Traditional Personalvermittlung agencies are good at what they were built for: filling well-understood roles at reasonable speed against a legible skill set. AI hiring breaks all three of those assumptions at once, the skill set isn't legible from a resume, the speed the market requires is faster than the agency's economics support, and the roles are specialized enough that generalist vetting produces false positives at scale. Here is the specific, not generic, breakdown of why this happens and what it costs.
Failure mode 1: generalist recruiters can't technically vet AI candidates
A Personalvermittlung consultant is typically excellent at reading people, managing a search process and negotiating between two sides. None of that helps when the actual question is whether a candidate can design a retrieval pipeline that holds up under production load, or whether their "3 years of ML experience" was fine-tuning a single classifier once. Without hands-on AI/ML experience, the honest answer is that a generalist recruiter cannot tell. The result is not that bad candidates never get vetted, it's that vetting happens at the wrong stage, after the client hires, instead of before the shortlist is delivered.
Failure mode 2: slow processes lose the best candidates
A traditional Personalvermittlung search, from brief to signed contract, commonly runs 4-6 weeks, sometimes longer for senior or niche roles. Strong senior AI candidates in Germany's tight market are typically fielding multiple offers within two to three weeks of going active. By the time a traditional agency's process reaches interview stage, the best candidate on the original shortlist has frequently already accepted somewhere faster. What's left, structurally, skews toward candidates who were available longer because they were harder to place elsewhere, which is not a promising signal.
Failure mode 3: volume-based economics push quantity over fit
Contingency-based Personalvermittlung agencies are typically paid only when a candidate is hired, and often compete against other agencies working the same role. The rational response to that incentive structure is to submit as many candidates as possible, as fast as possible, since more submissions mean more chances of being the agency that lands the fee. This is precisely backwards for AI roles, where the value is in a small number of deeply vetted candidates, not a wide net. Volume-optimized sourcing and depth-optimized vetting are in direct tension, and the fee structure resolves that tension in favor of volume.
Where the real cost lands, downstream
None of these three failure modes show up as an obvious problem at the point of placement, a plausible-looking candidate gets hired and everyone moves on. The cost surfaces three to six months later: a model that never makes it to production, a team that has to quietly route around a hire who can talk about AI but can't build it, or a resignation that restarts the search from zero, this time under more time pressure than the first attempt. Estimates for the fully loaded cost of a bad senior technical hire commonly run 1.5-3x annual salary once lost productivity, team disruption and the second search are counted, and generalist AI vetting is a structural driver of exactly this outcome.
What actually closes the gap
- Technical evaluation done by people who have shipped real AI systems, not recruiters reading resumes for keyword density.
- A compressed, parallel process (sourcing, screening, reference checks running together, not in sequence) that can produce a vetted shortlist in days, not weeks.
- Fee and process structures that reward accurate fit over submission volume, fewer candidates, each one already past a real technical bar.
- Sub-specialty awareness, treating ML engineering, MLOps, applied research and LLM/agent engineering as genuinely different roles requiring different vetting, not one "AI" bucket.