Not If, But How: Preparing International Family Mediation for Multi-Agent AI 

Many mediators already use large language models (LLMs) — ChatGPT, Claude, Gemini, Mistral — to support everyday work, but the level of adoption is still uneven and often depends on individual curiosity rather than shared professional standards. At the same time, we have just received a very clear signal from Europe’s justice sector that generative AI is already mainstream in day-to-day professional work.  

The Council of Europe’s CEPEJ “Guidelines on the Use of Generative Artificial Intelligence for Courts” (adopted in Strasbourg, 4–5 December 2025) reports that 68% of law firms and 76% of corporate legal departments use GenAI at least weekly, with especially high daily use reported in the Netherlands (42%) and Germany (38%). CEPEJ also surveyed its European Cyberjustice Network: 46% of respondents confirmed GenAI use in courts—mainly for summarising documents (47%), hearing minutes and transcription (40%), and information extraction (40%). The Guidelines also stress privacy and sovereignty: prefer customised, secure tools over generic “off-the-shelf” chatbots, and keep data and infrastructure under the competent public authority’s control, including pseudonymisation and strong data-governance rules. This pushes us, as practitioners in the judicial system, away from generic chatbots and toward secure, customised GenAI tools—potentially including local or sovereign deployments—supported by strong governance, pseudonymisation, and clear human responsibility. Link 

But where are we now? The reality for most mediators today is to use mainstream tools – ChatGPT, Claude, Gemini, Mistral – while true ‘AI agents’ and private self-hosted systems are still costly and maintenance-heavy. For most mediation use cases, the ROI is therefore not in building custom agent stacks, but in using off-the-shelf tools responsibly for low-risk, high-value tasks (e.g., drafting, summarising, structuring agendas), with careful confidentiality, transparency and human oversight. 

A helpful reality check comes from the Gartner Hype Cycle for Generative AI (2025) figure Link that shows we’re essentially at the “Peak of Inflated Expectations.” That’s the stage where everyone talks about the technology, expectations are huge, and the success stories are louder than the everyday problems, while practical building blocks like prompt engineering, orchestration frameworks, and domain-specific models are still climbing. We know that only the use cases with clear ROI will survive. 

The real question isn’t whether we’ll use GenAI—it’s which use cases will survive. The train has not left, but it is already pulling into the platform, and the question is whether mediators will be trained conductors or only surprised passengers? Will we be forced into using unsafe tools without proper education and guidelines? Will we be sidelined by legal platforms that do not understand mediation principles and ethics? 

In my view, we should prepare not just for “one AI assistant,” but for a multi-agent future. In other industries—for example in customer service—we already see systems where multiple specialised AI components are coordinated: for example, orchestration layers with policy and guardrails, tools that summarise live conversations and cases, and real-time conversational analytics that can flag keywords to avoid emotional escalation. Link 

Mediation has a clear rationale for a similar division of labour. A mediation-aligned multi-agent setup would separate roles such as: an orchestrator that routes tasks; a confidentiality and compliance gatekeeper that blocks risky prompts or outputs; a neutral summariser and note-taking assistant; an issue mapper that proposes agendas, or open questions; a drafting co-pilot for non-binding wording and checklists; and an escalation monitor that flags rising tension and suggests de-escalation language for the human mediator. These tools will remain just decision-support, with heavy human supervision and accountability staying firmly with the mediator. Link 

So in my opinion, we have to prepare not just for one but for a multi agent future, even if the ROI is not good for mediators. Yet. 

So let’s ask ourselves: what can we do now? How can we prepare our practice today? 

1. Define your own AI boundary policy—what you will use GenAI for (e.g., summarising, structuring, drafting neutral text or agreement drafts, creating scenario tables, etc.) and what you won’t use it for yet because you’re not sufficiently educated or practised. This should be paired with a clear “must not” list—for example: legal advice, outcome predictions, coercive suggestions, diagnosis, or entering strongly sensitive or identifiable data. Your policy will be unique and will depend heavily on your current theoretical knowledge and hands-on experience with GenAI. 

2. Standardize your mediation data structures – AI becomes powerful when your inputs are structured. So create templates, make a MoA skeleton, make a parental plan checklist, make a financial inventory table…etc. Once these building blocks exist, you can reuse them consistently across cases and let AI help you fill, compare, and summarise them without reinventing the wheel each time. 

3. Learn “AI language” (prompt discipline) – You already know how to speak and write neutrally as a mediator. The extra skill is learning how to translate that professional intention into clear instructions a machine can reliably follow. LLMs can score surprisingly high on emotional-intelligence tests Link —this can be misleading, so don’t worry: you’re not the first who asked everyday-life questions from an AI Tool. But it is still a machine—a model, to be precise.  

So when I hear “it’s useless, it hallucinated, it gave a false legal reference…etc” the missing piece is often prompting competence. You have to specify the role, the task, the limits (“must not”), the structure, the verification rules, and a lot more. Getting comfortable with that may take dozens of hours of learning (often 50–70 hours), but it makes the difference between random chatter and an AI-assisted professional workflow. 

4. AI adoption is a team sport. Join—or build—a peer group and create a safe-to-fail environment for experimentation. Pick one low-risk use case to trial—for example, producing a mid-session neutral summary, drafting an agenda, or generating three options for a parenting schedule. Then evaluate the output and share what you learn with your peers. R&D takes time and capacity, and the AI landscape is too large for exploring, testing, and implementing it alone. Prompts, templates, safety checks can be shared; failure modes can be documented; and best practices must become repeatable. Done well, this builds a mediation team that becomes more tech-savvy and more adaptive over time. Nurturing an AI-ready culture means continuous upskilling, so a peer group helps you stay current. 

5. Build a “multi-agent mental model” – Start thinking in specialised roles, not “one AI assistant.” Even with off-the-shelf tools, you can simulate this separation by using separate prompts, separate checklists, and different output formats for each role—so a “drafting” request never gets mixed with “process” or “numbers,” and your risk controls stay intact. In practice, you remain the orchestrator—the human who assigns tasks, checks quality, manages compliance and confidentiality, and decides what (if anything) enters the negotiation room. 

A mediation-ready “agent team” could look like this: 

• Orchestrator (you): routes tasks, sets boundaries, verifies, and makes all professional judgments. 

• Compliance gatekeeper: applies your red lines (no legal advice, no unsafe suggestions, no sensitive-data leakage) and forces transparency and consent checks. 

• Live note-taker and neutral summariser: produces short, balanced summaries of positions and proposals  

• Process and de-escalation co-pilot: flags rising tension and suggests neutral micro-interventions, reframes, or breaks—always as optional prompts for you. 

• Jurisdiction and terminology clarifier: helps translate terms across systems (e.g., custody/parental responsibility concepts), highlights where meanings differ, and suggests questions to confirm parties’ understanding. 

• Parenting plan specialist: generates structured parenting schedule options and “what to decide” checklists (holidays, travel, handovers, school, communication, contingencies). 

• Child support and budget specialist: drafts budget tables, scenarios, and assumptions for discussion (with clear caveats and verification). 

• Assets, tax & property-division specialist: produces asset inventories, disclosure checklists, and option tables for dividing property across jurisdictions—while also flagging tax-sensitive touchpoints for follow-up with qualified advisors (e.g., capital gains exposure, transfer taxes/stamp duty, pension taxation, residency/domicile implications, and timing effects). It can help structure “what to ask” and “what to verify” (valuation dates, liquidity, hidden liabilities, cross-border enforcement, and documentation), but it does not provide tax or legal advice—only structured support for your discussion and your experts’ review. 

• Agreement drafter: turns the parties’ agreed points into plain-language clauses and a coherent draft for human review. 

Start small: pick one low-risk use case, test it in a safe-to-fail way, and document what actually worked and what didn’t. Turn the wins into repeatable assets—prompts, checklists, templates, “red line” rules—and share them with peers so the whole community learns faster and more safely. The mediators who build ethical, reliable workflows today will be ready for multi-agent systems tomorrow—not because they chased hype, but because they built professional standards early, and won’t be pushed into adopting someone else’s tools on someone else’s terms. 

References 

Agrawal, G., De Maria, R., Davuluri, K., Spera, D., Read, C., Spera, C., Garrett, J., & Miller, D. (2025). 

  Redefining CX with agentic AI: Minerva CQ case study (arXiv:2509.12589). arXiv. 

Chandrasekaran, A. (2025, July 29). 

  The 2025 Hype Cycle for GenAI highlights critical innovations. Gartner. 

Coshow, T., & Zamanian, K. (2025, December 18). 

  Multiagent systems: A new era in AI-driven enterprise automation. Gartner. 

European Commission for the Efficiency of Justice (CEPEJ). (2025, December 19). 

  Guidelines on the use of generative artificial intelligence for courts (CEPEJ(2025)18Final). Council of Europe. 

Schlegel, K., Sommer, N. R., & Mortillaro, M. (2025). 

  Large language models are proficient in solving and creating emotional intelligence tests. Communications Psychology, 3, Article 80. 

Wolters Kluwer. (2024, October 24). 

  Wolters Kluwer’s 2024 Future Ready Lawyer Survey: Legal professionals confident in managing AI-driven changes to business of law. Wolters Kluwer. 

Blanka Illés

Dr. Blanka Illés is an international family lawyer and mediator with 26+ years of experience in complex, multi-jurisdictional family disputes, specialising in cross-border matrimonial property division and high-asset settlements. She graduated cum laude from Eötvös Loránd University (Budapest), founded her own law firm focused on international family and inheritance matters,… MORE