From Documents to Artifacts

1. The trap of individual acceleration

The first wave of AI adoption in legal work is individual. A lawyer opens a chat, gets a draft in minutes instead of hours, and feels the speed-up. That speed-up is real — and it is exactly what creates a new systemic problem.

Legal work is a relay race, not a sprint. A lawyer's memo goes to the head of department, an opinion to a committee, a memorandum to a partner, a draft decision to a judge. And at every handover of the baton the "person plus their chat" model breaks down: the recipient sees the final text and nothing more.

Text production has sped up by an order of magnitude. Verifiability has stayed manual. A team in which everyone has accelerated individually but still exchanges work the old way has not become faster — it has become riskier: more drafts, less confidence in them, and the same reviewing link in the chain.

2. What is lost when a document is handed over

A finished document is a conclusion with the evidence wiped off. The recipient cannot quickly establish:

• which sources were used and which were discarded;
• which versions of the legislation were analysed, and as of what date;
• which research queries were run;
• how facts and rules were matched;
• which alternative hypotheses were considered and why they were rejected;
• which limitations and uncertainties were identified;
• which checks were actually performed;
• why this particular conclusion was reached — and who is responsible for it.

Hence three chronic losses in team work: loss of context (each next participant reconstructs the picture from scratch), false confidence (smooth text masks unverified grounds) and diffused responsibility (it is unclear what a human checked and what was generated).

The classic response — "write more detailed memos" — does not work: it increases the volume of text without making it verifiable.

3. Changing the unit of exchange

The proposed approach is not an abandonment of documents but a refusal to treat the final document as the sole and principal object of exchange.

The document is the interface. The work package is the legal object.

The final text remains — people find text convenient to read. But it is handed over inside a structured work package (Legal Work Package, LWP), suitable both for human reading and for processing by agentic systems. The package holds the result together with its grounds: sources, facts, reasoning, checks, history and open questions.

This is no novelty alien to law. A lawyer is already obliged to separate established facts from asserted ones, to cite sources and to record the applicable versions of the rules. The artifact approach is the digitisation of a procedural discipline the profession has known for centuries, into a form that machines can verify.

4. The chain of provenance of a conclusion

Every material conclusion in the package carries a verifiable chain:

Source → fact extraction → interpretation → matching → intermediate conclusion → final statement.

In terms of the VeriLex data model: source → evidence → fact → rule → issue → conclusion → decision_log → handoff.

A point of principle: this is not about preserving the model's "train of thought". An LLM's hidden reasoning is legally useless and non-reproducible. What is recorded are verifiable working traces: which sources were opened, which fragments were extracted, which statuses were assigned to facts, which matching rules were applied, which checks were run and what they showed. This is the same logic as in the W3C PROV data-provenance standard and in the culture of reproducible scientific research — carried over to legal work.

The invariants of the chain (machine-checkable):

• a fact with the status established must have at least one piece of evidence and a human confirmation;
• a conclusion with no facts or evidence attached cannot hold "strong" status (check_unsupported_claims);
• a legally significant decision reaches confirmed status only when the "who confirmed" field is filled in — and that is always a human.

Hence the two principles that hold the whole construction together:

No source — no strong conclusion. The agent proposes — the human decides. Always.

5. Checking the process instead of re-reading the text

When a colleague receives the package, their agentic environment checks it automatically: the completeness of the source register; the currency of versions as of legally significant dates; the letter-for-letter correspondence of quotations to primary sources; the connectedness of "fact → conclusion" chains; the presence of the mandatory critique package (weak points, counterarguments); changes relative to the previous version; adherence to the internal research standard.

Comments come back not as a note saying "take another look" but as a structured critique package that the author addresses point by point. The team moves from the question "is this text written correctly?" to the question "how was the information it rests on gathered, verified and analysed?".

We pass on not only the answer. We pass on the verifiable process by which it was reached.

Team efficiency arises when the agents of different participants can understand and verify each other's results. That is the definition of the team effect in the agentic era.

Importantly, the agentic environments of both participants are ready-made tools (OpenAI Codex, Anthropic Claude Code or open-source analogues), not purpose-built software. The team layer is created on top of them by the artifact standard, instructions and skills — that is, by files that can be published and put to use today.

6. Sufficient transparency instead of total logging

The main risk of the artifact approach is degeneration into a bureaucracy of logs. The defence is the principle of proportionality: three levels of detail.

Basic (everyday work): the final text, a list of sources, key facts, a short log of checks, open questions. Overhead — minutes.

Extended (complex research): plus the structure of the argument, versions of the legislation, alternative positions, an evidence table, a graph of connections, a detailed verification protocol.

Audit (decisions with a high cost of error): plus the full change history, an action log, model and instruction identifiers, control hashes, participant roles, human confirmations — a reproducible research package.

The measure of an artifact's usefulness is not the number of logs but its participation in real checks and decisions. Anything that is read neither by a human nor by an agent is surplus.

7. Team memory: from archive to asset

Structured artifacts accumulate into an institutional memory with different properties from a network folder of documents.

Faced with a similar task, an agent: finds the past research; determines which of its parts are still current; checks changes in the legislation since the date of the previous analysis; reuses confirmed sources; compares new facts with old ones; builds an updated version and shows exactly what has changed.

Every completed project becomes a reusable intellectual asset. For Ukrainian institutions with high staff turnover, this is a mechanism whereby a specialist's departure no longer means the departure of expertise.

The manager gains a new management interface: a short conclusion + a map of grounds + risks + alternatives + a degree of confidence + a list of unconfirmed assumptions + questions that require a human decision. Facts, interpretations and recommendations stop being glued into one inseparable text.

8. The success criterion

Pilot No. 11 (a working group on full artifact exchange) tests five measurable quantities:

Metric	What it measures
Source coverage	Share of statements linked to sources
Unsupported claim rate	Share of "strong" conclusions with no grounds
Handoff completeness	Completeness of handover packages (accepted without return)
Time to review	Time to check someone else's result
Reproducibility score	Recoverability: from which input versions any conclusion was assembled

The hypothesis: time to review drops several-fold with only a slight increase in preparation time; the unsupported claim rate tends towards zero. If the pilot does not show this, it is the artifact standard that is to be simplified — not the team that is to be coerced.