Global Mediation

Global Mediation has been mediating in the United Kingdom since 1999. It began with a simple idea: that people deserve better ways to resolve conflict. Twenty-five years on it delivers thousands of mediations a year for local authorities, businesses and individuals, from a London head office with a national team. It works across community, civil and legal, workplace, business, family and SEND disputes, and trains mediators too. It is accredited by the Civil Mediation Council, and holds ISO 9001, ISO 27001 and Cyber Essentials. It reports that 95% of its mediations reach full or partial resolution.

SEND mediation is the part of the business this work touched. When a family and a local authority disagree about an Education, Health and Care Plan, a school placement or the support a child receives, mediation is how they try to settle it before a tribunal. Getting the meeting in the diary is its own job. A date has to suit a parent, the child or young person, the local authority officer, the school's SENCO and one of Global Mediation's mediators, and it has to be found while a family is already waiting on a decision about their child's education.

Bitcube built the engine that does that, from January 2024.

Setting up a mediation is a negotiation, not a lookup. Availability arrives one reply at a time over days, from five kinds of participant who do not know each other's diaries, and the engine had to run that exchange to a confirmed date or hand it to a person having tried. It could not work to one rulebook either. Global Mediation's own policy sets the notice periods, the rescheduling limits and the charges, and each local authority area adds rules of its own on top, so the rules had to be data the business could change rather than logic in a release. The cases themselves lived in Global Mediation's case management system, which had to stay the single record while the engine held the state of a negotiation that system knew nothing about. The correspondence was written by a third-party AI service that returns free-form text, so its output had to be constrained before anyone would call it an email — and the people receiving those emails were parents, children and young people in the middle of a dispute about a child's education, where a wrong or badly timed message is a harm rather than a defect. None of that could be proved by unit tests. A conversation running for days across five parties had to be demonstrated end to end, repeatedly, before a live case went near it.

Alongside the main line, a separate console harness was built to replay scenarios against the assistant on its own, and the behaviour tests were carried into our shared end-to-end test repository when the in-project Playwright suite was retired, so they outlived the project they were written for.