Build private AI document search for law and accounting firms
Small law firms, accountants and surveyors sit on thousands of documents they can't paste into ChatGPT. Install private search in their own office: a plain-English question returns the exact paragraph with a citation, limited to the matters each person may see, with every search logged. Few people can do the build, the security and the paperwork properly, which is why firms pay for it.
- A way in: PECR-safe outreach by firm type, a paid pilot offer, pricing with the hours behind it, a weekly routine and a first 30 days
- The basics first: Build your own "chat with my files" app in Python, step by step. This mission assumes you know embeddings, chunking and top-k search, and goes further
- uv (as in that lesson). The demo downloads about 155MB of small models and runs on a normal laptop CPU. No GPU, no Ollama needed for the demo
- Comfort with Python, a Linux server and basic networking (firewalls, ports, HTTPS). A few weeks of building practice covers any gaps
- The kit: download it all as a zip, or read the kit overview and grab just the demo
- About 3 hours for the mission, then around 8 hours a week of research, letters, calls and demos
Spot the gap, and know what it takes
Know exactly what firms are worried about, what you're selling, who you're up against and what it will cost you in hours.
Every small firm has the same two facts. Its best work is buried in years of precedents, letters, policies and closed files. And its staff are told, rightly, not to paste any of it into ChatGPT. That gap between "AI could find this in seconds" and "we're not allowed to use it" is your opening.
The regulators have made it concrete this year:
- Solicitors. The SRA's Misuse of AI warning notice ↗ (17 August 2026) says client information should only go into AI systems where appropriate contractual, technical and organisational safeguards are in place, and warns that both free and paid tools may put confidentiality at risk and that privilege may be permanently waived. The SRA had 42 reports of possible AI misuse in a year, with investigations ongoing ↗. The Law Society's Generative AI: the essentials ↗ gives small firms a checklist covering vendor data security and checking outputs.
- Accountants. ICAEW warns that putting client data into tools like ChatGPT is a potential breach ↗, and its July 2026 piece on protecting client confidentiality ↗ asks firms to know where data is stored, who can access it and whether it trains models, and to limit internal access to people who need it.
- Surveyors. RICS's Responsible use of AI in surveying practice ↗ standard has been mandatory since 9 March 2026: written AI policies, risk assessment before use, and telling clients in writing when AI is used.
- Everyone. It's personal data, so UK GDPR applies. The ICO lists AI as innovative technology that usually needs a DPIA ↗ when combined with another risk factor, and client files are full of them.
The offer, in one line"Private AI search for your firm's documents. It runs in your office, shows the exact paragraph it found, respects who can see which matter, and logs every search. Fixed-price pilot first, so you see the accuracy on your own questions before you commit."
Notice what it is not: a chatbot that writes legal advice. It's search with citations, with an optional short summary that cites every sentence. That keeps the fee earner responsible, which is what every regulator above asks for, and it's far easier to make reliable.
What you're up againstKnow this before a managing partner says it to you:
- Microsoft 365 Copilot. UK guides quote Microsoft's add-on at about £23 per user per month ↗, billed annually, on top of Microsoft 365, and resellers offer cheaper small-business bundles. If a firm already keeps everything in SharePoint with clean permissions, Copilot may be the better choice. Say so.
- Legal AI platforms. Harvey publishes no prices; reported figures are around $1,200 a seat a month with minimums of about 25 seats ↗, which puts it out of reach for a 15-person firm. Lexis+ AI is quote-only, with third-party estimates of $250 to $500 a user a month ↗ all in. These sell research over legal content; you sell search over the firm's own documents. Different jobs.
- The firm's IT provider. Many small firms pay a managed IT company. Vendor guides put UK managed IT at roughly £40 to £95 per user per month ↗. They can block you or bring you in. Step 6 makes them a partner.
Your edgedocuments that live on a file server or case management system rather than SharePoint, partners or insurers who don't want client files in anyone's cloud AI, paragraph-level citations, measured accuracy on the firm's own questions, and a fixed cost instead of a per-seat bill that grows with headcount.
TipWhat it takes. This is serious engineering: Python, Linux, networking, permissions and evaluation, done carefully. A permission bug here is a breach of client confidentiality, possibly across an information barrier, so the build and testing in steps 3 and 4 are the heart of the service. Partners decide at monthly meetings and compliance leads want paperwork, so plan on two to four months of steady letters, calls and demos before the first paid pilot. Average cold email reply rates are low (4.1% ↗ across 11 million emails in Hunter's study), which is why this mission leads with letters, introductions and talks. Firms will ask about professional indemnity and cyber insurance, so get quotes early. Who this suits: engineers who enjoy compliance detail and can give it around 8 hours a week for a few months. Keep your day job while you land the first pilot.