Three Companies, One Lesson: When AI Judges People, Trust Is the Business Model

Signup Write New Story Three Companies, One Lesson: When AI Judges People, Trust Is the Business Model by Valery Meshkov by Valery Meshkov | @ meshkov IT entrepreneur and the founder of Pravoved.ru and Jinn

IT entrepreneur and the founder of Pravoved.ru and Jinn

I have started three companies, and in all of them, software makes some kind of judgment about people. At Pravoved.ru, the online legal marketplace I co-founded, it was about matching people with lawyers and answering their legal questions. At Jinn, an HR-tech platform, it was assessing employees and candidates for corporate clients. Now at Narvalus, which I co-founded in Miami, we are building an AI recruiting agent for U.S. hiring teams.

When I started each of them, I thought the main risk was technical. Will the model be accurate enough? It almost never turned out to be the problem. The problem was that people didn't believe the system, even when it was right. And that changed not only the product but how we sold it, priced it, and decided where to start.

When we launched Pravoved, legal help in Russia mostly meant going to an office, paying an unclear price, and hoping the lawyer was good. Nobody could compare. Our bet was that people would ask a question online if they could see several lawyers answer it and check their ratings from other clients.

That bet worked, but the interesting part came later. Most incoming questions were standard, the same consumer-law issues again and again. So we built an AI assistant trained on our lawyers' answers to handle them. Technically, it was fine. What made users accept it was that a real lawyer was always one click away and the reasoning was visible. When an answer came without context, people doubted it, even when it was correct.

The business lesson for me was simple. Automate what is standard, keep people for what is not, and never hide which is which. It also pushed us toward corporate clients. Companies wanted predictable, documented legal support, and over time B2B grew to about 65% of our revenue. Investors from AddVenture and Target Global backed that model, in part because trust was built into it and not added later.

With Jinn, I moved into HR. The platform runs assessments, 360-degree feedback, and engagement surveys, and uses AI to turn that data into recommendations about people. We went from one corporate client to more than seventy in a few years.

For an HR director, the accuracy percentage of a model is not the real question. The real questions are "Can I explain this to the manager?" and "What do I tell the employee if they ask?" A score alone did not survive those questions. A score with clear reasons behind it did. So the explanation became the thing we really sold, and the score was more like a summary of it.

Recruiting is where this matters most, because the AI influences who gets an interview and who doesn't. And in the U.S., the rules are getting stricter. New York City's Local Law 144 requires an independent bias audit, a published summary, and notice to candidates when automated employment decision tools are used. The EU AI Act puts AI used in recruitment in the high-risk category. And existing anti-discrimination law applies whether a person or a program made the decision.

So when we designed the Narvalus agent, we started from trust and worked back to the model. A few decisions came out of that.

The agent works from the job description, criterion by criterion. It turns the job description into concrete requirements and checks each candidate against each one, with a short reason for every match or gap. The recruiter can see exactly why someone is near the top, without writing Boolean search strings.

It flags questions; it doesn't reject. If two jobs on a resume overlap in time, or a career path is unusual, the agent marks it as something to ask in the interview. People make every hiring decision. This was not a hard choice for me after Pravoved: the unusual cases are exactly where a human is worth the most.

Scoring and auditing are kept apart. The scoring step is designed never to receive age, photos, gender, or other protected attributes. But just removing fields doesn't remove every proxy; a graduation year can still hint at age. So we are building the audit side separately: where candidates choose to share demographic data, it is stored apart from scoring and used only to report selection rates for bias audits. If you skip that part, you can't actually prove the first part works.

This is the part I think founders underestimate. Explainability is shaping our go-to-market plan as much as our architecture.

First, the starting market. We chose to begin in Miami with staffing agencies and small and mid-sized employers. They have a lot of applicants and small teams, often without a dedicated sourcer or in-house counsel. A big company can afford to check a vendor's claims itself. A ten-person agency can't, so a tool that explains itself is worth more to them, and they can decide quickly.

Second, pricing. We plan a simple per-seat subscription with a free trial on real open roles. If the value is that recruiters trust the shortlist, the fastest way to prove it is to let them try it on their own vacancies, not to show a demo.

Third, the enterprise tier. Bias-audit reporting and documentation will go into the higher plan. Larger employers carry more legal exposure, and for them compliance is not a nice extra. It decides whether the deal moves forward at all.

If your product judges people in some way, for jobs, loans, housing, or legal help, build for the most skeptical person who will read the output: a regulator, a lawyer, or the person being judged. If the explanation doesn't hold up for them, the feature is not finished. And design your business around that skeptic too, because they are often the one who signs off on the purchase.

At Pravoved, the AI never replaced lawyers. It took the routine questions so lawyers had more time for the hard ones. I believe recruiting will go the same way. In a year or two, we will see if I was right.

Valery Meshkov is a serial technology entrepreneur. He co-founded Pravoved.ru, an online legal marketplace backed by AddVenture and Target Global, founded the HR-tech platform Jinn, and is a co-founder of Narvalus, an early-stage company developing an AI recruiting agent for U.S. hiring teams.

How I Productized Legal Advice Into a Scalable Marketplace