CASE STUDY 04 — KNOWLEDGE SEARCH

Internal Knowledge Search

How a 150-person accounting firm gave every employee a research assistant that never phones home.

The firm

Solstice Financial is an accounting and advisory firm with 150 staff across two offices. They advise 400+ clients on tax, audit, corporate finance, and advisory work. Their internal knowledge is vast — 15 years of engagement files, tax research notes, client histories, precedent documents, internal policies, and regulatory briefings. It's all stored on a shared drive and in their document management system. Nobody can find anything.

The problem

When a junior accountant needs to know "How did we handle a similar R&D tax credit claim last year?" they have three options: search the shared drive (25 minutes, 400 irrelevant results), ask a senior colleague (10 minutes of re-explaining), or Google it (generic guidance that doesn't reflect the firm's experience).

25-45 min
Time to find info
5-10h
Senior/week repeat Q&A
4-6 mo
Onboarding to independent
£600-900
Monthly API cost

The firm's partners have said no to using ChatGPT or similar tools for internal research. Client financial data, tax strategies, and advisory work product can't go through a third-party API. The ICO would have questions. The clients would have concerns. The professional indemnity insurers would have opinions.

What Foundry does

Foundry runs on a Mac Studio in the firm's server cupboard (it doesn't need a rack — it sits on a shelf). It's plugged into the firm's network and has read access to the document management system.

It reads everything, understands it, and answers questions — locally.

What it looks like day to day

Tuesday, 10:15 AM. A junior accountant, Priya, is working on an R&D tax credit claim for a software company. She's never done one before.

She types into the Foundry search bar: "R&D tax credits for software development — have we done this before?"

Foundry responds in 3 seconds:

Yes. The firm has handled 7 R&D tax credit claims for software companies since 2022.

Most relevant: The 2024 claim for BrightStream Ltd (engagement file 2024-0341) involved similar circumstances — a SaaS company claiming for backend algorithm development. The senior on that engagement was Michael Chen. Key precedent: HMRC accepted that server optimisation algorithms qualified, but frontend UI work did not.

Relevant documents: BrightStream R&D claim summary (2024-0341, p. 12-18) | HMRC correspondence (2024-0341, p. 31-34) | Internal R&D software sector guidance (April 2024)

Priya reads the BrightStream summary, finds the exact approach the firm used, and adapts it for her client. She books a 15-minute call with Michael Chen to confirm her approach rather than a 2-hour "start from scratch" meeting.

The transformation

MetricBeforeAfterChange
Time to find internal info25-45 min30 sec - 3 min-90%
Senior repeat Q&A5-10 hrs/week1-2 hrs/week-80%
Onboarding to independent4-6 months2-3 months-50%
Duplicate researchFrequentRareNear eliminated
API cost£600-900/month£0 (local)-100%

Annual impact: 8,000+ hours of professional time recovered across the firm. At an average charge-out rate of £150/hour, even recovering 10% of that for billable work is £120,000/year.

Foundry cost: £999 setup + £99/month = £2,187 first year.

What the team says

"I used to spend half my morning looking for a precedent file. Now I ask Foundry, get the answer, and I'm working on the actual problem in under a minute."Priya, junior accountant (6 months qualified)
"The real surprise was onboarding. New joiners are useful in week two instead of month three because they can actually find how we do things here."Senior manager
"We banned ChatGPT for client-related research in 2025. We lost nothing — Foundry does it better because it knows our firm's experience, not just the internet's."Partner, risk and compliance

Technical details

Hardware: Mac Studio M3 Ultra, 512GB unified memory
Model: Qwen3 (30B or 235B MoE depending on collection size) via MLX or llama.cpp
Embedding model: nomic-embed-text-v1.5
Index size: ~1GB per 10,000 documents
Query latency: 2-5 seconds for 50,000+ documents
No-cloud posture: All indexing, retrieval, and answer generation performed locally
Citation accuracy: Every claim links to a specific document and page/section

Is this right for your firm?

Works for professional services firms with 50+ staff and 5+ years of accumulated documents, accounting/legal/consulting/advisory firms with data confidentiality obligations.

Book a Foundry Fit Review →

← Back | Next: Code Review →