AI-written job applications: how a small business hires when every application sounds good

You post a job, and nearly every application sounds good: error-free, friendly, tailored precisely to your ad. Often, AI is behind it. 65 percent of job seekers registered with a Swiss regional employment center (RAV) have used AI for their applications, and a third use it regularly. That is the finding of a survey published in April 2026 by AMOSA, the labor market research unit of the cantons of Eastern Switzerland, Aargau, Zug and Zurich.

For your hiring, that means the cover letter has had its day as a selection criterion. Make it optional. Instead, test what demonstrably predicts job performance better and what AI can hardly polish: a short call with fixed questions early in the process, and a small task the candidate solves and explains during the interview. In the example further down, this takes the same time as before, and you talk to more people.

Two in three write with AI, and hardly any employer has changed how it selects

AMOSA surveyed 5'787 job seekers who were registered with a RAV in its cantons in early August 2025, plus 663 companies with at least ten employees. The figures that matter for your hiring:

• 61 percent of job seekers use AI for their resume and cover letter: to write, correct, translate or tailor them to the job.
• Employers estimate that 39 percent of applications were written partly or entirely with AI. Small companies estimate 34 percent, mid-sized ones 43 and large ones 49 percent.
• Just over half of the employers receiving such applications say the documents increasingly look standardized (52 percent) and that it has become harder to judge how much of the work is the candidate's own (51 percent).
• Only 7 percent of these employers have changed their selection process as a result. 3.6 percent of all employers surveyed use AI for their own screening.

Two caveats apply. The survey covered people registered with a RAV, not everyone who applies for jobs. And the sample of employers is not representative. Even so, the direction is clear: most candidates write with AI, and almost everywhere, selection works the way it always has.

Using AI is not misconduct. Roughly one in ten employers surveyed uses AI to write or improve its own job ads. Candidates do the same with their applications. A good process is built with that in mind.

The effect is strongest for jobs that attract many applications. According to the jobs.ch AI Report 2026, roles in administration, finance, sales, marketing and IT have become much easier to fill. The scarcity index for these occupations, which AI affects most, fell by 69 percent between 2023 and 2025. The index compares demand with the talent available. In occupations that AI affects less, such as nursing and the skilled trades, it rose by 19 percent on average. There, the challenge is finding people at all, and this article only helps at the margins.

What a cover letter still tells you

The AMOSA data show a striking pattern. Job seekers who use AI regularly receive around 48 percent more interview invitations than those who don't use it. The figure comes from statistical models that account for differences between the groups. As a simple average, it was 1.23 versus 0.86 invitations over four weeks. But a first analysis about six months after the survey suggests that regular AI use does not shorten the job search. The study therefore describes AI as a door opener.

This is a correlation from a survey, not an experiment. It does, however, fit with two research papers from US universities that analyzed data from a large international freelance platform. Both are working papers, which means they have not yet been published in a peer-reviewed journal:

• In April 2023, the platform introduced an AI tool that tailors cover letters to the job post. Afterward, a well-tailored cover letter said much less about who got an offer: the correlation fell by 79 percent. Employers paid more attention to signals that are harder to fake, such as a freelancer's track record on the platform (Cui, Dias and Ye, 2025).
• A second team used an economic model to simulate what happens when written applications no longer reveal anything about ability. In that simulation, the strongest fifth of candidates is hired 19 percent less often and the weakest fifth 14 percent more often (Galdin and Silbert, 2025).

Freelancers on a global platform are not the same as applicants for an administrative role in Winterthur. Still, AMOSA sees its findings as consistent with an international study: the cover letter is losing value as a signal, and employers are turning to other criteria.

How well does a cover letter predict who will do the job well in the first place? There is no validity estimate for the cover letter. It does not appear as a separate method in the major review of selection methods discussed below, and we found no study that establishes such a value. Inviting candidates on the strength of their cover letter today therefore selects, to a large extent, for skill with writing tools. That is a useful skill. For many roles, it is not the deciding one. And the trend is set to continue: AMOSA expects AI agents that could search for suitable jobs, write applications and submit them on their own.

Several large Swiss employers have already made the cover letter optional, and one recruits its apprentices without any application file at all. Many companies, especially small and mid-sized ones, still ask for one as standard, the Swiss business magazine Bilanz reported in February. An HR expert quoted there explained why (our translation): "There simply is no reasonable alternative to the cover letter yet." This article describes one.

Why detecting AI text doesn't help

The obvious reaction is to detect AI-written text and screen it out. That works poorly, and it hits the wrong people.

• OpenAI, the company behind ChatGPT, released its own detection tool in early 2023. On an English test set, it correctly identified 26 percent of AI-written text and wrongly labeled 9 percent of human-written text as AI. According to OpenAI, it was very unreliable on texts under 1'000 characters and was recommended for English only. OpenAI shut the tool down in July 2023 because of its low accuracy (OpenAI).
• Researchers at Stanford University ran 91 English essays by non-native speakers and 88 essays by US students through seven widely used detectors, a small sample. On average, the detectors flagged 61 percent of the non-native speakers' essays as AI-generated, while classifying the US students' essays as human almost without error. The simpler the word choice, the more likely a text was flagged as AI (Liang et al., 2023).

The tools have evolved since 2023. We found no independent measurement for German-language applications. In a multilingual labor market like Switzerland's, suspicion that hinges on simple language can hit people who write in a second language particularly hard.

Some also treat the German sharp S ("ß") as a sign of AI text, because Swiss Standard German doesn't use it. At most, an "ß" shows that nobody checked the text for Swiss spelling, or that the person learned to write in Germany or Austria. It is no reason to turn anyone down.

Then there is the law: anyone who processes personal data must make sure it is accurate (Art. 6(5) of the Federal Act on Data Protection). A note saying "AI text" in a candidate's file, produced by an unreliable tool, is therefore on shaky ground. The better question is whether what the application says is true. You check that in the interview.

What actually predicts job performance

Research expresses how well a selection method predicts later job performance as a correlation: 0 means no relationship, 1 a perfect prediction. A major review comes from Sackett, Zhang, Berry and Lievens, published in 2022 in the Journal of Applied Psychology. It recalculated the established values. Most turned out considerably lower than previously assumed, and the structured interview came out on top.

Method	Value	What to do
Structured interview	0.42	make it the core of your process
Job knowledge test	0.40	ask fixed job questions in the interview
Work sample, solved and explained on site	0.33	one task of 30 to 45 minutes
References	0.26 (from 1998)	only with consent
Unstructured interview	0.19	replace it
Years of experience	0.07	check as a minimum only
Cover letter	no value	make it optional

How to read the table: it lists the methods that are realistic for a small business. The value is the predictive validity, an average across many individual studies in different occupations, and it varies with the role and with how the method is used. Structured means the same questions for everyone, with scoring set in advance. The 0.40 applies to knowledge tests built specifically for the job; there is no separate value for job questions asked in an interview. The structured interview's value is more than twice that of the unstructured interview. Note the years of experience: at 0.07, they say almost nothing about later performance. For references, the review did not calculate a new value for lack of data; the 0.26 comes from the earlier review by Schmidt and Hunter.

The three strongest methods in the table have one thing in common: the answers happen in the conversation or on site, and AI is of little help there. With the cover letter, it is the other way around. That is our own assessment, based on the AMOSA usage figures.

A four-step process for a small business

You need neither an assessment center nor new software. The four steps build on the three strongest methods in the table.

1. Three questions instead of a cover letter

State in the job ad that a cover letter is optional, and list the three main tasks of the role. Ask for a resume and answers to three questions, three sentences at most each:

1. When can you start, and what workload are you looking for (as a percentage of full time)?
2. Which of the three main tasks of this role have you already handled yourself? One example in two sentences.
3. What needs to be right about this job for you to still enjoy working here in two years?

The answers may be written with AI. They give you statements you can check in the interview, and three short answers are quicker to read than a page-long letter.

2. A 15-minute screening call

Early in the process, call more people than you will later invite, for example ten instead of six. Decide in advance on two or three criteria the role truly requires, and ask everyone the same four questions in the same order:

1. What was your most recent task, and what exactly was your part in it?
2. About the example in your application: what was difficult, what did you do, and what was the result?
3. Workload, salary expectations and notice period.
4. What would you like to know about us?

Rate each criterion on a scale of 1 to 5 right after the call and write a short note. A phone call is enough; there is no need for video. Questions unrelated to suitability for the job, such as family planning or health, don't belong in it. More on that in the legal section below.

3. The interview with a small task

Invite the best three or four to a structured interview: the same questions for everyone, scoring set in advance, two people listening. Ask about specific situations from past work, and add two or three fixed questions on job knowledge from the role's daily work, asked orally or as a short test on site.

Add a task of 30 to 45 minutes that the candidate solves on site and explains to you. For an administrative role, for example: answer a made-up customer complaint and explain what they would clarify first. Three rules apply: a made-up case, no real work for your company, and a clear statement beforehand that the task is unpaid and how long it takes. Pay attention to the explanation. That is where you see who understands the work.

4. Written references, reference checks and a probation period with goals

In Switzerland, the application file usually includes Arbeitszeugnisse, the written references that former employers must issue on request (Art. 330a of the Code of Obligations). They come from the former employer, so a candidate's use of AI doesn't change them. The Sackett review gives no separate value for them. Read them for gaps and for the tasks they describe, and follow up in the interview and with the reference checks. According to AMOSA, closer checks of application documents, such as Arbeitszeugnisse, are among the changes some employers have made.

Before you hire, check two references. You need the candidate's consent first. The easiest way is to contact the people the candidate names. Two fixed questions are enough: What tasks did this person handle for you, and how well? Would you hire them again?

When the person starts, write down what should be achieved after one, two and three months. For a permanent position, Swiss law sets the probation period at the first month of employment. It can be extended in writing to a maximum of three months, and it is extended accordingly if the person is out sick or injured (Art. 335b of the Swiss Code of Obligations). During this time you see the real work. With clear goals, it becomes a fair test for both sides.

Does this take more time?

An example based on assumptions you should replace with your own figures: an administrative role, 60 applications, two people conducting the interviews.

Step	Before	New
Reading applications	60 × 5 minutes = 5 hours	60 × 3 minutes = 3 hours
Screening calls including notes	none	10 × 20 minutes = 3 hours 20 minutes
Interviews	6 × 60 minutes × 2 people = 12 hours	4 × 75 minutes with task × 2 people = 10 hours
References	none	2 × 20 minutes = 40 minutes
Total	17 hours	17 hours

The effort stays the same. But you speak to ten people yourself instead of six, and your decision rests on methods with a predictive validity of 0.33 to 0.42 rather than on written documents and an unstructured interview at 0.19.

What Swiss law requires

• Only questions about suitability. An employer may only process data about employees insofar as it concerns their suitability for the job or is necessary to perform the employment contract (Art. 328b of the Code of Obligations). The Federal Data Protection and Information Commissioner (FDPIC, known in German as EDÖB) applies this principle from the application stage. Questions about illness or pregnancy are only permissible if the role justifies them for specific reasons.
• Inform candidates. Candidates are told who processes their data and for what purpose. This also applies to information you obtain from third parties (Art. 19 of the Federal Act on Data Protection).
• References only with consent. Contact current or former employers only once the candidate has agreed. They may only share what matters for the new role, such as performance and conduct, and may not disclose the terms of the employment contract (FDPIC, in German).
• Unpaid work samples only if agreed and limited. In a case about unemployment insurance, the Swiss Federal Supreme Court stated that an agreement that certain work is unpaid is permissible within limited bounds if it is made expressly (overview of trial work, in German). If you use someone as a regular worker, however, you owe wages: an employment contract is deemed concluded when the employer accepts work that, in the circumstances, would only be expected for pay (Art. 320(2) of the Code of Obligations). A made-up case during the interview is therefore the safe route.
• Return or delete documents after a rejection. The FDPIC considers it acceptable to keep application data for up to three months after the rejection, so you can defend yourself against a discrimination claim under the Gender Equality Act (FDPIC FAQ, in German). You should keep it longer only with the candidate's consent and up to a date set in advance (FDPIC, phases of the employment relationship, in German).
• Automated screening. If a system screens out applications without a person making the decision, the company must inform the candidate, who can ask for the decision to be reviewed by a person (Art. 21 of the Federal Act on Data Protection). What this means for AI tools in recruiting is covered in our article on AI in hiring and the EU AI Act.

What you can do this week

1. In your next job ad, make the cover letter optional, list the three main tasks and add the three questions.
2. For the open role, define two or three criteria and write down the four questions for the screening call.
3. Prepare a task from the role's daily work as a made-up case that can be solved in 30 to 45 minutes.
4. Decide who deletes application documents after a rejection, no later than three months afterward.

If you'd rather not set this up yourself

clever hr works as your recruiting team, on your behalf, in your systems and under your name: applicant management, job postings, direct sourcing, and interviews and selection through to a shortlist with a clear, documented assessment. You can book a single module or the whole process. Every application stays in your systems, and hiring decisions remain your call. We bill by the hour at our published rate of CHF 175, with no fee tied to a hire. Our article on recruiting by the hour explains why we bill this way, and our recruiting service page gives an overview.

Let's talk about your next opening for 30 minutes: book an intro call.

---

Note: This article reflects the position in November 2026 and does not replace legal advice in an individual case. The time calculation is based on assumptions.

Sources (13)

Sources: AMOSA, KI in Jobsuche und Recruiting, April 2026 (in German) · Sackett, Zhang, Berry and Lievens, Journal of Applied Psychology 107(11), 2022 · Cui, Dias and Ye, Signaling in the Age of AI, working paper, 2025 · Galdin and Silbert, Making Talk Cheap, working paper, 2025 · OpenAI, AI text classifier, 2023 · Liang et al., Patterns, 2023 · jobs.ch AI Report 2026 (in German) · Code of Obligations, Art. 320, 328b, 330a and 335b · Federal Act on Data Protection, Art. 6, 19 and 21 · FDPIC, phases of the employment relationship (in German) · FDPIC, data protection FAQ (in German) · Arbeitsrecht aktuell, trial work, with Federal Supreme Court ruling 8C_726/2011 (in German) · Bilanz, February 20, 2026 (in German)