Consulting Firm Research: Private AI for Knowledge Management and Report Drafting

The firm had been operating in Bangkok for over fifteen years. Forty consultants across six practice areas had produced thousands of deliverables: market assessments, due diligence reports, competitor analyses, operational reviews, strategy decks. The work was good. Finding it again was the problem.

Engagement files lived in nested folder structures on a shared drive. Naming conventions had changed three times over the years. Some reports were in Word, others in PDF, a few only existed as email attachments that never made it to the server.

When a new engagement came in, the standard process was to ask around: "have we done something like this before?" The answer was usually yes, but finding the actual work meant hours of digging through folders and hoping someone remembered which partner led the similar project in 2019.

The research problem was similar. Consultants gathered market data, regulatory filings, and interview notes during engagements. That research disappeared into project folders when the engagement ended. Three months later, a different team working on a related sector would start from scratch.

The firm had considered cloud-based AI tools but the partners rejected them. Client deliverables contain sensitive financial data, strategic plans, and confidential interviews. Uploading that corpus to a third-party API was not an option.

The system runs on a server in the firm's office. No data leaves the local network. The private AI layer sits on top of the firm's existing file storage and indexes everything into a searchable knowledge base.

A pipeline crawled the shared drive and processed every document: Word files, PDFs, PowerPoint decks, Excel models. Each document was chunked, embedded, and stored in a local vector database alongside its metadata: engagement name, client sector, practice area, date range, authors. The full corpus came to roughly 2,400 engagement deliverables and several thousand supporting research documents.

On top of the indexed corpus, OpenClaw orchestrates the conversational layer. Consultants interact through an internal messaging channel. A query like "what do we know about logistics infrastructure in the EEC" returns relevant passages from prior deliverables, research notes, and email threads, ranked by relevance and tagged with source documents.

Report drafting works as a structured workflow. A consultant provides an engagement brief: client sector, scope, key questions. The system searches the knowledge base for relevant prior work, assembles a draft outline with references, and generates section drafts that draw on the firm's own language and frameworks. The output is a first draft, not a final product.

When consultants gather new market data or regulatory updates during an engagement, that material is tagged and indexed automatically. The knowledge base grows with every project.

A principal preparing a proposal asks the system to show prior work in Thai healthcare, especially hospital operations or procurement. Within seconds, the system returns a ranked list of relevant deliverables with summaries, dates, and links to the original files. What used to take an afternoon of folder diving takes less than a minute.

A junior consultant starting research on a market entry analysis asks what the firm already knows about the target sector. The response includes passages from three prior engagements, two internal research memos, and a relevant email thread from a partner who worked with a similar client two years ago.

During report drafting, a consultant feeds the engagement brief into the workflow. The system identifies structurally similar past deliverables, pulls the firm's standard analytical frameworks, and produces a first draft with inline citations. The consultant spends time refining arguments rather than writing boilerplate.

Partners use the system before client meetings. "Summarize our engagement history with this client" produces a briefing covering past deliverables, key findings, and follow-up recommendations. The system handles Thai and English documents equally, treating the firm's bilingual corpus as a single searchable body of work.

The most immediate change was speed. Research that previously took a full day now takes minutes. The firm estimates consultants spend roughly 65% less time on the "finding what we already know" phase of new engagements.

Proposals started referencing specific prior work more consistently. Reports built on the firm's existing analytical frameworks instead of reinventing them. Junior consultants produced stronger first drafts because they had access to the same institutional knowledge that senior partners carried in their heads.

Knowledge retention improved structurally. When a senior consultant left the firm, their engagement history and research remained fully searchable. Individual expertise became organizational memory without requiring handover documents or wiki updates.

The Bangkok office runs the system on its own hardware. There is no per-query fee, no usage-based pricing, no external dependency that could change terms. Client confidentiality is straightforward because nothing leaves the server.

The system runs on a server in the firm's office. Document processing uses a pipeline of deterministic extraction (Apache Tika for parsing, standard chunking for segmentation) followed by local embedding for vector search. The vector database stores document chunks alongside structured metadata: engagement ID, practice area, document type, date, author.

Search combines vector similarity with metadata filtering. A query like "due diligence reports in energy sector from 2022 to 2024" uses semantic search against document content and structured filters against metadata fields. Results include relevance scores and direct links to source documents on the shared drive.

Report drafting uses a structured pipeline rather than open-ended generation. The system identifies relevant prior deliverables, extracts the firm's analytical patterns, and generates section drafts with inline citations. Every claim in the draft traces back to a cited source document.

OpenClaw provides the orchestration layer: plugin registration for search and drafting tools, cron scheduling for periodic re-indexing as new documents arrive, and messaging integration so consultants interact through existing channels. The system does not require internet access to operate. All inference, search, and document processing happens on local hardware.