MMOOZZGG.com

Analysis

AI in recruitment: what real hiring and onboarding cases show in numbers

Medium19 min read20 sources

By the companies’ own numbers, AI in recruitment helped Fix Price and X5 where candidates waited too long for a reply. At Dodo Pizza it exposed the interviewers’ mistakes, and where the cases say who decides, a person does. Nearly all of it comes from high-volume hiring of frontline staff, and most numbers are the companies’ own claims. My advice to an HR director: start at the step where applicants drop out, not with choosing the candidate.

AI in recruitment: what real hiring and onboarding cases show in numbers
On this page · 11 sections

What recruitment with AI has shown: the short answer

By the companies’ own numbers, AI in hiring helped where it ended the wait. At Fix Price, a Russian discount chain, the robot replies within about a minute [1]. The company says its robots selected 23% of all frontline candidates in 2023 and sent 50% of candidates to interviews in 2025 [2]. The company does not say 50% of which candidates. In the cases that disclose it, people kept the decision about the person.

AI in recruitment is software that takes over steps of the hiring process a recruiter did by hand: resumes, first calls, booking and interview review. Here I count only published cases that name the step, the period and a number. X5, a Russian grocery retail group, tested an AI recruiter in one unit. There, AI fills more than 20% of office vacancies, and the hiring manager still decides (Retail.ru, 26 January 2026) [5].

For 20 years I led software development for Walmart, Disney, Apple and Microsoft. I also hire people myself, so I read these cases as an employer. I believe AI in HR already has a track record. Apart from the X5 office trial, nearly all of it comes from high-volume hiring of shop staff, warehouse workers, cashiers and couriers. The numbers are mostly claims by the companies and their vendors, and below I look at what stands behind them.

Five cases: who did what, and which numbers hold up

Of the five cases, four are Russian, and all four rest on company claims. The fifth, from the United States, ended in a 2023 settlement of a lawsuit brought by the EEOC, a U.S. regulator [10].

The tools in these HR processes are of four kinds, and three of them use AI. Along the funnel, X5’s AI recruiter sources candidates, and Fix Price’s robot screens and books them. Dodo Pizza, a Russian pizza chain, uses a system that reviews interviews, and iTutorGroup’s filter was a fixed age setting. None of these cases uses AI to predict how a hire will perform.

Company and startHiring stepWhat the system doesNumberType of numberSource and date
Fix Price, since December 2022Application, call, interview bookingA robot screens resumes, calls, books interviews and calls back those who missed one1.5 million resumes, 1 million calls and 90,000 messages in total. 23% of all frontline candidates selected in 2023; 50% of candidates sent to interviews in 2025, base not stated. About 2,500 bookings a month. Set up in 3 monthsCompany claimThe cloud platform vendor’s own case study, not an independent audit, 13 June 2024 [1]; ComNews, 4 September 2025 [2]
X5, trial in one unitSourcing, resumes, first interviewAn AI recruiter finds candidates and runs the first interview; a manager decidesOver 20% of office vacancies; selection quality improved “by up to 85%”, with no method; hiring pipeline doubledCompany claimRetail.ru, 26 January 2026 [5]
Dodo Pizza, since June 2024InterviewAnalyzes recordings of in-person interviews and recruiting center calls135 interviews in a month, mostly by phone; over 2,000 ratingsCompany claimComNews, 31 July 2024 [3]
Lenta, Fast Hiring projectCandidate paperworkThe company has not said how AI is usedPaperwork in 15 minutes instead of three daysCompany claim, no baselineComNews, 4 August 2026 [4]
iTutorGroup, United StatesApplication intakeRule-based software, not AI, rejected applicants by ageOver 200 rejected; $365,000 settlementOutcome of a regulator’s caseEEOC press release, 11 September 2023 [10]

The details are in our case studies of Fix Price, Dodo Pizza and Lenta, a Russian hypermarket chain. Our news story shows how X5 moved part of its hiring to an AI recruiter. What follows is not a retelling but the causes: why it worked in these companies and where checking the numbers stops.

Reason one: AI won where people could not reply in time

As I read the case study, AI won at Fix Price because the high-volume hiring team could no longer keep up with the chain’s growth. In the five years before the case study, that talent acquisition team doubled to 70 people [1]. Demand for shop and warehouse staff kept rising [1], and ComNews presents Dodo Pizza’s project as a fight against the staff shortage [3].

The robot closed exactly this gap. The company says it replies in about a minute, handles thousands of resumes a day and calls the chosen candidates itself [1]. If a candidate misses an interview, the robot gets in touch again and offers another time. The recruiter sees a contact report in the database (ComNews, 4 September 2025) [2].

The company describes the effect on headcount this way. HR director Anton Maksimenko said the chain opened about 1,900 of its own stores in two and a half years [2]. In his words, the robot lets the company avoid growing the HR team and takes routine work off its specialists. By the company’s account, the robot replaced the recruiters it would have hired, not the ones it already had [2].

X5 reached the same point from another side. In its trial, applications are handled around the clock across time zones, and the company reported that its hiring pipeline doubled [5]. These are office roles, not frontline ones, but the mechanism is the same: the candidate does not wait for the morning.

In my view, the cause is simple. In high-volume hiring, a candidate goes to whoever answers first, and AI removes the wait rather than choosing better than a human. One caveat matters: Fix Price explains its 2025 figure of 50% with three things at once [2]. They are AI, wider sourcing and a better way of working with candidates, so not all of it can go to the model.

The 2024 case study adds one more detail [1]. The chain writes and tests its bot scripts itself and sees conversion at every step of the conversation. That is how it finds errors in the robot’s logic and corrects the scripts. It brings in more candidates when new stores open and fewer once vacancies are filled.

Reason two: a person makes the hiring decision, and Russian law makes that the default

In every Russian case that discloses it, a person decides whom to hire or reject. Russian law allows a machine-only decision only with written consent or under a federal law. X5 left the hiring decision to the hiring manager [5]. Fix Price chose a system that leaves the live interview to its recruiters [1].

The provision is Article 16 of Russia’s personal data law (152-FZ of 27 July 2006, version of 26 July 2026) [6]. Part 1 bans decisions that have legal effects on a person, or otherwise affect their rights, when based only on automated processing. Rejecting a candidate by software alone falls under this provision.

Part 2 allows such a decision only with the person’s written consent or where a federal law provides for it along with safeguards. Part 3 requires the company processing the data to explain how the decision is made and what it may lead to. The company must also let the person object and explain how to protect their rights. Under Part 4, an objection must be answered within 30 days.

For employees, Russia’s Labor Code (Federal Law 197-FZ of 30 December 2001, as amended on 25 May 2026) says the same [7]. Article 86, point 6, bars an employer from basing a decision on data obtained only through automated processing or electronic means. Point 3 of the same article requires collecting data from the employee directly. Our checklist on 152-FZ lists what to ask a vendor about personal data.

The market seems to agree. The Russian outlet Sostav retold an Izvestia report on a survey by two Russian HR services [11]. In it, only 5% of employers trust an algorithm with the final hiring decision (Sostav, 12 April 2026) [11]. The survey’s sample size and dates are not given.

In the EU, GDPR Article 22 already bars machine-only decisions that significantly affect candidates, except for contract, law or explicit consent [17]. The AI Act, Regulation (EU) 2024/1689, adds a second layer. It lists AI systems that analyze and filter job applications as high risk in Annex III, point 4(a) [8].

High-risk rules for hiring apply from 2 December 2027 under Regulation (EU) 2026/1744 of 8 July 2026, the Digital Omnibus on AI [19]. It amended Article 113; the European Commission page, updated 19 October 2026, gives the same date [9].

Reason three: almost all case numbers are claims, not measurements

Nearly all the numbers in these cases are claims by companies and vendors. None reports how many recruits stay after three months or what recruitment cost before and after. That is the key answer to the argument that “it already works for everyone.”

Take the Fix Price case study on the cloud provider’s site, dated 13 June 2024 [1]. The headline promises almost 15,000 candidates in 1.5 years, while the text itself gives about 2,500 interview bookings a month [1]. Spread over 18 months, 15,000 gives about 830 a month, a third of 2,500. The case study does not say whether both count the same thing.

A second example comes from the same chain. The phrase “efficiency doubled” means the share of candidates the robot sent to interviews, not recruitment quality [2]. X5 says the AI improves resume selection quality “by up to 85%” but does not say how this was evaluated [5].

Large surveys do not give a baseline either. The SHRM study The State of AI in HR 2026 draws on answers from 1,908 HR professionals (SHRM, 31 March 2026) [12]. Its public page shows no shares: the full study is open to members only.

I see the cause in who writes the case studies. Companies and vendors publish them, and they care about scale, not a baseline for comparison. My advice: evaluate outcomes against your own baseline, not other people’s percentages.

Where automated screening failed: a case that reached a regulator

AI in recruitment: what real hiring and onboarding cases show in numbers

The clearest failure that reached a regulator is iTutorGroup’s: it agreed to pay $365,000 to settle the EEOC’s hiring discrimination lawsuit [10]. The U.S. Equal Employment Opportunity Commission (EEOC) said its software rejected women aged 55 or older and men aged 60 or older. That way, the company rejected more than 200 qualified applicants in the United States [10].

Besides the payment, the settlement requires the company to provide training to everyone involved in recruitment and adopt a new anti-discrimination policy. It also bars asking candidates for their date of birth. The EEOC will monitor compliance for at least five years. The suit rested on the U.S. Age Discrimination in Employment Act (ADEA), which bans age discrimination against people 40 or older, hiring included [10].

In my view, the failure did not come from a complex model but from a rejection setting that nobody checked. The software was a fixed age filter, not a learning model. With no human involved, it decided exactly what it was told to decide. So AI does not remove bias by itself: it applies the bias of its rules to every application, as iTutorGroup’s filter did [10].

In Russia, such a rejection with no human involved and no written consent would breach Part 1 of Article 16 of 152-FZ [6]. The lesson: check filter settings before you start, and review rejections by sampling, split by age and gender. X5 describes a similar check: when it started its AI recruiter, the team checked every sample, according to the managing director of X5 Business Support [5]. The model was then retrained, and its algorithms and prompts were refined.

The math: at what volume a robot recruiter makes sense

On Fix Price’s numbers, the robot makes sense for the chain, not one store: a store gets under half a booking a month. This is my estimate, not the company’s measurement. The inputs are about 2,500 bookings a month, 70 high-volume hiring specialists and more than 6,600 stores [1].

MetricValueHow it was calculatedWhat it means
Interview bookings a monthabout 2,500Cloud provider case study, 13 June 2024 [1]The company’s own number
Bookings per store a monthabout 0.42,500 / 6,600 storesEach store sees few bookings; the chain creates the value
Bookings per HR specialist a monthabout 362,500 / 70 specialistsFewer than two bookings per person per working day
A sample flow of 10,000 candidates a month2,300 selected and 5,000 sent to interviews10,000 × 0.23 and 10,000 × 0.5An illustration of the shares, not the chain’s flow: 23% is of all candidates, 50% has no stated base, so the two do not compare
Candidates a month, by the case study’s headlineabout 83015,000 / 18 monthsA third of 2,500: the two numbers have different bases
Hours on first contact a week, hypothetical companyabout 33200 applications × 10 minutes / 60Almost a recruiter’s full week; with 20 applications, about 3.3 hours
Break-even price of a tool a monthyour numberhours a week × cost of a recruiter’s hour × 4.3 weeksIf the tool costs more, it does not pay off on first contact alone

The main difference from hiring without AI is the clock. A reply in a minute versus a reply in office hours makes the difference of a whole night. For example, a candidate applies at 20:00. A recruiter who starts at 9:00 answers 13 hours later, while the robot answers in a minute. In that time, a candidate for a frontline job can go to whoever replied first.

My threshold estimate: below 4 hours of first contact a week, as with 20 applications in the example, a robot recruiter is not needed. A chatbot with an interview calendar and a quick reply from the owner will do. This is my math, not a measurement, and our news story on X5 covers what a company with five vacancies can take from it.

Before you start, count three numbers: applications a week, the share left without a reply after a day, and the cost of a recruiter’s hour. If many go unanswered and first contact eats up the week, the robot closes exactly that gap.

An unexpected result: AI checked the recruiters, not the candidates

For an HR director, I find Dodo Pizza’s insight the most useful: interview analysis exposed the interviewers’ own mistakes. Since June 2024, the chain has recorded in-person interviews for three roles, manager, cashier and courier, and calls from its recruiting center [3]. In the first month after the June 2024 start, the system analyzed 135 interviews and produced more than 2,000 ratings.

The finding is about the interviewers. Only half of the candidates hear a full account of the company culture and bonuses, and fewer than 50% are asked about their motivation [3]. In the same 2024 conversations, the standards on paperwork, schedules and pay were met in 100% of cases. So the AI did not make hiring fairer by itself; it showed where the people running the interviews were uneven.

Here is a cause that rarely gets named: a company loses some candidates at the interview itself, because of what its recruiter says. The vendor’s analysts estimate such losses at up to 30% of applicants, but this is their estimate with no baseline [3]. You can only see it by analyzing the conversations themselves.

Anna Lifanova, executive director of Dodo Pizza’s corporate network, gave the reason the chain spends on this. She said that after construction costs, the HR unit takes the largest share of investment, and that share keeps growing [3].

For hiring in the EU, this kind of technology has a limit. Article 5(1)(f) of the AI Act has banned emotion recognition in the workplace since February 2025 [8]. A system that analyzes a candidate’s tone of voice has to take this into account.

AI in HR after the hire: what onboarding has confirmed and what is still a plan

I found no onboarding case with retention numbers in open primary sources: the tools are confirmed, the results are not. Super Lenta, part of the Lenta group, built a chatbot for routine staff questions [4]. When its knowledge base has no answer, the question goes to an expert, who replies in real time.

At Dom Lenta, neural networks help build e-learning courses for staff, adapt the material to different categories and analyze the results [4]. Decisions on range, prices and strategy stay with people. The group’s chains created no new AI positions, but specializations appeared within existing roles.

Lenta adds that work with neural networks depends on setting tasks well, then checking and refining the result. Dodo Pizza only plans to digitize the employee path from the first call to training and growth [3]. It is still thinking about onboarding and internships. How an AI agent for onboarding new hires works is covered in a separate analysis.

Why hand routine work to a machine at all? In my book How Not to Go Bankrupt!?, I wrote that automating repetitive, low-value tasks frees time for creative work and personal involvement [13]. In the same book I wrote about retraining, which concerns the labor market, not a new hire’s first weeks. Encouraging people to retrain and learn new AI-related professions helps them adapt to changing market demands and softens the downsides of automation.

So the honest conclusion is this: AI already supports onboarding, but open sources hold no retention numbers for new employees. I do not quote numbers without a baseline here.

What an HR director should do: six steps and a rethink of hiring

  1. Measure your baseline before you start: time to reply, the share of applicants who reach an interview, and the cost per hire. Without these three numbers, you will have nothing to compare with after three months.
  2. Start with the reply and the interview booking, as Fix Price did. The chain set up its system in three months, and its recruiters kept the live interview [1]. Before signing, ask the vendor in writing who makes each rejection, what candidate data it stores and where, and whether you can export the rejections.
  3. Leave the hiring decision to a human. In Russia, Article 16 of 152-FZ requires this unless the candidate gives written consent or a federal law provides for it [6]. Such a decision also needs an objection process and a reply within 30 days [6]. For employees, Article 86, point 6, of the Labor Code adds the same rule [7]. Abroad, check GDPR Article 22 in the EU [17], UK GDPR Articles 22A–22D [20] and the AI Act [9].
  4. Check rejection settings and samples before and after you start. X5 reviewed every sample when it launched its AI recruiter [5]. Without such a review, your filter may repeat the iTutorGroup story, where unchecked software rejected candidates by age. Take date of birth out of the form your filter reads: the EEOC settlement bars iTutorGroup from asking for it [10].
  5. Review your own interviews, as Dodo Pizza does [3]. There, only half of the candidates or fewer heard about the culture or were asked about motivation. Recordings and ratings show where you lose people after they apply.
  6. After three months, compare the outcome with the baseline from step one. Look at time to reply, the share who reach an interview and retention of new staff, not at the number of resumes processed.

My view is simple: use AI in hiring to rethink how you hire, not to hire the old way faster. In a 2026 talk, my first advice to CEOs was to freeze hiring and rethink how they hire staff [14].

For me, the key in hiring is not a diploma but skills. In a 2025 conversation I put it this way. The most important task is to see whether a person can do the job you are hiring them for [15]. In my book I wrote about the same idea: in hiring, I follow the same logic and try to pick smart, capable people [13].

As an employer, I now get resumes written by AI, so candidates send me videos of their work. I described this in a 2025 talk [18], when the queue for an interview with me ran to the end of December. Ask for a sample of real work, not only a resume.

Responsibility also stays with the person. In a 2025 conversation I recalled my talk at the Financial University, a Russian university [16]. There I said that if I use AI as a tool and chose it myself, I am responsible for the decision. That is why, for me, AI is a professional’s assistant, not a replacement.

The bottom line: AI takes the flow, people keep the decision

The conclusion is the same as at the start: AI takes the flow of applications, and people keep the decision about the person. The four Russian cases used AI at different steps: the reply at Fix Price and X5, the interview at Dodo Pizza, paperwork at Lenta. Each is a step where applicants drop out, so start there and measure your baseline first.

What comes next for AI in hiring already has dates and plans. In the EU, high-risk rules for these tools apply from 2 December 2027 [19]. Dodo Pizza plans to digitize the employee path [3], and Lenta’s chains created no AI positions, only new specializations within roles [4]. Fix Price says the robot lets it avoid growing the HR team [2]. In my view, the recruiter’s work moves from the first call to the decision and its check.

Questions and answers

Did AI really cut candidate paperwork at Lenta to 15 minutes?

That is the company’s own claim, and there is no baseline to check it. According to ComNews on 4 August 2026, AI cut candidate paperwork in Lenta’s Fast Hiring project from three days to 15 minutes [4]. How many candidates, which stores and what period this covers is not stated [4].

How many companies in Russia already use AI in hiring?

There is no exact share, and the estimates count different things: X5 cites about 5% of companies using AI systematically but names no source [5]. The survey Sostav retold from Izvestia on 12 April 2026 gives no sample size or dates [11]. In it, 12% of companies use AI at several hiring stages, 23% in single tasks and 18% are testing it [11]. Most often, AI handles first resume screening: 68% of respondents said so [11].

Can AI reject a candidate on its own?

In Russia, as a rule, no: Article 16 of 152-FZ bans decisions on a person’s rights made only by automated processing [6]. The exceptions are written consent or a federal law. In the EU, GDPR Recital 71 names e-recruiting without human intervention as such a decision, with a right to human review and challenge [17]. UK GDPR Articles 22A–22D, from the Data (Use and Access) Act 2025, have governed UK automated decisions since 5 February 2026 [20].

What are the most common AI mistakes in hiring?

The survey Sostav retold names one: 28% of respondents pointed to algorithm errors that can cost a company suitable candidates [11]. Its sample size and dates are not given. A new risk runs the other way: resumes written by AI, and X5 says its AI recruiter flags them [5].

Should you tell candidates they are talking to a robot?

I would tell them: silence is risky, although I found no direct Russian rule on disclosing a robot. If a machine makes the decision, Part 3 of Article 16 of 152-FZ requires explaining the procedure and its consequences to the candidate [6]. According to the European Commission page updated 19 October 2026, the AI Act’s transparency rules in Article 50 apply from August 2026 [9]. People must be told they are talking to a machine.

Can you record interviews and analyze them with AI?

Dodo Pizza has recorded interviews since June 2024 [3]. Its vendor says it stores no biometric data and identifies people by their application form details [3]. Recording needs a lawful basis such as consent the candidate may withdraw “at any time” (GDPR Articles 6(1)(a), 7(3)) [17]. Identifying someone by voice biometrics needs explicit consent or another exception (Articles 4(14), 9). Keep recordings for a stated purpose, no longer than necessary (Article 5(1)(b), (e)).

What can a company with five vacancies take from this?

It probably does not need a robot recruiter: the robot makes sense with a volume of applications that a small company does not have. A quick reply and a same-day interview booking help more, and the owner keeps the hiring decision. If five vacancies bring, say, 20 applications a week, the math above gives about 3.3 hours of first contact a week. A chatbot with a calendar can handle that.

Sources and fact-checking

Links lead to the sources’ own pages, checked on 20 November 2026. Checked by the editors on 21 November 2026.

  1. yandex.cloud (yandex.cloud)
  2. ComNews (comnews.ru)
  3. ComNews (comnews.ru)
  4. ComNews (comnews.ru)
  5. Retail.ru (retail.ru)
  6. consultant.ru (consultant.ru)· Article 16 also on legalacts.ru: https://legalacts.ru/doc/152_FZ-o-personalnyh-dannyh/glava-3/statja-16/
  7. consultant.ru (consultant.ru)· Article 86 also on legalacts.ru: https://legalacts.ru/kodeks/TK-RF/chast-iii/razdel-iii/glava-14/statja-86/
  8. eur-lex.europa.eu (eur-lex.europa.eu)
  9. digital-strategy.ec.europa.eu (digital-strategy.ec.europa.eu)
  10. eeoc.gov (eeoc.gov)· The ADEA rule: EEOC page ‘Age Discrimination’, https://www.eeoc.gov/age-discrimination, read on 21 November 2026.
  11. Sostav (sostav.ru)
  12. shrm.org (shrm.org)
  13. litres.ru (litres.ru)
  14. YouTube (youtube.com)
  15. YouTube (youtube.com)
  16. YouTube (youtube.com)
  17. eur-lex.europa.eu (eur-lex.europa.eu)· Original text as adopted on legislation.gov.uk, read on 21 November 2026, Article 4: https://www.legislation.gov.uk/eur/2016/679/article/4/adopted. Article 5: https://www.legislation.gov.uk/eur/2016/679/article/5/adopted. Article 7: https://www.legislation.gov.uk/eur/2016/679/article/7/adopted. Article 9: https://www.legislation.gov.uk/eur/2016/679/article/9/adopted.
  18. YouTube (youtube.com)
  19. eur-lex.europa.eu (eur-lex.europa.eu)
  20. legislation.gov.uk (legislation.gov.uk)

Found a mistake? Tell us

Publications

Mike Mozg

AI for accountants: what to hand an assistant every day and what to check yourself

AI for accountants means an assistant that drafts letters, summarizes new guidance and matches two exports. A person still signs, posts and files. The UK audit regulator FRC put it plainly on 30 March 2026: the human auditor always answers for the work, whatever tools were involved [6]. In my view, an assistant pays off where checking its answer is faster than doing the work yourself.

Article16 min
Contact us