Applying Engineering Judgment to Your Hiring Tech Stack
AI in hiring: what actually works — and what’s just vendor marketing. Start with where your tools’ data came from, not what the demo promised.
Every major sourcing platform you are paying for is built on data that was, at some point, scraped, inferred, purchased, or aggregated from sources the candidates themselves never explicitly consented to share with a recruiting tool. The profiles you’re searching, the contact data being “enriched,” the engagement signals being surfaced — almost none of it came from people raising their hand and saying “yes, please share my information with anyone who pays for a sourcing license.”
Understanding how the data got there tells you a great deal about how reliable it is, how fresh it is, and why two different tools searching for the same profile return completely different results.
If you want to understand your tools, start by understanding where their data comes from. Most vendors won’t tell you directly. But it’s not hard to figure out, if you know where to look.
That’s engineering judgment applied to hiring technology. And it’s what most TA teams are not doing.
The “AI” Problem in Recruiting Tools
Every recruiting tool sold in the last two years has AI in the description. It is now as meaningful a differentiator as “cloud-based” was in 2012 or “mobile-first” was in 2015. Which is to say: it describes an implementation detail, not a capability.
Here is what AI in a recruiting tool usually means in practice, without the vendor language.
Résumé screening AI is, in most cases, a weighted keyword frequency model trained on your historical hiring data. It learns what your past hires looked like on paper and scores new candidates accordingly. The problem is structural: it encodes whatever biases existed in your previous hiring decisions and treats them as ground truth. It also systematically underserves non-linear careers, career changers, and people whose vocabulary doesn’t match the job description’s vocabulary — which at senior and principal levels is a significant proportion of the most interesting candidates.
I’ve seen tools in demos that were confidently scoring a Principal Engineer candidate in the bottom quartile because their résumé described distributed systems work in the language of the problem, not the language of the tools. The ATS didn’t recognise Kafka by its architectural description. It needed the word Kafka.
Candidate matching AI in sourcing platforms is largely semantic search — comparing vectorised representations of a role brief against vectorised representations of candidate profiles. It has gotten genuinely better in recent years. Semantic matching can surface candidates whose skills are adjacent to what you described, which is useful. But the quality of the match is entirely dependent on the quality of the brief you feed it. Vague in, vague out. And most hiring briefs are vague.
Engagement and conversation AI — tools that score interview responses, analyse tone, assess “culture fit” signals from video — is the category that should be raising the most flags in European markets right now, and largely isn’t.
The EU AI Act: This Is Not a Future Problem
The EU AI Act is in force, and high-risk AI system requirements apply from August 2026. The Act explicitly classifies AI systems used in employment and worker management decisions — including recruitment, promotion, and task allocation — as high-risk.
What that means in practice: high-risk systems require conformity assessments, human oversight, transparency to affected individuals about the use of automated decision-making, documentation that would survive regulatory scrutiny, and registration in the EU database before deployment.
When you go through the documentation of the tools being sold into the European recruiting market — the interview analysis tools, the automated screening tools, the AI scoring systems — very few of them have clear, auditable answers to these requirements. The sales deck has an AI Act compliance section. The actual technical and legal documentation frequently does not.
Ask your vendor for their EU AI Act conformity documentation before your next renewal. Watch the response carefully. Uncertainty here is your answer.
If you are operating in Czechia, Slovakia, Germany, or anywhere under EU jurisdiction and you are using automated scoring in a hiring decision, you have legal exposure that most TA leaders are not yet aware of. That is worth saying plainly rather than leaving it quietly in the background.
Most Tools Not Mentioning About Their Data
Sourcing platforms compete on data coverage and freshness. When a vendor tells you they have 700 million profiles or 1 billion data points, the useful follow-up questions are not “how many?” but “from where?” and “how old?”
Most aggregated sourcing databases are built from some combination of public social profiles scraped at some point in the past, data purchased from third-party brokers, job board applications aggregated over time, email address graphs built from professional communication metadata, and inferred attributes layered on top of all of the above.
The inferred attributes are where things get interesting. A platform may show you a candidate’s inferred current employer, inferred seniority level, inferred likelihood of being open to a move — none of which the candidate provided directly. These inferences are generated from behavioural signals: what content someone has engaged with, what job postings their IP address has visited, what email patterns their address appears in.
This is not necessarily wrong or illegal. Under GDPR’s legitimate interest provisions, much of it is defensible. But it means the profile you’re looking at is partly real and partly a model. The “likelihood to be open to a move” score your sourcing tool is showing you is not the candidate telling you they’re open. It’s the platform guessing, based on signals the candidate probably doesn’t know they’re emitting.
Knowing this changes how you use the data. It’s a lead, not a guarantee. A signal, not a fact. And it means your outreach message should not assume openness — it should create space for it.
The Platforms Underneath the Platforms
One thing experienced sourcers understand that most TA buyers don’t: the tools you are paying for are, in many cases, querying the same underlying data through different interfaces.
Several of the major sourcing platforms license data from the same two or three underlying aggregators. The differentiation is in the interface, the workflow, the filtering logic, and the CRM built on top — not in the fundamental data. You may be paying for three tools that are showing you the same universe of profiles through different windows.
This is worth testing empirically. Run the same search in your top two or three tools and compare the overlap. If the result sets are highly correlated, you are paying for redundancy.
Where genuine differentiation exists — and it does exist in some categories — it tends to be in real-time data signals, proprietary community data, or specialised domain coverage that the broad aggregators don’t index well. Niche platforms for deep tech profiles, for security engineering, for ML research backgrounds sometimes have genuine coverage advantages because they’re embedded in communities the general platforms don’t reach.
The public internet is larger than any commercial database, and the candidates you most want to find are often the ones who have left the fewest footprints in the places everyone is looking. Knowing how to read what’s publicly visible — and what’s deliberately not visible — is a skill that separates effective sourcers from people running Boolean searches and hoping.
Evaluate for Lock-In Before You Evaluate for Capability
A senior engineer evaluating a new dependency doesn’t just ask “does this do what I need today?” They ask what it costs to leave.
Hiring tools have the same property, and most buyers don’t interrogate it.
Your ATS accumulates your entire candidate relationship history. Every pipeline stage, every note, every rejection reason, every past applicant who might be right for the next role. The data is technically yours. But whether you can actually get it out in a useful format, and what it costs to migrate it, is a question most vendors answer vaguely during the sales process and specifically in the contract.
Before you sign anything with a meaningful term commitment, get written answers to the following: What does data export look like in terms of format, completeness, and cost? Do you own your historical candidate data in a portable format on day one? What are the exit provisions if the product changes materially or the company is acquired? How does pricing change as your hiring volume grows by two or three times? What integrations exist and what does it cost to maintain them when either system updates?
The vendors with the best demos frequently have the vaguest answers to these questions. That ratio should concern you.
The Adoption Curve Tells You What the Tools Are Worth
When every company in your market has the same sourcing tool, the sourcing tool is table stakes. It is not competitive advantage.
Automated outreach at scale reached near-universal adoption across most markets in 2023 and 2024. The documented result: Gem analysed over four million recruiter messages and found single-touch email response rates hovering at just 4.9% per message, with full 4-stage sequences peaking around 32% to 35%. (Similarly, broader market InMail reply averages linger between 10% and 18% for cold outreach). The technology made the candidate experience worse for everyone and erased the advantage for anyone using it. The stadium stood up to see better and nobody could see better.
Be precise about what you’re buying.
You are not buying competitive advantage when you buy a tool that every other TA team in your market has had for two years. You are buying operational parity — the right not to fall behind. That’s a legitimate purchase. Just be clear-eyed that it’s what it is, and don’t expect it to solve a reach or quality problem it isn’t designed to address.
Competitive advantage in 2026 comes from inputs no tool can replicate: a precisely defined role that gives candidates enough information to self-select accurately, relationships with passive talent that predate the urgency of an open role, and sourcing methods that operate outside the channels everyone else is using. None of those ship in a SaaS subscription.
There’s a broader structural shift worth naming here. The era of static hiring documents — generic job descriptions, keyword-stuffed CVs, boilerplate LinkedIn summaries — was designed for a slower, less complex world. Today’s most effective hiring infrastructure is dynamic: it gives candidates real information upfront, it surfaces the right talent before a role is urgent, and it creates conditions where both sides can make genuinely informed decisions before investing time in interviews. The tools are part of that picture, but they’re not the system. The system has to come first.
The System Determines Which Tools You Need — Not the Other Way Around
This is the point everything else leads to.
Most TA tech stacks were assembled reactively. A role took too long, so someone bought a sourcing tool. Interview quality was inconsistent, so someone bought a structured interview platform. Scheduling was a mess, so someone bought a coordination tool. Each purchase addressed a symptom. The underlying system that produced the symptoms was never redesigned.
A functioning hiring system — clear role definition, documented sourcing approaches for each candidate profile type, a structured and consistent evaluation process, a feedback loop that actually improves over time — generates a specific, defensible list of tools it needs and where they fit. The system tells you which tools to buy. Without the system, you keep assembling tools in search of a process that never quite comes together.
The TA teams making genuinely good hiring decisions in 2026 are not the ones with the largest tech stacks. They tend to have fewer tools, used more deliberately, inside a process that was designed before the tools were selected. They also tend to own their candidate data, their sourcing logic, and their evaluation frameworks outright — rather than renting capability they lose the moment a contract expires.
The data confirms this paradox: while HR AI adoption jumped from 26% to 43% in twelve months, SHRM benchmarks show the average time-to-fill actually expanded from 33 to 44 days. Simultaneously, Gartner HR Research found candidate offer acceptance rates plummeted to 48%—demonstrating that tech stack inflation often slows down cycles and introduces friction rather than improving hiring quality.
Before you sign the next renewal or take the next demo, it’s worth mapping your process stage by stage: where the friction actually lives, and whether the problem is a process gap, a judgment gap, or a genuine tool gap. In many cases, the answer is upstream of anything a tool can fix.
If You’re Building Hiring Infrastructure in 2026
The companies getting the most value from their hiring technology investments share a few things in common. They define the process first and select tools to serve it. They interrogate vendor data provenance before signing. They build in contract exit provisions as a non-negotiable. They treat EU AI Act compliance as a current legal consideration, not a future one. And they hold automated outreach to the same standard they’d hold any other communication — is this actually going to create a conversation, or just add noise?
They’re the same questions a good engineer asks before accepting a new dependency. The discipline of applying that thinking to hiring infrastructure is what separates teams that are genuinely improving from teams that are perpetually shopping for the tool that will finally make everything work.
If you’re a TA leader working through this kind of systems-first evaluation — trying to build hiring infrastructure your organisation actually owns rather than rents — Recberry’s Talent Engineering work is worth a look as a practical starting point.
Part of an ongoing series on TA strategy: what the data — and the details vendors don’t lead with — actually say about hiring in 2026.
Barbora Jenšík is the founder of Recberry. She has spent twenty years placing engineers into companies including Skype, Avast, Barclays, and Kiwi.com as an external recruiter. Recberry deploys sourcing and vetting systems inside technology companies so their internal TA teams can fill senior engineering roles without agencies — then leaves.