Legora CTO Jacob Lauritzen says the company raised its AI agent’s citation accuracy from 70% to 99% in a single release, without changing anything else. The interview also describes how Legora tests legal agents when client confidentiality means it cannot inspect the agents’ work directly.
Watch Desk analysis
What happened
In a Latent Space interview published on 2 November, Lauritzen said Legora builds test environments modelled on customers’ work, using examples from legal practice to assess its agents. He described a citation problem that went from 70% accuracy to 99% after one release, without a change to the surrounding system.
Lauritzen also discussed the difficulty of evaluating an agent when its work involves confidential law-firm client data. The approach he described includes simulated environments and narrowing failures down to a manageable set of types. These are his account of Legora’s methods and results, not independently measured performance figures.
Why it matters
Legal AI needs more than convincing prose. A citation that looks plausible but points to the wrong authority can undermine the work around it. Legora’s reported improvement is a striking example of how much a targeted release can matter, while its testing challenge shows why evaluation gets complicated when the most useful real-world data cannot simply be inspected.
The practical lesson is that buyers should ask not only which model a legal tool uses, but how its provider tests citations and investigates failures without exposing confidential client material.
Our read
The 70% to 99% jump is the headline number, but the more durable point is the testing problem behind it. If legal agents are to earn trust, vendors need credible ways to measure errors while respecting the confidentiality that makes legal work, well, legal work. Lauritzen’s account gives a concrete example, though the interview summary does not provide the test set or measurement method behind the figures.
What to watch
- Whether Legora publishes more detail on how it measures citation accuracy.
- How its simulated test environments compare with performance on real legal work.
- Whether other legal AI providers explain how they test systems without inspecting confidential client traces.
Discussion spark: Should legal AI providers publish detailed accuracy tests even when client confidentiality limits what outsiders can inspect?
Sources and evidence
- Building Legal Agents When You Can't See the Data – Jacob Lauritzen, CTO of Legora (2 November 2026, 22:54 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.