Catherine Bamford, CEO and Founder of BamLegal, has run the consultancy for fifteen years, advising law firms and in-house legal teams on legal technology, with a focus on drafting legal documents. She previously worked as a legal engineer and, before that, as a commercial property lawyer. She was joined by David Howorth, co-founder and CEO of a document automation vendor; James Tapscott, who ran drafting experiments at his law firm; and Redmond Lee, who described his vendor’s AI build tools.
Overview
Document automation has always produced consistent first drafts, but the effort of setting it up stalled many firms. Bamford argued that AI now removes much of that burden. The harder question is whether generative AI can simply replace automation, and the panel’s answer was no: the two work best together.
On the build side, AI can review a precedent and suggest the questionnaire, since the information needed is largely the same regardless of a firm’s wording (for a lease: parties, premises, rent review, alterations). Some tools can then build the automation itself, turning weeks or months of work into minutes. Howorth described “headless drafting”, where AI gathers information in a conversation and answers the questionnaire in the background, while the logic engine inserts the clauses. The AI never drafts the document, which reduces review time and token costs. Tapscott’s firm tested pure generative drafting and found about 97% accuracy, but errors ranged from a missing comma to a missing guarantor. Formatting such as cross-references and numbering dropped out, and time to a usable first draft was slightly longer. On a 28-page lease, it also made an average of five unrequested changes, which forces a line-by-line review.
Tapscott’s conclusion, offered as true for today only, is a “sandwich”: AI extracts and analyses information, the automation layer inserts it without touching boilerplate, and AI then adds bespoke clauses. He also found that giving the AI just the options it needed outperformed adding commentary and guidance, and expects prompt libraries to matter less over time. Bamford said automation specialists will spend less time on admin, citing one practitioner who uses a simple AI prompt to sweep a finished automation for spacing, punctuation and grammar errors. Howorth cautioned that the barrier is often organisational: firms hesitate to give fee earners time to build documents, though a couple of hours can make the return on investment easy to prove. Bamford noted that fixed fees for transactional work mean efficiency improves profitability, while clients may bring more work in-house.
Key Takeaways
- Use AI to answer the questionnaire rather than write the document. This keeps clause logic deterministic and avoids reviewing an entire AI-generated draft.
- Judge accuracy by severity, not percentage. A 97% score can hide an omitted guarantor, so ask what the errors would mean in practice.
- Measure time to a usable draft. Formatting repair and line-by-line checks can outweigh the speed of one-click generation.
- Keep inputs simple. Tapscott found supplying the options a model needs worked better than heavy commentary or guidance.
- Prepare a clear answer for the board on why automation still earns its budget alongside a general AI tool: consistency every time, and lower usage costs.
- Redefine the automation role and invest the build time. Use AI for routine checks so specialists focus on logic, workflow design, and integration.