Hudson Labs launches a no hallucination guarantee, which CEO Kris Bennatti says no other market competitor will match, after a test showed state-of-the-art models returning wrong numbers 30% of the time beyond four to eight periods.
Bennatti says Hudson Labs' forensic risk score converts qualitative governance risks into machine-learning variables and carries about a one-in-three chance of SEC enforcement at 70 or higher.
Bennatti says Hudson Labs' subscriptions are profitable, so prices will not suddenly rise, while inefficient AI use becomes costly as developers chafe at Anthropic lock-in.
Hudson Labs bets on guaranteed numbers while financial AI hallucinates
Financial AI still gets numbers wrong, and Hudson Labs is betting that pre-processing, retrieval and checks, not bigger models, make figures trustworthy. Models cannot tell when they are wrong, so long spans of data need search rather than larger context. Hudson pairs this with a forensic risk score built from qualitative disclosure signals. Bennatti expects generalist models to need years to catch up.
Takeaways
Accurate financial numbers need pre-processing and built-in checks. Bennatti says models fail worst when they do not know an answer and do not know they do not know. In a test of leading models, asking for a metric over more than four to eight periods returned a wrong number 30% of the time.
Hudson Labs offers a no hallucination guarantee. Bennatti says Hudson is the only company willing to offer one. It covers restatement-adjusted, error-free numbers over five years and rests on the whole architecture, not just the last generative step. It does not apply through MCP, which is not deterministic.
Retrieval metadata enables searches that keywords cannot. Hudson stores SEC filings and transcripts in vector databases with tags for KPIs, timing and sentiment, holding 52m data points across 10,000 to 20,000 companies. That allows sentiment screens, such as the most stressed CEOs. NotebookLM wins on better search, not a bigger context window.
The forensic score turns qualitative risks into variables. Older fraud models used days sales outstanding and accruals and were noisy. Hudson's score converts management integrity, turnover at the top, aggressive accounting and off balance sheet risk, among others, into variables. Machine learning sets weights using companies with SEC investigations or fraud settlements.
High risk scores predicted underperformance and enforcement. In an older backtest, high-risk companies lagged share prices by about 15% and were three times more likely to face a securities class action. A score of 70 or higher carries roughly a one in three chance of SEC enforcement. D and O insurers use it.
Finance AI vendors must integrate or die. Caughran finds logging into several finance chatbots cumbersome. Bennatti says 2026 is integrate or die. Some enterprise customers can only use Copilot, so Hudson builds connectors in both directions. It is not prioritising a CLI because many customers do not know what MCP is.
General models will take years to close the gap. Bennatti expects generalist models to catch up on hallucination in about five years. LLMs read tense from verbs and miss guidance, such as American Eagle's capital expenditure statements. Hudson's guidance model assumed only a nine-month moat, as big providers invest in finance.
Tap a name to go there · tap a line to see why they are linked