50,000 Companies Analysed by an AI Robot — and What It Found That Nobody Expected

The starting point: a business that only grew as fast as its founders’ network

A web agency ran entirely on networking — referrals, contacts, events. Quality leads, but with a structural flaw: intermittency. Three or four months could pass without a single lead entering the funnel. There was no prospecting process, no list of potential clients, no predictable growth channel.

The system we built didn’t make an existing process more efficient. It created the first active acquisition channel in the agency’s history.

The idea: the market had already signalled its interest — it just had to be read

The starting point was Google My Business. The logic: a company that set up a profile there has already demonstrated it cares about being found. And for an agency that builds websites, GMB has a rare property — you instantly see who has a website and who doesn’t.

On top of that we layered financial-data sources (Listă Firme, Risco), which opened sales angles you don’t have otherwise: you can see, per industry, whether companies with a website have higher average revenue than those without. You’re no longer selling “a website” — you’re selling a measurable difference within their own market.

The initial filtering was done by industry: sectors with many listed companies and real competition — exactly the ones where a website genuinely helps the business, rather than being sold for the sake of selling.

How it works: an AI robot that doesn’t classify information — it gathers it

For every company on the list, an AI agent actually visited the website and ran a battery of checks:

• compliance with Google’s guidelines (relevant for businesses that depend on organic traffic);
• presence of terms & conditions and a privacy policy;
• the company’s legal identification on the site (registration number, identifying data);
• legal obligations for dealing with consumers — including the mandatory information about alternative dispute resolution (SAL, under Romanian/EU consumer law).

For companies without a website, the robot matched the business name from GMB to the company in the public registries and checked its revenue and profit. Inactive companies or ones with heavy losses were excluded automatically. But a healthy company, smaller than its competitors who do have a website? An ideal candidate for a brand-new site.

Everything flowed back into the CRM, which automatically calculated a score per lead — alongside the issues found, the financial data, and every discovered contact (emails, phone numbers). The salesperson opened the record and already had everything.

The numbers

• 50,000+ companies checked in depth, including actually visiting the site;
• under 15 seconds per site, versus the 30–60 seconds it would take a human — and, being asynchronous, the system ran up to 50 sites at once. In practice: a volume that would have meant months of human work compressed into a few hours of processing;
• calibrated accuracy on random samples of 20 companies each, iterating on the robot until its divergence from a human’s assessment dropped below 5%;
• final result: 400–500 companies with an existing site and concrete problems identified, plus a nearly-as-large segment of qualified companies with no site at all (~60/40) — that is, two distinct offers, already segmented: fix/rebuild vs. brand-new site.

False positives happened — at worst, under 1 in 10. But their cost was a one-minute human check; the savings across the rest of the process covered them hundreds of times over.

The surprises — the part that wasn’t in the plan

1. Nearly 30% of the identified websites were, technically, breaking the law. They were missing the company’s identifying data and the mandatory alternative-dispute-resolution information — fineable violations, in fields that work almost exclusively with consumers: lawyers, dental practices. The companies didn’t know. Which turns the first sales contact into a real service: “you have a problem that could cost you a fine” is a completely different conversation from “we’ll build you a nice website”.
2. Of the decent sites, ~90% looked abandoned for over 5 years. Built once, then forgotten — no updates, no care. The market wasn’t full of companies without a website; it was full of companies that thought they’d ticked the box.
3. For the team that built it: the first time the AI gathered information rather than just classifying it. And the fact that a system like this can get within 5% of a human’s accuracy, running 3–4× faster and with dozens of sites in parallel at a comparable cost — is a level of scale you never reach with people alone.

The counter-intuitive decision: no automated outreach

With 400–500 qualified leads and all the contact data in the CRM, the obvious temptation was to automate the outreach. The decision was the opposite: every contact is made by a human. The AI does the part people do badly and slowly (gathering and verifying at scale), and people do the part the AI shouldn’t do for them — the relationship.

What remains

An AI-powered CRM isn’t “a smarter database”. In this case it was the difference between a business dependent on luck in networking and one with a funnel built on data: who’s in the market, who has concrete problems, who can afford the solution, and who’s worth calling today.

And the most valuable discovery wasn’t the lead list — it was the picture of the market: a third breaking the law without knowing it, 90% with an abandoned digital presence. The kind of thing you never see from the ground, but that becomes obvious when someone — or something — looks at 50,000 companies at once.