Why Writer’s $2B generative AI platform matters for regulated

The reality of production hallucinations

I view the industry focus on model leaderboards as a strategic error for anyone managing enterprise risk. While top models score under 2 percent on grounded summarization benchmarks, hallucination rates in production workflows run 10 to 40 times higher. Stanford RegLab found that general purpose models hallucinate on 69 to 88 percent of legal queries. An industry average hallucination rate of 20 percent means an organization processing 10,000 tasks per day generates 2,000 incorrect outputs every day. Because error rates compound in multi-step agentic workflows, a model with a low benchmark score still fails when it handles messy retrieval or specialized domains. OpenAI’s o3 reasoning model hallucinated on 33 percent of PersonQA prompts, which is double the 16 percent rate of its predecessor o1. This occurs because standard evaluation methods reward guessing rather than admitting uncertainty. A federal court in New York sanctioned attorneys after a chatbot in the Mata v. Avianca case produced six fabricated judicial decisions. A personal-finance publisher also had to correct 41 of 77 AI-assisted articles because the model produced wrong compound-interest arithmetic. In Germany, a court ruled that an AI chatbot for Aesthetify was responsible for inventing false specialist qualifications for clinic founders.

Regulatory pressure and insurance shifts

The legal pressure on AI deployment arrived with the EU Product Liability Directive, which member states must transpose by December 9, 2026. The EU Product Liability Directive, which member states must transpose by December 9, 2026, treats software as a product subject to strict liability and prohibits manufacturers from using contracts to limit this liability. California’s AB 316, effective January 1, 2026, prevents companies from asserting that an AI’s autonomous conduct caused harm instead of their own choices. You should understand that these regulations make traceability a legal requirement rather than a technical preference. State legislatures in the US have also moved to regulate AI. Colorado’s law, with implementation delayed to June 30, 2026, targets high-risk systems used for consequential decisions like financial lending. Texas passed the Responsible AI Governance Act in June 2025, which prohibits intentional discrimination and manipulative conduct. Utah’s law requires disclosure when people interact with generative AI and extends its implementation to July 1, 2027. The New York Department of Financial Services issued guidance in October 2024 to assist regulated entities in mitigating cybersecurity risks from AI. Insurance markets are also adjusting, with ScienceSoft projecting 60 to 80 percent of new policies will include AI risk assessment by 2028. Does the current insurance landscape offer enough coverage to offset these legal shifts?

Defensibility through specialized architecture

Writer solves the defensibility problem with an architecture built for regulated sectors. The company reached a $1.9 billion valuation after its $200M Series C round in November 2024. Investors like Salesforce Ventures and Adobe Ventures participated in the round. Its Palmyra Med model uses 70 billion parameters and achieves 85.9 percent accuracy on medical benchmarks. The platform uses a proprietary Knowledge Graph to anchor responses in verified internal data, such as clinical trial results or approved medical libraries. Writer provides a full audit trail for agentic workflows and includes every tool call and intermediate step. This allows a compliance officer to trace an output back to its specific source document. The system shows how much each source influenced an answer so reviewers can rank the highest-impact sources first. Writer leads the enterprise generative AI category and serves customers like Mars, Ally Bank, Lennar, Prudential, Qualcomm, Salesforce, Uber, Vanguard, Accenture, and Intuit. In healthcare, users see 70 percent productivity gains in medical record summarization. The Palmyra X4 model also achieves high accuracy in tool calling, which is a skill necessary for executing complex actions. Palmyra Med costs $10 per 1 million output tokens, which is much lower than the $60 rate for GPT-4.