Sid Ali Boutellis, a freelance legal technology professional with over ten years of experience and newly acquired legal credentials, moderated this EU AI Act Forum session with Alexander Deicke, Legal Interim Manager at K11 Consulting GmbH. Deicke trained as a lawyer and now supports organisations with AI governance and strategy, including banking sector clients in Germany. Boutellis brought a hands-on developer’s view of building with AI agents, while Deicke offered an advisory perspective on how organisations are adopting and governing them.

Overview

The speakers asked what the shift from prompt-by-prompt tools to AI agents means for legal practice by 2027. Deicke observed that early reluctance to accept governance has given way to demand for guidance, yet most companies still struggle to move from use case to business case. Boutellis noted that agents deliver faster output on tasks such as due diligence, but at the cost of visibility into what happens behind the scenes.

Both stressed that shortcuts carry risk. Boutellis explained that a workable application can take five hours or twelve depending on how much time goes into security, scalability and documentation that makes the system explainable. Deicke argued that work is shifting from drafting towards investing time at the start and checking quality at the end, and that humans tend to avoid reviewing their own output. His example was a lawyer asking AI to change one clause in an SPA and finding other clauses altered because the instruction had not been limited. On data, both urged a single source of truth, such as one agreed M&A form rather than each partner’s preferred precedent, or a structured matter intake as the starting point.

On regulation, Deicke noted that the AI Act is based on products rather than processes, so organisations need an inventory of the tools they use, risk classification and trained staff. He expects governance to become increasingly tool-supported but not tool-only. Looking ahead, the discussion suggested that technology spend no longer differentiates firms, and that clients still want commercial, trusted advice. Deicke reported firms banning first-year associates from using AI to protect their training; Boutellis countered that this risks losing talent. Both favoured a middle path: limited use for defined tasks, quality checks and senior partner oversight.

Key Takeaways

• Do not mistake speed for readiness. Time spent on constraints, security and documentation is what makes agent-built systems explainable and resilient, so leaders should resist pressure for shortcuts.
• Build quality control into the workflow. People avoid checking their own work, so assign independent responsibility for reviewing AI output, as the automotive industry does with dedicated quality departments.
• Fix the data before the tools. Agree on one authoritative precedent per document type and start from structured matter intake, so AI is not trained or prompted on inconsistent versions.
• Constrain AI instructions precisely. Unrestricted prompts can lead to unrequested changes, so specify exactly what may be altered.
• Inventory your AI tools to meet AI Act obligations. Because the Act is product-based, knowing what is in use, classifying risk, and training staff comes before any compliance technology.
• Train juniors with guardrails rather than bans. Limiting AI to defined tasks with senior oversight protects professional development without sacrificing the efficiency and talent that AI fluency attracts.