A personal AI workstation for every lawyer
National legal AI infrastructure can be built not as one heavy product, but as a network of adaptable workstations unified by quality data, shared standards and a professional community.
A concept for judges, government lawyers, advocates, universities and digital transformation teams. Author — Sergej Avdejcik, VeriLex project.
- Ready-made infrastructure — start without development
- Models at the level of domain experts
- 120+ million court decisions in open access
Why “ordering one big system” no longer works
Expensive.
Tens of millions before the first result.
Slow.
Years of approvals and procurement — while the technology changes every six months.
Averaged out.
The requirements of thousands of specialists are reduced to “average” scenarios that fit no one.
Hard to change.
Every change goes through a contractor and a new budget.
Dependency.
On a single vendor, a single platform, a single contract.
Obsolete at launch.
The system is designed around the technology of two years ago.
OECD, analysis of 200 AI deployments in the public sector: most initiatives get stuck at the pilot stage.
Three shifts that opened a window of opportunity
Language models learned to work with legal text
Independent blind benchmarks: legal AI tools solve research tasks at 74–78% versus 69% for human lawyers. Frontier models match domain experts on well-specified tasks.
Working environments appeared — not just chats
Agentic environments read case folders, follow written instructions, use tools, keep memory and check their own results. The AGENTS.md instruction format is an open standard supported by tens of thousands of projects. This is ready-made, industrially proven infrastructure: there is nothing to develop.
Data and tools connect securely
Open protocols connect the model to legislation, registries and case materials at the level of a specific workstation — without sending data to third-party clouds.
The model is a reasoning engine
Not an oracle. Not a knowledge base. Not a replacement for the lawyer. An engine that needs fuel and a driver.
The model is trained on vast corpora of text, including legal text. So it understands the structure of legal argument, the language of legislation and standard legal constructions. But it does not “know the law” — it reasons over the data it has been properly given.
A language model has a powerful general-purpose mechanism for language and reasoning — but it needs current, verified and properly organized data.
A strong model is not yet a legal assistant
Outdated editions.
The model learned from a data snapshot that was already stale on release.
Blind to the case file.
Until it gets controlled access to the materials, it reasons in the dark.
Mixes sources.
The training data contains different versions of the same acts.
No citation guarantee.
A plausible reference is not the same as an existing one.
Ignores institutional rules.
Your regulations, positions and standards are unknown to it.
Carries no responsibility.
The model does not decide when human review must step in.
58–82%
hallucination rate of public LLMs on legal questions (J. Legal Analysis)
17–33%
error rate of commercial legal RAG tools (Stanford)
1,733
court cases involving fabricated AI citations across 40 jurisdictions (July 2026)
The formula for a full legal copilot
Click a component to see what it contributes.
The main effect comes from the synergy of a general-purpose engine and properly organized data.
Five layers of data — different access rules
Public law
The Constitution, codes, statutes, case law, registries. Open to everyone, machine-readable.
- Access mode
- Open access
- Responsible
- State
Public law
The Constitution, codes, statutes, case law, registries. Open to everyone, machine-readable.
- Access mode
- Open access
- Responsible
- State
The state answers for layer 1. The institution — for layers 2–4. The practitioner — for layer 5. Nobody has to build everything at once.
We hand over not just the answer. We hand over a verifiable process of reaching it.
Split view: a single .docx file on the left unfolds into a seven-folder work package on the right.
Traditional exchange
- sources used
- rule editions
- discarded hypotheses
- completed checks
The recipient sees only the outcome. Sources, rule editions, discarded hypotheses, completed checks — everything stayed with the author. Review means redoing the work.
Agentic teamwork
- result/
- sources/
- facts/
- analysis/
- verification/
- logs/
- metadata/
The recipient sees the result — and its grounds: sources, facts with statuses, the logic, the checks, the uncertainties, the change history.
The document is the interface. The work package is the legal object.
Open formats: Markdown, JSON, YAML, Git. The package is readable by a human and checkable by an agent.
Every conclusion knows its origin
Civil Code of Ukraine, Article 651, edition of 1 May 2026, verified on 8 July 2026 (stable link).
These are not the “model's thoughts” — they are verifiable working traces: sources, statuses, checks. Like scientific reproducibility standards, but for legal work.
Team effectiveness emerges when the agents of different participants can understand and verify each other's results
Review stops being a repetition of someone else's work and becomes an audit of the process. Comments stop being “please take another look” emails and become targeted. Experience stops leaving with the employee and stays in the packages.
What a lawyer's workstation looks like
This is configuration, not development: the workstation is assembled from plain-text instruction files on top of existing agentic environments (OpenAI Codex, Anthropic Claude Code, open-source equivalents). Deployment can start today.
A stylized mockup, not a screenshot of an existing product.
Starter configurations — open and free
legal-copilot-ukraine/
- The foundational instruction file: how the agent works, cites and verifies in this repository. AGENTS.md is an open standard supported by tens of thousands of projects.
AGENTS.md
The foundational instruction file: how the agent works, cites and verifies in this repository. AGENTS.md is an open standard supported by tens of thousands of projects.
5–10 ready configurations: municipal lawyer, judge's assistant, advocate's assistant, legislation expert, contract work, EU-law harmonization, a student workstation. Download, install locally, and get a working workstation within an hour. Everything runs on top of existing agentic environments — the repository contains no custom code: only instructions, roles, skills and training courses.
Thousands of specialists develop the system — the best solutions return to everyone
Base configuration
Adaptation by the user
Skill improvement
Artifact creation
Community review
Best practice returns to the repository
Each workstation becomes a reflection of one practitioner's experience. The instruction written by the best lawyer in the department becomes a file that works for every colleague — and survives staff rotation. No centralized contractor can evolve at this speed.
The state builds conditions, not a product
The state provides
- machine-readable law (stable identifiers, editions, dates)
- data and artifact standards
- security requirements
- certification of configurations
- pilots in institutions
- training
- interoperability
- support for the open ecosystem
The state does NOT
- design a single “perfect Legal AI product”
- pick a single vendor
- replace professional judgment with an algorithm
The task of the coordinating center is to create conditions in which many quality solutions emerge, get verified and evolve quickly.
Technology spreads at the speed people learn
Universities
- develop legal skills for agents
- test models on Ukrainian legal tasks
- build teaching cases on open data
- train students
- set standards of legal AI literacy
- evaluate the quality of artifacts
Mentors
- help formalize processes
- teach the user instead of configuring for them
- instill the practice of verification
- teach teamwork with artifacts
- gradually hand full control to the practitioner
The ability to organize the work of a legal agent and hand colleagues a verifiable artifact is a core professional competence of the coming decade.
Year one: a controlled chain reaction, not a megaproject
Months 0–3
- Open repository
- 5–10 configurations
- A minimal artifact standard
- Training materials
- Selection of pilot institutions
Months 3–6
- Workstation rollouts
- A team artifact-exchange pilot
- Mentor training
- Quality evaluation
- Feedback collection
Months 6–12
- Scaling up
- New legal specializations
- Interoperability standards
- A national network of practice
Day one is the publication of the repository: the technology base already exists, no development is required. The timeline is a working hypothesis, to be refined with pilot institutions. The speed benchmark: Diia.AI's path from strategy to a working service.
Centralized SaaS versus a personal agentic ecosystem
Before
Requires years of custom development
After
Runs on existing agentic environments today
Before
A single averaged product
After
An individual workstation
Before
A long development cycle
After
A fast start from configurations
Before
Changes go through a contractor
After
Changes made by the user
Before
Exchange of final files
After
Exchange of verifiable artifacts
Before
A hidden preparation process
After
A reproducible chain behind the result
Before
Centralized scenarios
After
Distributed skills
Before
High cost of evolution
After
Evolution through the community
Before
Platform lock-in
After
Freedom to switch models
Before
Limited personalization
After
Accumulated personal and team experience
We know the risks — and design the defenses from day one
Verified data
Source binding
Logging
Access control
Testing
Human confirmation
Quality metrics
Local deployment
Disclosure levels
Confidentiality rules
The compliance frame: EU AI Act (human oversight), the CEPEJ principles of the Council of Europe, the UNESCO guidelines on AI in courts, and the guidance of the Ministry of Digital Transformation and the Ministry of Justice for legal professionals.
Don't start with a national megasystem. Start with ten quality workstations, open data, a shared artifact standard and a community of professionals.
The technology base already exists: the agentic environments OpenAI Codex, Anthropic Claude Code and open-source equivalents work today. Launch means publishing instructions, skills, roles and training courses — not developing a product.
Author of the concept — Sergej Avdejcik, VeriLex project. We answer every message.