An open position costs you every day. Where automation in the hiring process actually saves time.

Why manual screening puts you at a disadvantage

Anyone who has screened fifty applications in one afternoon knows the pattern. You read the first ones carefully. From the twentieth on, the CV layout, your energy level and the name on the first page decide. That is human. It is also why good candidates get overlooked.

The second disadvantage is the wait. In many companies, days pass between an application arriving and the first substantive reply, sometimes weeks. Good candidates rarely apply to one company only. Whoever replies last often ends up talking to someone who has already signed elsewhere.

Both cost money, even though neither shows up on an invoice. An unfilled position means work is left undone, colleagues pick it up on top of their own or an order is turned down. How much that is per day depends on the role. For a developer missing from a client project you can roughly derive it from the day rate. For a back-office role, from the overtime of everyone else. Run that calculation for your own open positions before you read on.

In projects we treat recruiting like a sales process. The application is an enquiry. The signed contract is the close. And as in sales, the time to the first reply decides more than the salary does.

Why the maths has changed

The economist Ronald Coase described in 1937 why companies do some tasks themselves and buy others in. It comes down to what the route via the market costs in effort, compared with the effort of doing it internally.

Applied to recruiting: for a long time it was sensible for many companies to hand screening to a recruitment agency. The software to do it in-house was expensive and the staff to run it were missing. The agency charged a share of the annual salary for it, which is a substantial amount for well-paid roles.

Since language models can read and classify CVs, the internal route has become far cheaper. Reading and pre-sorting one application costs a fraction of a cent in computing. What remains is the effort of building the workflow properly once. Coase’s calculation therefore tips towards doing it yourself. That does not hold for every company. If you fill two positions a year, you do not need an automated workflow. If applications come in every month, you do.

Why keyword search overlooks good candidates

Classic applicant tracking systems search by keyword. If the job ad says “project management” and the CV says “Scrum Master”, the system does not find the person. Conversely, someone who copied every keyword from the ad into their CV ends up at the top.

Language models compare meaning instead of strings of characters. To do that, a text is translated into a series of numbers, a so-called embedding, in which similar meanings produce similar numbers. That is how the system recognises that “React experience” and “frontend development with modern frameworks” mean almost the same thing. The workflow we build in projects has four steps:

1. Intake. The application is captured whether it arrives by email, through a form or via a job board. At night and on weekends too.
2. Extraction. A language model pulls the important details out of the PDF: skills, work experience, availability and salary expectation, if stated.
3. Matching. The profile is compared with the job description, by meaning rather than by keyword.
4. Short report. HR does not receive fifty PDFs but a sorted list with a summary per candidate. For example: strong technical fit, salary expectation above budget, three months notice.

That takes seconds instead of days. In fairness: the match is only as good as the job description. Feed in an ad with twenty mandatory requirements and you get a list on which nobody fits.

A reply on Saturday night

A strong candidate sends her documents on Saturday at 11 pm. In most companies nothing happens until Monday, often until Wednesday. In an automated workflow she receives a reply within minutes that responds to her actual application. If the profile fits the role well, the reply contains a link where she picks a slot for a first conversation.

A human still conducts that conversation. The machine only made sure it happens before the competition opens its inbox on Monday. In projects we see the time from application to first interview shrink from weeks to days. How much depends on how long your team takes today. The logic is the same one we use for following up on sales enquiries.

Step	Manual	Automated
First screening	A few seconds per CV, depending on the day	Every application is read in full
Matching against the role	Keywords	Meaning
First reply	Days	Minutes
Scheduling	Emails back and forth	Booking link in the first reply
Decision	Human	Human

Anonymisation, discrimination law and the rules for AI

A common objection: does the machine not discriminate? It can, if it learns from old hiring decisions in which humans discriminated. That is why we do not let the model learn from your hiring history. It only compares application and job description.

On top of that comes a step that is rarely done in manual screening. Before matching, the workflow removes photo, name, date of birth, gender and origin from the documents. The sorting is based on experience and skills. The team sees the identity only when it invites someone to an interview. Germany’s General Equal Treatment Act, the AGG, demands exactly that: no disadvantage because of these characteristics. A workflow that never sees the characteristics is easier to prove than the good intentions of a case worker.

Two rules you need to know. Article 22 of the GDPR does not allow a rejection made by a machine alone. The software may sort and suggest. A human has to make the decision and take responsibility for it. One click confirming fifty suggested rejections at once does not count as a decision. And the EU AI Act classifies AI used in candidate selection as a high-risk system. The obligations that come with that, such as documentation, logging and human oversight, are coming. The exact date is in motion, the direction is not. If you build the workflow today with logging and a human decision, there is little to retrofit later. How to protect such workflows against manipulation is covered in the article on the security of automated workflows.

What this means for management

Your HR team was hired to win people for the company, convey its culture and negotiate contracts. Instead it spends a large part of its time opening PDFs, coordinating appointments and typing data from one system into another.

Automation does not replace the recruiter. It gives back the hours that currently disappear into routine. Companies that have built their workflows this way do not hire fewer people. They hire faster and with less effort per position. Why that applies to the whole company is covered in the article on revenue per employee.

What you can check this week

1. Take the last ten applications and look up how many days passed until the first substantive reply. Not the automatic acknowledgement, but the reply from a human.
2. Ask your HR team how many hours per week go into screening and scheduling. The answer is usually higher than expected.
3. Check whether your rejections are currently decided by a human or whether a filter in the applicant tracking system already does it. If the latter, you already have the Article 22 problem, without any AI.

If you then want to know which steps can be automated in your case, write to us. We will also tell you if the setup is not worth it at your application volume.