools — see our Enterprise AI Solutions page, our blog post on Enterprise AI Adoption, and our analysis of AI Agent Security.