Building a safer AI agent for Crown Court Order analysis

AI is often discussed in criminal practice in broad terms: efficiency, automation, innovation. Those are all relevant but for defence solicitors the real question is more practical: can AI help?

 

Stone King's Crime Team have recently built an AI-assisted agent designed to do exactly that. Its purpose is not to replace legal judgment or advocacy, but to support it – by bringing structure, consistency and speed to the review of proposed Crown Court Order terms.

 

The starting point is the form. Requests are captured through a simple structured input – whether a single term or an entire draft order. This ensures the agent is working from the operative wording, rather than narrative material and reflects how practitioners actually approach these documents in court.

 

From there, the process becomes an automated workflow. The input is passed through a controlled system which retrieves the relevant material and sends it to the agent for analysis. The key here is that the workflow is deliberately narrow and predictable: it avoids unnecessary duplication, removes administrative friction, and ensures each request is handled in a consistent way.

 

At the centre of this is the AI agent itself. Crucially, it is not allowed to “roam”. It is grounded in a curated repository of legal materials – statutes, appellate authorities, and approved guidance – and is restricted to those sources only. The agent is required to adopt an authority-first approach: it must identify relevant case law, provide a neutral citation, refer to the paragraph where available and extract the relevant passage before making any comment. If no authority is found, it must say so expressly.

 

This leads to the output, which is structured and usable in practice. Each term is set out, followed by a response that highlights any supporting or problematic features by reference to authority. The output is not a piece of advocacy in itself; rather, it is a disciplined piece of research support, designed to sit alongside the solicitor’s own analysis.

 

The safeguards are as important as the functionality. The agent is not permitted to invent cases, quotations or legal principles. It cannot give definitive recommendations on whether a term should be accepted or opposed. And every output is subject to lawyer review – the solicitor must check the full judgment, the current law and the specific facts before relying on it.
 

Used properly, this approach offers real benefits: faster identification of issues, greater consistency across cases, and more focused preparation for court. Just as importantly, it demonstrates that AI in criminal practice can be deployed in a way that enhances, rather than undermines, professional standards.