Legal
AI that cites its working.
Practical AI for UK law firms and in-house teams. Retrieval over your own files, not a black box trained on public case law. Every answer linked back to the source paragraph. Nothing privileged leaves your network.
Two weeks. Fixed price. Honest answer.
Where it works
Four applications
we've shipped.
01
Precedent and case-file retrieval
The work your firm has done sits across PDFs, scanned bundles, Word documents, and partner inboxes. Public databases (Westlaw, LexisNexis) cover public case law — your internal precedents are the bit that's actually privileged and competitive, and the bit nobody can find. We build retrieval systems that read legal context, not keywords, and cite every answer back to the source paragraph.
→ Typical result: Around 70% reduction in precedent research time, measured on partners' own time logs (73% on the public RAG build).
02
Contract drafting and review
Standard agreements (NDAs, supplier terms, engagement letters, employment contracts) get redrafted from templates that drift. We build systems that assemble the right template, flag missing or unusual clauses, and surface the diffs against your firm's preferred positions. The partner still drafts and negotiates — the system handles the mechanical work around it.
→ Typical result: 4 hours to 45 minutes for standard agreements. Senior time freed for the genuinely bespoke work.
03
Compliance and regulatory monitoring
FCA, SRA, ICO and sector-specific updates land daily. Keeping up costs a compliance officer's week and still misses things. We build systems that ingest the updates, classify them against your client base and matter types, and surface a daily two-minute review queue rather than a weekly half-day.
→ Typical result: Weekly manual review collapses into a daily two-minute scan. Misses drop, audit trail improves.
04
Document intake and triage
Inbound correspondence — court orders, statements, schedules, requests — arrives in mixed formats and gets manually classified before it reaches the right person. We build classification systems that read the document, identify type and urgency, extract the structured fields, and route with the original attached. The lawyer doesn't lose context; they save the triage time.
→ Typical result: Triage time falls from minutes per document to seconds. Misrouting drops below 2%.
A project in the open
Fifteen years of case files, searchable in seconds.
A UK commercial litigation firm with 2.3 million pages of internal case files. Partners spent half a day per matter finding analogous precedents — by memory, by email thread, by asking the senior associate who happened to be there. A larger consultancy had quoted them £180k for a fourteen-month "knowledge management transformation" with no commitment to a working interface.
Eight weeks later we'd shipped a citation-first retrieval system over the whole corpus, on the firm's own infrastructure, behind their SSO.
Read the full case study- 73%
- Research time cut
- 2,000+
- Queries per day
- 2.3m
- Pages indexed
- 8 wks
- To production
Constraints we design around
Legal AI fails differently to other AI.
Citation is non-negotiable
No retrieved fact without a clickable source. No generative answer without the underlying passages. We don't deploy anything where a partner can't verify the working in seconds.
Privilege stays inside the firm
Self-hosted embeddings. Vector store on your infrastructure or your private cloud, behind your SSO. Nothing privileged goes to a third-party API. UK data residency by default.
Hallucinations are a deployment risk
Retrieval-first architecture means the model surfaces what exists rather than inventing what doesn't. Confidence thresholds and "we don't know" responses are part of the design — not an afterthought.
Audit trail for regulators
Every query, every retrieved passage, every model response logged. If the SRA or your insurer asks how the system reached a conclusion, the answer is in the logs, not in a black box.
Got an archive worth searching properly?
We start with a two-week diagnostic. If the numbers don't work, we tell you and refund the second week.