Legal AI in 2026: what is actually changing for Belgian lawyers

In one year, legal AI has moved from conference gadget to working infrastructure. 69% of legal professionals surveyed for the 8am 2026 report use generative AI, up from 31% a year earlier. Over the same period, courts worldwide have issued close to 2,000 decisions flagging AI-hallucinated content, and the European AI regulation passed its 2 August 2026 milestone.

For a Belgian lawyer, the real work this autumn is sorting. The trends that matter are not the keynote ones: they are the ones with dates and sanctions. We kept five: adoption crossing the majority line, the partial entry into application of the AI Act, courts sanctioning hallucinations, the shift from chatbots to agents, and the deontological framework taking shape at the Belgian bars.

Every section relies on dated, verifiable facts. No predictions for 2030.

Prerequisites before evaluating a legal AI tool

Before comparing solutions, four points to validate internally. Without them, any trial will stall.

• Basic AI literacy in the team: knowing what a language model is, what it does well and where it makes things up. Article 4 of the AI Act made this an obligation in February 2025, since softened into an obligation of means.
• A map of your data: which documents may leave the firm, to which hosting, under which contract (GDPR, data processing agreement).
• A realistic budget: count a few dozen euros per month per lawyer for the public plans of specialised assistants, more for quote-based licences.
• A written framework: an internal usage charter and, if the firm wishes, a client information clause. Belgian templates exist, we come back to them below.

Legal AI adoption tipped over in a single year

The surveys published in 2026 converge: individual use of generative AI has become the majority among legal professionals. The 8am 2026 report (over 1,300 respondents) measures 69% of users, against 31% in 2025 and 27% in 2024. Wolters Kluwer's Future Ready Lawyer 2026 survey (810 lawyers) counts 92% of respondents using at least one AI tool daily, and 62% estimating they save 6 to 20% of their working week.

Two caveats. These figures come from American and international surveys: no published equivalent exists for the Belgian bar. And the gap between studies comes down to the question asked: "using an AI tool" says nothing about depth of use.

The figure with the heaviest consequences for business law firms sits elsewhere. According to the ACC/Everlaw survey, generative AI adoption in corporate legal departments doubled in one year, from 23% to 52%, and 64% of those teams expect to depend less on outside counsel. The client is equipping itself faster than its lawyer.

The governance gap remains: 43% of firms surveyed by 8am have no AI policy and no plans to write one, and 54% offer no training. Individual use is running ahead of internal rules.

AI Act: what applies since 2 August 2026, what has been postponed

Since 2 August 2026, the transparency obligations of Article 50 of the AI Act apply: a chatbot must tell the user they are talking to an AI, and generated content must carry machine-readable marking. On the same date, the European AI Office received its supervision and sanction powers over providers of large models, with fines of up to 35 million euros or 7% of worldwide turnover.

The rest of the calendar moved at the last minute. The so-called "AI omnibus" regulation, published in the Official Journal on 24 July 2026 and in force three days later, postpones the obligations for high-risk systems under Annex III (recruitment, credit, education, biometrics) to 2 December 2027, and those for systems embedded in regulated products to 2 August 2028. It also softens AI literacy into an obligation of means: the firm must take documented training measures, not guarantee an individual level.

timeline
    title AI Act application calendar, 2026 omnibus version
    2 February 2025 : Prohibited practices, AI literacy
    2 August 2025 : Obligations for general-purpose models
    2 August 2026 : Transparency (Article 50), AI Office sanction powers
    2 December 2026 : Marking of content from pre-existing systems, new prohibitions
    2 December 2027 : High-risk systems under Annex III
    2 August 2028 : High risk embedded in regulated products (Annex I)

For a law firm, the reading is simple as the texts stand: assisted documentary research and drafting support do not fall under Annex III high risk. The obligations already touching lawyers are AI literacy and, for those exposing a chatbot to their clients, Article 50 transparency.

The concrete trap: invented case law now costs money

Symptom. Well-built submissions cite judgments with plausible references: court, date, docket number. The judge or opposing counsel looks the judgment up in JUPORTAL or EUR-Lex. It does not exist.

Diagnosis. A general-purpose language model produces statistically plausible text. Without grounding in a documentary corpus, it fills the gaps with references that look like real ones. The AI Hallucination Cases database records 1,996 court decisions worldwide in which a court found hallucinated content, including 8 in Belgium, as of 31 August 2026. In a judgment discussed by La Tribune of AVOCATS.BE, the Antwerp Court of Appeal fined a lawyer 2,500 euros for abuse of process after he relied on "non-existent case law and legal sources invented out of whole cloth", plus 7,500 euros in damages per opposing party.

Fix. Two measures, deployable in a week. One: every reference cited in a submission is verified in the official source (JUPORTAL, Justel, EUR-Lex) before filing. A few minutes per reference, built into the review workflow. Two: prefer tools that cite the exact passages each answer relies on, which turns verification into a click rather than a blind search.

From chatbots to agents: the technical maturity of 2026

In 2026, the reference architecture of legal tools is no longer the bare chatbot but retrieval-augmented generation (RAG): the question triggers a search in a controlled corpus, and the answer cites the passages retrieved. The next wave is already reaching legaltech: agents, able to chain several steps (search, compare, draft) with human checkpoints. It was one of the dominant themes of Legalweek 2026.

For the lawyer, the practical consequence fits in one sentence: a tool's value is measured by the quality and freshness of its corpus, more than by the size of its model. An agent searching JUPORTAL (around 145,000 decisions since 1990) or the rulings of the Council for Alien Law Litigation (around 268,000 since 2007) returns verifiable answers. The same agent without a corpus returns elegant ones.

Why we would never run a firm on a tool without citations

Our position is blunt: between a brilliant model that does not cite its sources and an average model that does, a firm should pick the second. Responsibility for the submission stays entirely with the lawyer, the Belgian charters keep repeating it, and a tool's real cost is measured by the verification time it imposes. An answer without references is verified by redoing the whole search: the promised time saving disappears. An answer pointing to precise passages is verified in a few clicks.

This criterion has the merit of being binary at purchase time: the tool shows its sources, or it does not. Promises of "enhanced reliability" without a citation mechanism cannot be tested. A clickable citation can.

The most concrete move comes from the bars. The French-speaking Brussels Bar circulated an AI usage charter for firms, and a model client information clause is in circulation, specifying that the use of AI "in no way amounts to a delegation of the legal work". A model charter is publicly available. For a firm, adopting it takes one meeting: appoint a referent, list the authorised tools, require internal marking of AI-assisted documents and systematic human verification.

Legal AI in 2026: fewer demos, more method

The underlying trend of 2026 is not one more model, it is a change of standard. Adoption is a given, the framework is being written, sanctions are landing: legal AI is now judged on the record, like the rest of the profession. The firms getting value out of it are those with a charter, a verification workflow and tools that cite their sources.

Jef (https://www.jef.chat) fits that logic: the assistant turns a firm's documentary base into a conversational agent whose every answer cites its sources, backed by the official Belgian and European sources. The catalogue of covered sources, with volumes, is public on the sources page. Registration is free, with 50 messages included.