The sales engineer had the job title typed into the box before I had finished saying it. Senior platform engineer, payments, Kubernetes. Eleven profiles came back with a small green score beside each one, and I knew six of them, because I had spent InMail credits on all six inside the previous month. Two had declined. One declined twice.
Nobody else in the room noticed. The demo was going fine.
Twelve years in, and that pattern has not moved. The demo req is a req the vendor has already solved, run against a market the vendor has already indexed, in front of a room that cannot check the answer. Not dishonest. It is a sales meeting. The real trouble is that one category label is being asked to cover three separate jobs, and those jobs have different inputs, different failure modes, and almost nothing in common at the point where the money moves.
Research, sourcing, screening, and only one of them is photogenic
Research is what happens before any individual name matters. Where does this skill set physically sit? Which employers run a team doing it at the scale the brief assumes. What the adjacent titles look like when the obvious one does not exist locally, which in regulated work it very often does not. I spent a stretch on pharmacovigilance roles and learned that the title on the req and the title in the market agreed roughly never.
Sourcing turns that picture into named people, with contact details attached and a reason for the approach written down.
Screening sorts people already in front of you, inbound or replies, against criteria somebody defined. Or did not define, which I see more.
Most of the grumbling I hear about AI recruitment tools is really about buying strength in one of those three and expecting it in the other two. A screening product will not tell a hiring manager that his market is a third the size he believes. A market map will not sort four hundred applications by Friday. Both are decent at their own job. Neither was the thing that was needed on Tuesday. If the slate you wanted was a market read and what turned up was a tidier inbox, the product did not fail. The purchase did.
I was wrong about boolean for years
My strings used to be beautiful. Nested brackets, four spellings of the same framework, an exclusion tail bolted on after every bad batch of results. I thought the string was the craft. It is mostly a comfort ritual. A long string describes my vocabulary, not the market’s, and the market’s drifts constantly, because the same work gets written up one way by somebody out of academia, another way by somebody out of consultancy, and a third way by somebody who has only ever sat in a product team.
The ceiling was never the string anyway. It was the index underneath it. Run everything through LinkedIn Recruiter and the candidate universe becomes the population that keeps a profile current on LinkedIn, which in narrow fields excludes precisely the people worth ringing. The principal engineer who last updated hers two jobs ago has not left the field. She stopped getting anything useful through that channel and went back to ignoring it.
Her work is public. It is just scattered. Commit and issue history on GitHub or GitLab. Answers on Stack Overflow with ten years of dust on them. A talk sitting in a conference archive, preprints on arXiv, papers through Google Scholar, PubMed for anything clinical, a patent with her name third on it, a standards working group whose mailing list goes back further than her current employer, a Discord that nobody outside the field has heard of. Everybody says look at the GitHub. Almost nobody reads the GitHub. You can tell inside a minute of talking to someone whether they have opened a repository themselves or whether they are repeating a line they picked up at a meetup.
Reading across all of that by hand is possible for one req, on a quiet week, if the phone stays silent. So it does not happen. The argument for adopting AI recruitment tools in the research layer is throughput and nothing loftier than throughput, and throughput is enough.
Contact data is where sourcing quietly fails
A list of names is not a deliverable. I have been handed gorgeous slates of unreachable people. Work address bouncing, mobile recycled to a stranger who was getting fed up, personal address accurate in 2021 and dead since. You find that out on a Monday, in batches, when the drip you set running on Friday reports a send rate that looks respectable and buries the bounce log two screens down.
Vendors differ most here and talk about it least. The questions worth asking are procedural and dull enough that nobody raises them in a demo:
- Where did this address come from, and when was it last confirmed as actually delivering?
- Two independent sources, or one pattern guess off the company domain?
- What happens when it stops resolving? Does anything recheck it, or does the record just sit there looking healthy?
- Can the trail be opened up for one specific person, or is there only a badge reading verified?
Verified with no method behind it is a word. Two confirmations and a date is something to act on, including the decision to ignore the file and reach the person through a mutual contact instead.
A ranking is an opinion with a number bolted on
A ranked slate looks like measurement. Underneath it is a set of choices about what counts as relevant, applied at speed and without variation. Consistency is worth real money on volume reqs, where human attention falls off a cliff somewhere around three in the afternoon and never properly comes back. Consistency tells you nothing about whether the rule is any good. A rule that penalises a career break will penalise every career break in the pile, politely, at speed, and nobody in the loop will watch it happen.
Which is why I ask to see the discards. Vendors will scroll the top of a ranked list all afternoon. The interesting material is underneath. Read twenty profiles the system set aside, and if three of them are people you would ring, the ranking has just said something honest about itself.
The same question works at every layer, and these days it is close to the only one I open with. Can it show where the answer came from? The sources behind the market picture. The evidence that this person does this work. The trail behind the number. The reason a candidate landed in one pile and not the other. A sourcer who can see why somebody surfaced writes a first line that lands and briefs a manager with specifics. A sourcer holding an unexplained list forwards it and hopes, which is the position of anyone who has ever been emailed a spreadsheet by an agency and told to trust it.
Trial it on the req nobody wants
Never trial on an easy req. The pull towards it is enormous, because an easy req makes the tool look clever and makes whoever championed it look clever, and the whole exercise turns into theatre that ends in a signature.
Take the search that has been open since spring. The one where the hiring manager said, out loud, that he wanted a Stripe engineer but cheaper. Then sit with four questions that have uncomfortable answers. How many of these names did I have already, and if the overlap is high, fine, that means the thing agrees with me, but it had better also be putting up people I could not have reached on my own. Of the new names, is the reason legible without ringing the vendor? Pick five and try to make contact today, not next week, today. And where it got somebody wrong, was the mistake explainable? An explainable miss gets corrected.
An arbitrary one comes back every month for the length of the contract. Run that on two reqs rather than one, and make the second a role you have never filled before. A tool that only performs where you already hold the judgment is a tool you are subsidising with your own experience.
Recruiters who put AI recruitment tools through a trial shaped like that decide faster and regret it less, because the trial is measuring the work rather than the pitch.
A hard req also exposes the overlap already sitting in the stack. A year in, the research tool has grown contact features, the screening tool has started sourcing, both push into Greenhouse in slightly different shapes, and someone on your team has gone back to a spreadsheet because it is the only place the full picture lives. Settle on which layer is the record for a candidate and which one owns the contact data. Settle where a sourcing note goes so the person screening can read it. None of that gets demoed. It is what breaks when whoever set it up leaves.
That payments req from the demo took another four months. The search that eventually cracked it started with a fortnight of reading conference proceedings, went through a maintainer thread about a memory leak, and finished on a Thursday phone call with a man who has never kept a profile anywhere.
