In a Latent Space podcast interview, Jacob Lauritzen, CTO of legal-tech company Legora, described how the firm builds AI agents for lawyers. According to Latent Space, Legora started in Stockholm with a product combining ChatGPT, Swedish law, and a search tool. Its agent now drafts and edits Word documents, reviews thousands of contracts at once, and coordinates work between lawyers, clients, and other agents. Latent Space reported the company passed $100 million in ARR in April, less than 18 months after launch, has more than 1,800 customers, and raised $550 million in October at an $8.8 billion valuation.
Lauritzen said Legora cannot read any user traces because uploaded data is confidential and often belongs to the firm's clients. Instead, three engineers and three lawyers build simulated environments, such as a 10,000-file data room for due diligence, and plot agent performance by practice area. The company draws on about 100 legal engineers and classifies production failures into 10 to 20 known modes.
Lauritzen estimated that current models handle about 60% of legal work well, with complex analysis across million-token contracts still requiring more capability. Everything the agent writes receives a citation linking to the source document. He also said US firms are shifting routine work to flat project fees, pressuring the billable-hour model.
- Legora passed $100M ARR in April, less than 18 months after launch, with 1,800+ customers (per Latent Space)
- Raised $550M in October at an $8.8B valuation; Wexler was its fifth acquisition this year
- Lauritzen estimates models handle ~60% of legal work well; everything the agent writes gets a source citation
- Legora can't read user traces due to confidentiality, so it builds simulated RL environments to test agents
What it means for you
This is a vendor-founder interview about one legal-AI company's internals, so most operators can read it as context rather than a call to act. The genuinely useful part for anyone running operations: Legora can't see its own customers' data, so it tests its AI against made-up scenarios that imitate real work. That's a reminder that an AI tool's accuracy on your actual work is something you have to verify yourself, not something the vendor can promise. The other durable point — that every AI output carries a citation back to the source document — is a sensible standard to demand from any tool you let touch customer or legal material.
Try this
If you already use or are trialing an AI tool on any high-stakes documents, check one thing this week: when it makes a claim, can you trace that claim back to the exact source? If not, treat its output as a draft a person must verify, not an answer.
Who should care
Owners and operations leads at firms that handle contracts, compliance, or confidential client documents — and anyone evaluating an AI tool for legal, financial, or regulated work.
Skip this if
You're not in legal services and aren't evaluating an AI tool for document-heavy or regulated work. This is industry-insider detail, not something that changes your week.
Sources: Latent Space — read the original