The case for court-hosted AI family mediation in Canada

Log in to listen to this article

When a family unravels, the house must be divided, the support calculated, and the children’s lives rebuilt around two homes. And before any of it can gain traction, the waiting begins.

Months pass before anyone sits across from a judge in court. Months more go to sworn financial disclosure, much of it producible in an afternoon, had the parties still shared an address. If they cannot agree, a trial awaits, often years away. None of this is exceptional. In fact, it’s the system working exactly as designed.

The delay is not a problem of legal complexity. Most separating families do not need a judge to decide their future; they need structure, time, and information about what the law says, what a court will likely do, and what options exist. Give them that, and most parties will settle. The bottleneck is not the law; it’s the supply of neutral human time, which is finite and measured in appointments, overburdened schedules, and waiting lists.

This bottleneck need not be permanent. It can be cleared, but so far, no jurisdiction has built the system to do it.

The gap 

Last year, more than 278,500 family law cases were active in Canada’s civil courts—nearly a third of the caseload. By some counts, as many as 80 per cent of the people in them had no lawyer. Contested matters over children or property routinely cost people tens of thousands of dollars and involve many court appearances, well short of a trial. The deepest cost falls on the children, who fare better the less conflict their parents’ separation produces.

We have already decided, as a country, that this is wrong. A Saskatchewan statute now requires that separating families attempt dispute resolution before a contested matter proceeds. In British Columbia and Alberta, it’s required by court rules. While mediation is voluntary in Quebec and Ontario, couples are required to attend mandatory resolution information sessions before litigating. 

The principle — that structured resolution belongs at the start and not the end — is no longer in dispute. What defeats it is rationing. A resolution can’t always be reached, limited by the hours in a working day and, in places, walled off from property and pensions.

The components exist; the assembly is missing 

The engine that guides couples to agreement is proven, but partial. In Australia, National Legal Aid, with government support, developed a low-cost, AI-assisted online dispute resolution tool to help separating or divorcing couples agree on parenting arrangements and property division without going to court. Since launching in 2020, amica has helped more than 17,000 couples work through property and parenting. Drawing on what was developed down under, ODR.com launched an AI-assisted family dispute resolution platform in 2024, which is now marketed to family courts in Canada and the United States.

At the People’s Law School in British Columbia, Beagle+ answers everyday legal questions, including about family law, from a closed, curated corpus of vetted provincial legal content rather than the open internet. It informs rather than mediates, but proves that carefully sourced legal AI is already working for Canadians. The spousal-support ranges lawyers calculate with software, not by hand, are already built into the tools they use daily. Document assembly and case management exist, too.

These are important bits, but what no one has built is the whole. The Netherlands came closest, creating a system that produced sound, court-ready agreements, but it died when it had to fund itself as public support fell away. England requires couples to consider mediation before court, but has built no public system to deliver it. 

The standards for AI in dispute resolution are only now being written, at UNCITRAL and among international bodies. But there is still no mandatory, court-hosted, publicly funded, covering an entire separation, open to every family at no cost. That is for Canada to build, nationally, not in patches.

This is not the AI you have been warned about

When our courts sanction lawyers for citing cases that do not exist, they are sanctioning open-internet language models generating text from unverified sources. A 2024 Stanford study found that such models hallucinate on the majority of legal questions. The proposed system is a different architecture: it would retrieve from a closed, verified corpus of statutes, guidelines, and decisions rather than inventing. It would cite a source for every claim, and be unable to conjure a case not in its index, which is the retrieval design Beagle+ already uses. It would present the same calculation to both spouses, thereby dissolving the information asymmetry that underpins unrepresented litigation. The real hallucination risk is the status quo, where self-represented parties already turn to consumer chatbots with no guardrails. This proposed design narrows that risk; it does not create it.

How it would work

Legislation would make structured AI mediation a step in every proceeding; a judge would decide when it applies; and urgent matters would go straight to the bench. Each spouse would complete a sworn intake alone, with safety screening, under the same penalties as any financial statement. The system would lock in what the parties agree on and narrow the rest to what genuinely remains, usually less than either feared. 

On each open question, it would state the law and the outcome a court could realistically reach—the same figure for both sides—so each knows what litigation would cost and what it would take. It could also propose options drawn from comparable outcomes. When agreement is reached, it would draft a complete separation agreement, every clause anchored to the law behind it. Either party could retain a lawyer in the background throughout, to review every proposal — a human supervisor watches for trouble. 

If the process failed, it wouldn’t tip the parties back into the queue, but instead hand them to a judge already holding the full record of what was agreed and what remains — a swift, paperless process. The judge would then weigh all options to finalize the matter, including offering advice or guidance. No hearing would be required, but could be called. No existing system does that.

The hard objections, answered

The most serious concern is not technological; it’s intimate partner violence and coercive control, which seasoned professionals miss, and no checklist reliably catches. The proposed design does not pretend to solve what cannot be solved. The visible danger is caught by structured screening, applied more consistently than the human kind. The subtler danger is watched by a human supervisor alert to the process being turned into a weapon. And one principle governs without exception: wherever a party signals a safety concern, by any route, the mediation stops, and that party’s own judgment about whether to go on prevails. Participation is never required where safety is in doubt.

Other concerns have answers. 

• Algorithmic bias: A model trained on past outcomes can inherit the disadvantages embedded in them, so any deployment must be trained against that bias and audited publicly, broken down by gender, income, and region. 
   
• Privacy: Family data is among the most sensitive there is—an argument for a court-administered trust rather than a commercial product, with Canadian data residency and contractors who never touch a party’s file. 
   
• Indigenous jurisdiction: Where a matter engages Indigenous child and family law, the standard design may not apply at all; co-design with First Nations, Métis and Inuit governments is a precondition for launch, not an afterthought. The digital divide demands funded, staffed places for those who cannot manage on their own. The Constitution is no barrier, and mandatory pre-trial steps are nothing new, from the mediation Ontario has regionally required in civil cases since 1999 to the family information programs now mandatory in several provinces. Human mediators lose nothing as the system carries the families they never had time to see and returns the hardest cases to them.

What should happen now?

This is national work, and the machinery to coordinate it is already in place. The federal, provincial and territorial justice ministers should agree on a single standard for governing court-hosted family data. One province—British Columbia, the obvious first—should run a court-integrated pilot on today’s technology, with no new legislation needed to begin. The law societies should settle what independent legal advice means for an AI-mediated agreement. Quebec’s Cyberjustice Laboratory, 15 years into this work, should anchor the evidence. Indigenous co-design should start now, in parallel, concluding before anything goes live. And the service should be free where families meet it, paid for out of the court time it saves, promised openly at the outset.

No country has built the whole of it, made it national, and made it the default. Canada can, and Canada, almost alone, already holds every piece: the provincial mandates, the proven technology, some of it already built for Canadian law, the institutions to move in concert, and the constitutional footing to require it.

The unrepresented majority who go to family court are not waiting for perfection. They are waiting for what represented parties already have: the same law, the same numbers, and the same honest account of what a Court would do.

The technology is here. The architecture allows it. What is missing is not capacity, or knowledge, or means.

What is missing is the decision to lead.