We run a technical recruitment agency. Eighteen years of hiring engineers for companies that expect a lot from their vendors. For most of that time, the hardest part of the job was finding good people. Somewhere in 2025 a second problem showed up next to it: telling whether the person in the pipeline exists at all.
It didn’t arrive as one dramatic case. It arrived as a pattern. CVs that read perfectly and matched the job ad a little too well. Phone numbers that were never mobile. Three applicants for the same role with the same skills section, word for word. A candidate who did great on the first call and was a different person on the second.
Gartner expects one in four candidate profiles worldwide to be fake by 2028. The US Department of Justice has charged the people behind a scheme that placed North Korean IT workers into more than 300 US companies. Those are big numbers, and they are easy to nod at. What made it real for us was seeing it on our own jobs, with our own clients’ names on the line.
What we did first
We did what every agency does: we tightened the process. More checks before the interview, more questions on the call, a shared list of things to look for. It worked, and it didn’t scale. Every check took a recruiter’s time, and the patterns kept changing.
So we started writing the checks down as rules. Does the file’s author match the candidate? Do two full-time roles overlap? Is the phone number a VoIP line? Is there text in the PDF that a reader can’t see but a filter can? Then we ran the rules on every application that came in, before anyone looked at it.
That was the moment it stopped being a process and started being a product. Not because we went looking for a SaaS idea, but because the checks were only useful if they ran on everything, every time, without a recruiter remembering to do them.
The one rule
Early on we made a decision that shaped everything after it: the system never says “fake”.
Not because the word is impolite. Because it is wrong. A tool that reads a CV and public data can tell you an application is inconsistent. It cannot tell you why. Honest people have messy CVs, gaps, odd file names, a phone number from a previous country. A single signal is a note to look closer. What separates a fabricated application is a cluster of signals across different areas, and even then the right response is a human looking at the evidence, not a rejection.
So TalentShield produces signals with reasons. Every signal says what we found, why it matters and how you can check it yourself. The score sorts the pipeline so recruiters know where to start. The badge says “worth a closer look”, not “reject”. The verdict, when there is one, is recorded as a person’s.
This turned out to be the right call for a second reason. Recruitment tools are high-risk under the EU AI Act, and the obligations that come with that are exactly the ones a “signals, not verdicts” design already meets: human oversight, explainability, logging. We didn’t design for the regulation. We designed for our own recruiters, and the regulation agreed.
Where we are
TalentShield runs on every application our agency receives. It lives inside the ATS, so recruiters don’t open another tool. When a new pattern shows up in our pipeline, it becomes a signal the same week. We are running pilots with teams that hire at scale, and each of those pipelines makes the library better.
If you hire in tech, remotely or across borders, and you would like to see what the checks find on your own applicants, book a demo. Real signals on real candidates, and a weekly call with people who still recruit every day.
