In late 2025, a Bristol-based commercial law firm came to us with a problem. Their junior solicitors were spending roughly 12 hours per week answering the same client questions: what documents do I need for a lease assignment, what's the typical timeline for a share purchase, how does the SRA regulate X. The answers were all in their internal knowledge base. But finding them required searching through 400+ documents across three different platforms.
What we built
We built a custom GPT trained on their internal knowledge base. Not their client files. Not their case documents. Their knowledge base: template guides, process checklists, FAQ documents, and regulatory summaries. The distinction matters because no confidential client data was ever used as training material.
The tool sits in a private interface that only the firm's staff can access. A solicitor types a question in natural language. The system searches the knowledge base and returns a sourced answer with references to the specific documents it pulled from. If the answer isn't in the knowledge base, it says so rather than guessing.
What £8k covered
The total build cost was £8,000 including VAT. Here's where the money went. £2,500 on knowledge base audit and preparation: reviewing 400+ documents, identifying which ones were current and accurate, structuring them for AI ingestion. £3,000 on system build: configuring the GPT, building the search layer, testing accuracy across 150 sample queries. £1,500 on integration and deployment: setting up secure access, user permissions, and the web interface. £1,000 on training and documentation: showing the team how to use it, documenting the process for future knowledge base updates.
The results after 90 days
Junior solicitors reduced their FAQ-answering time from 12 hours per week to roughly 4 hours. That's 8 hours per week saved across the team, or about 32 hours per month. At an average internal billing rate of £150 per hour, that's £4,800 per month in recovered capacity. The system paid for itself in under two months.
Accuracy was the metric the firm cared about most. Over the first 90 days, they tested 300+ queries against known correct answers. The system returned accurate, source-referenced answers 94% of the time. The remaining 6% were cases where the knowledge base itself was incomplete or outdated, which actually helped the firm identify gaps in their documentation.
What we learned
The AI is only as good as the knowledge base. We spent 30% of the budget on document preparation, and it was the most valuable 30%. Garbage in, garbage out applies more to AI than to almost anything else.
Staff adoption depends on trust, not training. The solicitors who used the tool most weren't the ones who attended the training session. They were the ones who tested it, found it gave accurate answers, and started relying on it. Trust is earned through accuracy, not presentations.
