AI Case Study · Tax & Compliance · Professional Services
From six hours to seconds. Enterprise RAG for a leading international US tax advisory firm.
- Client
- Anonymized · international US tax advisory firm
- Industry
- Tax Advisory · Professional Services
- Service
- Sphere AI Engineering
Plain English
If you only read one box.
Their advisors lived inside a document retrieval problem — FATCA, FBAR, treaty law, and multi-jurisdictional guidance scattered across shared drives, a legacy DMS, and email archives. Sphere built a jurisdiction-aware enterprise RAG system that solved it in five weeks.
A leading international US tax advisory firm — serving American expatriates and corporations from its European headquarters across offices in Europe, Asia, and the Middle East — was drowning in its own knowledge. Regulatory publications, client case archives, advisory memos, IRS guidance, and treaty documents lived in a fragmented ecosystem of shared drives, a legacy document management system, and email.
The fix wasn't a better folder structure. Sphere built a production enterprise RAG system: a custom ingestion pipeline, hybrid vector + keyword retrieval with jurisdiction-aware metadata filtering, rigorous pre-production evaluation benchmarks, role-based access control, and a tamper-evident audit trail — designed in, not bolted on.
The result: document research that took advisors an average of six hours per engagement now takes seconds, retrieval accuracy improved by 66%, and the whole system reached production in five weeks.
Same advisors. Same knowledge base. A completely different working day.
Chapter 01 · The problem
The firm's most valuable asset was also its least accessible.
Growth multiplied the firm's knowledge faster than any human filing system could keep up. Six hours of research per engagement was the tax the firm paid on its own success.
The client is a leading US tax advisory practice for Americans abroad. Their team of CPAs and IRS Enrolled Agents guides clients — from individual expatriates to multinational corporations — through the most complex intersections of US federal tax law, local tax rules in their operating jurisdictions, and international treaty obligations. Every engagement touches FATCA, FBAR, Form 5471, FIRPTA, bilateral estate tax treaties, streamlined filing procedures, and a dense web of IRS guidance that changes regularly across multiple jurisdictions.
As the firm expanded from its European headquarters into additional financial centers across Europe, Asia, and the Middle East, so did the volume and complexity of the knowledge it needed to manage. Regulatory publications, client case archives, internal advisory memos, IRS announcements, OECD treaty documents, regional tax guidance, and prior engagement records had accumulated across a fragmented ecosystem of shared drives, a legacy document management system, and email archives.
The cost was concrete: advisors spent an average of six hours per engagement on document research alone — time billed against margins, not against insight.
Chapter 02 · The build
Discovery first. Architecture second. No stack proposed before the failure modes were understood.
Sphere mapped every knowledge source and its sensitivity level before writing a line of pipeline code — so access control and auditability were architecture, not afterthought.
Sphere began with a structured discovery process rather than immediately proposing a technology stack. Understanding the specific failure modes of the firm's existing retrieval workflows — and the compliance sensitivity of the data involved — was a prerequisite to designing a system that would actually work in a regulated professional services environment.
Knowledge source mapping. Sphere inventoried every repository in the firm's ecosystem: IRS publications and revenue rulings, OECD treaty documentation, regional tax guidance, internal advisory memos, client case archives, engagement letters, and research from prior matters — mapping sensitivity and access requirements for each source type.
Custom ingestion pipeline. Domain-optimized chunking and multi-jurisdiction document classification, tuned to how tax guidance is actually structured and cited.
Hybrid retrieval architecture. Vector + keyword search fused with semantic reranking and jurisdiction-aware metadata filtering, so a query about one jurisdiction never surfaces guidance governing another.
Role-based access control from the architecture stage. Advisors retrieve only within their authorization scope; client-specific records never surface in cross-advisor queries.
Tamper-evident audit trail. Every query, retrieval event, and generated response logged — source documents, cited chunks, and synthesis — giving the firm's leadership full visibility and an evidentiary record for any engagement subject to later professional review.
Chapter 03 · The proof gate
No interface shipped before the retrieval benchmarks passed.
Sphere's position: deploying RAG without rigorous pre-production evaluation is the single most common failure mode in enterprise AI. Every milestone had a quantitative pass/fail threshold.
Before deploying any interface, Sphere ran structured RAG evaluation benchmarks across four dimensions: retrieval precision, retrieval recall, answer faithfulness (whether the generated response was grounded in the retrieved documents), and answer relevancy (whether the response actually addressed the advisor's query).
Each milestone carried a quantitative pass/fail threshold before the build progressed. The result of that discipline: retrieval accuracy improved by 66% over the firm's previous keyword-based search — measured, not asserted — and the system went from kickoff to production in five weeks.
The impact of collapsing per-engagement research from six hours to seconds compounds rapidly across a professional services firm. Advisors take on more engagements without added headcount. Junior advisors access the same depth of institutional knowledge as senior partners. And the firm's multi-market expansion runs on a knowledge base that travels with the business — not expertise that stays at headquarters.
“Sphere deployed a production-ready RAG pipeline in five weeks. Document retrieval accuracy improved by 66% compared to our previous keyword-based search. Most strikingly, the time our advisors spend on document research dropped from an average of six hours per engagement to seconds. Sphere understood that responsible AI means the governance layer is designed in — not added later.”
Capabilities delivered
Enterprise RAG in regulated professional services
The client — a leading international US tax advisory firm — requested that their company name and the names of their stakeholders not be disclosed publicly. All results, timelines, and the client quote in this case study are real and verified by Sphere; only identifying details have been removed.
Enterprise Retrieval-Augmented Generation (RAG) combines semantic vector search with a large language model to retrieve contextually relevant document passages and synthesize a grounded, cited answer — rather than returning a list of document links. Unlike keyword search, RAG understands the meaning of a query and retrieves semantically similar content even when the exact words don't match. In professional services environments, production RAG adds domain-optimized chunking, jurisdiction-aware metadata filtering, faithfulness evaluation, and audit trail logging that generic implementations omit.
Sphere deployed this production-ready enterprise RAG pipeline in five weeks — covering knowledge source mapping, custom document ingestion pipeline development, embedding configuration, hybrid retrieval architecture, RAG evaluation benchmarking, role-based access control, and audit trail implementation. Timelines vary by knowledge base complexity and number of document sources, but Sphere's Precision-Driven Engineering™ framework is specifically designed to reduce time-to-production for AI systems in regulated environments.
Yes — when built correctly from the architecture stage. This implementation included role-based retrieval access controls (preventing cross-advisor access to client-specific records), a tamper-evident query and retrieval audit log, and deployment within a data governance framework aligned to the firm's professional confidentiality obligations.
The baseline — an average of six hours of document research per engagement — was established during discovery from advisor time records across the firm's prior keyword-based workflow. The post-deployment figure reflects measured query-to-cited-answer times in the production system across live engagements, alongside a 66% improvement in retrieval accuracy validated through Sphere's structured evaluation benchmarks.
Yes. The pattern — a governed ingestion pipeline, hybrid retrieval with domain-aware filtering, evaluation gates, and audit-grade logging — generalizes across regulated professional services: legal, accounting, compliance, insurance, and financial advisory. Sphere delivers with defined success criteria, integrated against your existing data and identity stack.
