Company databases still work just as they did in 1841: extracting and retyping registers. Instead, Romanian-based Veridion lets AI read the internet like a human does, but at a scale of 1.4 billion pages per month for 693 million companies. Today, everyone can do something with language models, so the question is why this company in particular would do it better.
Take the hotel in Ljubljana where we are speaking with Veridion. The Slovenian business register knows it as ‘MIKLOSIC 3 HOTEL d.o.o., founded June 15, 2022, formerly a consultancy firm’, with an address and an activity code. Veridion knows it as Eurostars uHotel, owned by a Madrid-based company without a website, part of a chain with 300 hotels in 23 countries and a Barcelona-based group with 1.65 billion euros in turnover, tenant of the building, 8.18 million euros in turnover, 4.4 stars from 1,700 Google reviews, and a rooftop pool.
“A company is not a row in a register,” says Alexandra Tofon of Veridion. “It is the sum of everything it leaves behind.”
Company data from 1841 #
The sector that sells business information, accounting for over 20 billion dollars in revenue, dates back to the year the Mercantile Agency of New York sent lawyers and accountants to registers and courts. The method hasn’t changed: a company submits its papers upon incorporation and updates them when the law requires.
What it actually does, where it operates, and how many people work there doesn’t need to be reported anywhere. Every step passes through human hands, and the sector resells the same data pool through resellers. Three out of five companies changed something about themselves in the past year; the database didn’t notice. The result: insurance priced on activities no one reported and credits that SMEs don’t receive because the bank doesn’t see them.
Of Veridion’s 693 million entities, only 180 million are actually operational. The rest are shell companies.
Read the internet like a human #
Veridion started in 2017 with a salesperson who couldn’t find a reliable prospect list. His solution: opening and reading websites. “Can a machine do that like a human, but at scale?” The answer went through three phases that coincidentally align with the evolution of AI. First, rules and classifiers, with the classic dilemma: scale or accuracy—pick one.
Two years ago, the batch era followed: re-reading the entire web weekly in a process taking six hours to two days, even though 95 percent of companies hadn’t changed. “We were wasting computing power on nothing,” says data engineer Alexandru Tofon.
Today, Orion is running—an event-driven engine on Flink and Kafka. Every website is its own stream that updates the graph in seconds.
Today, Orion is running—an event-driven engine on Flink and Kafka. Every website is its own stream that updates the graph in seconds. Pipelines are YAML recipes: cheap checks first, a language model only where it pays off. Over 90 percent comes from cache. As a result, open models of 9 to 27 billion parameters on proprietary GPUs suffice, supplemented by self-trained transformers and, where the task demands it, Gemini, Claude, ChatGPT, and Mistral.
The result: thousands to tens of thousands of pipelines per second and 100 million model predictions per day. Alexandru demonstrated it with a weekend project: 578,000 websites compared to their own previous versions. In one weekend, 377 companies changed a published price, 455 their contact details, and 51 their address; 1,990 rebuilt their site, and two websites turned out to be hacked, with the ransom note as the only verifiable trace.
Why not just ChatGPT? #
Veridion answers that question with four design choices. One: a page yields claims, not companies. A scraper reads a page and records a company. Veridion notes that one source made one claim, and a company only exists once multiple independent sources confirm it. “One wrong match poisons every decision downstream.”
Two: the old value remains until the opposite is proven. Let a language model read the same page twice and it sometimes answers differently. That answer would result in thousands of phantom moves per day. The burden of proof lies with the new run. “Everyone is building with these models now. Almost no one has built-in protection against their non-determinism.”
Three: every value carries its evidence, including source, date, certainty, and the model’s reasoning. “That’s why a chatbot answer doesn’t belong in an insurance file. Ask the same question twice and you get two answers, without provenance.”
Four: proprietary hardware with three replicated data centers in Germany. “In the cloud, the volume we work with would bankrupt us in a few months.” According to Veridion, this makes them half as expensive as established players.
The honest answer to ‘why them’ is therefore not AI. The models are open and available to everyone. The advantage lies in six years of validation, in discipline against hallucinating models, and in hardware. All performance figures are their own measurements: 30 percent more accurate, with an example of a match rate that went from 15 to 60 percent for a Canadian insurer.
The models are open and for everyone. The advantage lies in six years of validation
The boundaries are sharp: Veridion does not know what is not on the internet. It does not have audited annual accounts, credit scores, or DUNS numbers—exactly what the sector runs on.
Discipline is the advantage, and it is temporary #
Veridion does not sell software or seats, but data, per use case and without public pricing. It sells this to credit rating agencies, insurers, procurement platforms, and data houses that resell them under their own names.
The company proves that AI makes the 1841 method obsolete: a machine can read the internet like a human and do so 693 million times simultaneously. Yet Veridion does not focus on the models, but on trust. Claims only become facts after confirmation. Values do not change without proof.
Companies today are too quick to unleash a chatbot on their customer data without seeking structure and trust behind it. The question is how long Veridion can build an advantage now that the same open models are ready for everyone. Will everyone soon build their own Veridion? The company does not see itself getting into trouble that quickly. Just yesterday, it raised 20 million euros in a Series A funding round.
Until then, the lesson applies to every database of companies or contacts: the problem was never too little AI, but too much trust in what was typed in once.