George Hannah, Solicitor Apprentice at Lewis Silkin, Liam Fitter, Product Compliance Associate at Dataships, and Nestor Dubnevych, Business Owner at Legal Nodes (which builds agents for regulatory teams), reviewed October’s legal AI developments. Hannah offered the view of a trainee using AI tools at work, Fitter a product compliance perspective, and Dubnevych a builder’s view of agents. Audience member Andy McDonnell added pointed questions on adoption and pricing.

Overview

The central question was what remains for specialist legal AI vendors as frontier labs move into legal work. The speakers noted Claude’s improving legal capability and new artifact feature, Gemini’s legal offering and Microsoft’s Copilot for legal, which Fitter believes is the fourth frontier lab to release a legal-specific product. Dubnevych also noted Meta’s latest model scoring markedly higher on a legal benchmark than before. Their shared view was that many teams will ask at contract renewal whether a general model is good enough.

The reasoning went beyond features. Dubnevych argued that finance teams will scrutinise return on investment, and that the real cost of multiple tools is context switching: one interface with projects and connectors protects a lawyer’s focus. Fitter was sceptical of vendors’ own ROI dashboards, because the value of qualitative legal work is hard to quantify, but stressed that persistent context matters for principles-based regulatory judgement. Hannah cited a claim that regaining focus takes around twenty-three minutes, and observed that faster AI turnaround means more task switching. On artifacts, Dubnevych and Fitter agreed that they suit settled data as static snapshots, not live dashboards that teams rely on.

The practical implications centred on training and cost. Fitter described an Irish firm, from a conversation around two years ago, that withheld AI tools from trainees until senior associate level, while others give access from the start. Hannah described Ruby Legal Training, a sandbox that prepares associates to spot bad outputs, and Dubnevych proposed simulated “digital twin” environments with synthetic client matters. Fitter warned of becoming a “meat proxy” between AI systems. On pricing, speakers agreed token costs are heavily subsidised. Dubnevych expects vendors to post-train small models, while Hannah relayed a founder’s view that falling model prices may make that unnecessary. McDonnell argued that the true costs may not fall enough, and that Harvey’s approach suggests most work does not need a frontier model.

Key Takeaways

• Test whether a general model is good enough before renewing. Compare specialist tools with Claude or Copilot against real workflows, not vendor claims.
• Treat vendor ROI figures with caution. Legal work is qualitative, so seek your own evidence of value.
• Count the cost of context switching. Fewer tools with shared context may protect focus better than a larger stack.
• Use AI visuals for settled information. Static artifacts suit finished work, not live dashboards that need constant updating.
• Train juniors to supervise AI, not just receive its output. Simulated practice and spotting bad outputs may build judgement better than restricting access.
• Plan for token pricing risk. Subsidised costs may rise, so consider model routing, smaller models, or on-premise options, while recognising the speakers disagreed on which will prove best.