Executive Summary
A national Legal Copilot ecosystem for Ukraine
The problem. The volume of legal work Ukraine must complete in the coming years (harmonisation with EU law, justice reform, reconstruction) far exceeds the capacity of its available specialists: the courts have 2,250+ vacant judicial posts, and caseloads run five to ten times higher than the European average. Hiring the missing lawyers is impossible — the productivity of those already working must be raised several times over.
The window of opportunity. By 2026, three technological preconditions have converged: (1) flagship language models (GPT-5.5, Claude Opus 4.7, Gemini 3) have reached the level of domain experts on legal tasks — independent Vals AI evaluations show 74–78% against a 69% attorney baseline; (2) agentic environments (Codex, Claude Code, open-source analogues) make it possible to build complete workflows configured through text instructions, with no programming; (3) Ukrainian legal data is already open: 120+ million court decisions and a machine-readable legislation API.
Why not "just ChatGPT". A model without current data, case materials, instructions and verification produces plausible text rather than verified text: public LLMs hallucinate on legal questions in 58–82% of cases, and the body of court cases involving fabricated AI citations reached 1,733 (July 2026). The formula for a genuine copilot: model + legal data + task materials + instructions + verification + a qualified user.
Start today, with no development. The atomic unit of the system — a personal agentic workstation — is built on infrastructure that is already available and running in production: OpenAI Codex, Anthropic Claude Code or open-source analogues. No custom product needs to be built: to launch, it is enough to publish sets of instructions, roles, skills and training courses in an open repository.
The key decision. National Legal AI infrastructure is built not as a single centralised SaaS (expensive, slow, averaged-out, obsolete by the time it launches) but as a network of personal agentic workstations connected by machine-readable data, open standards and a professional community. The state centralises only what no one else can solve: data, standards, security, certification and training.
The team effect. Moving from individual to team efficiency requires changing the unit of exchange: not the final document but a structured work package of artifacts (Legal Work Package) — the result together with its sources, facts, reasoning, checks and history. Every conclusion carries a verifiable chain of provenance: source → fact → analysis → matching → conclusion. The recipient's agent verifies the package automatically; verification turns from repeating the work into auditing the process. Two principles: "no source — no strong conclusion" and "the agent proposes — the human decides, always".
The launch mechanism. A managed chain reaction instead of a multi-year reform: an open repository with 5–10 starter workstation configurations (local-government lawyer, judicial assistant, regulatory-act expert, EU harmonisation, and others), a minimal artifact standard, a network of mentors, university programmes and 3–5 pilot institutions with measurable impact.
The first-year plan. Months 0–3: repository, configurations, artifact standard, selection of pilots. Months 3–6: launch of the workstations, a pilot for team exchange, training of mentors. Months 6–12: scaling, new specialisations, a national network of practices.
What the start requires. A small team, 1–2 institutional partners, access to already-open data and support from international technical-assistance programmes. The cost of starting is incomparable with the budget of any centralised system.
Full text: programme article · Evidence base: theses and sources