Recruiting, ATS and Artificial Intelligence: the complete guide to candidate selection

1. Modern recruiting: what it is and why digitise it

Modern recruiting is the cycle that runs from the need for a person to their first day at the company, managed so that every step leaves a data point and has an owner. Digitising it serves, first of all, to make readable what today lives in mailboxes, spreadsheets and the memory of individuals. Speed comes afterwards, and only if the steps are clear.

Recruiting, selection and talent acquisition

According to the Wikipedia entry "Recruitment" (CC BY-SA 4.0, accessed on 29 October 2026), recruiting comprises the activities by which an organisation identifies, searches for, evaluates, selects and hires people, for permanent or temporary roles. The Italian entry "Selezione del personale" (personnel selection) calls selection the process by which the company, through the human resources function, searches for, evaluates and places people suited to the role, and lists its stages: analysis of the position, recruitment, evaluation, placement. The two entries overlap, so we set a convention.

"Recruiting" here means the complete cycle, from the hiring need to the start date; "selection" the evaluation and choice part that belongs to it; "talent acquisition" the continuous approach, which builds attractiveness and a pool of interesting people before the position opens. This is our own definition, not a quotation.

The selection funnel

From the candidates' side, the process is a funnel with five steps, and at each one some people drop out.

1. Applications received from all channels: job ads, spontaneous applications, referrals.
2. Applications that pass screening.
3. Candidates invited to interview, in one or more rounds.
4. Finalists who receive an offer.
5. Offers accepted and people who actually join.

How many drop out at each step matters less than why: a missing requirement, a withdrawal, an unreachable candidate, a rejected offer. Without the reason, nothing improves. Measures in Recruiting analytics and KPIs; operational stages in How the end-to-end recruiting process works.

Where it breaks down with email, Excel and disconnected tools

An example with made-up data: the line manager receives forty CVs as PDFs by email, forwards eleven to a colleague, replies to six with a single line and forgets the rest. The recruiter keeps a spreadsheet with a hand-written "status" column, in two copies, one on the desktop. A candidate asks where things stand and nobody can say in under half an hour.

• The same candidate appears in three places and nobody knows who has already spoken to them.
• The evaluation criteria sit in the head of whoever reads the CV and change from one candidate to the next.
• Calendar, evaluation forms and offer letter live in different tools; the new hire's data is re-keyed by hand into the employee records.
• If a candidate asks which data the company holds about them, finding it means a hunt through attachments and folders (Privacy, GDPR and the AI Act in recruiting).

Who does what

HR defines rules, budget and compliance. The recruiter runs the process day by day. The hiring manager, who is responsible for the position, states the need, evaluates on the merits and chooses together with HR. The candidate experiences timings, communications and outcome. In SMEs two roles often coincide; the responsibilities remain distinct and it is worth writing them down.

What you want to achieve

The goals are hiring quality, speed, candidate experience and compliance, and they partly contradict one another: speeding up without written criteria lowers quality, adding checks slows things down. Start from compliance and criteria, which cannot be recovered afterwards, and look for speed in the repetitive administrative work done by hand. The same applies to artificial intelligence: it supports whoever selects, and the decision stays with a person (AI in recruiting).

2. How the end-to-end recruiting process works

End-to-end recruiting has six stages: hiring need and requisition, job profile and publication, screening and shortlist, interviews and evaluations, offer and signature, onboarding. Each stage has an owner, produces a data point and hands a result on to the next. The points where someone approves or chooses, the gates, must be decided before the search is opened, because that is where the process stalls or goes wrong.

Stage by stage: owner, data, result

The table is an EcosAgile working model, not a published standard. In small companies several roles coincide; KPI names have different definitions in the sources; those in the table in Recruiting analytics and KPIs are defined there, the others are defined in writing with a start, an end and a unit.

Who decides, and when the rule is written

There are really five gates: who approves the position, who approves profile and criteria, who confirms the shortlist, who makes the final decision after the interviews, and who authorises the offer. The position we hold is that the evaluation criteria are fixed at the second gate, before seeing a single CV. Changing them midway because the candidate everyone likes does not meet them is a common way of making a selection indefensible, even in good faith. If the criteria change, note when and why.

Communications with the candidate cut across all the stages: acknowledgement of receipt, invitation, outcome. A rejected candidate who receives no reply is a cost the process does not record, but the candidate remembers (Candidate experience and employer branding).

Examples of pipeline statuses

Statuses describe where an application stands. They are examples, not a list to copy: each company chooses as few as possible, provided each one corresponds to a decision or a wait.

• Requisition: draft, awaiting approval, approved, on hold, closed.
• Application: received, in screening, shortlisted, rejected (with reason), withdrawn.
• Interview: to be scheduled, scheduled, held, evaluation completed.
• Offer: in preparation, sent, accepted, declined, signed.
• Start: documents being collected, ready, started.

A status such as "on hold" without a name and a due date next to it is the most reliable sign of a process that loses candidates. The rules for moving an application from one status to another (the workflow) are in What an ATS is and how it works. The personal data created along the stages and the rules for processing them are in Privacy, GDPR and the AI Act in recruiting.

3. What an ATS is and how it works

An ATS (applicant tracking system) is the software that keeps positions, applications, pipeline statuses, interviews, evaluations, offers and recruiting statistics in a single repository, and moves each candidate along a flow with defined responsibilities. It is the process system. Artificial intelligence is a capability that can support it in some activities: an ATS without AI is still an ATS, whereas an AI with no process underneath leaves no trace of anything.

The definition

According to the English-language Wikipedia entry "Applicant tracking system" (CC BY-SA 4.0, accessed on 29 October 2026), an ATS is an application that handles recruiting and hiring processes electronically, and the entry compares it to a CRM designed for recruitment. Among its functions it lists posting job ads, reading CVs, searching the database, tracking, scheduling interviews, automated communications and reporting.

ATS, HRIS/HRMS and CRM for recruiting

The table is an EcosAgile working model, not a market classification: vendors use the acronyms in overlapping ways. On HRIS and HRMS, see the guide on the digitisation of HR processes.

Our position: ATS and HRIS should be kept distinct in the data model even if the same vendor sells both, because they have different lifecycles and retention rules; the unsuccessful candidate does not enter the employee records. The fragile point is the handover from one to the other: the new hire's data is either transferred or re-keyed by hand. With few recurring roles, a well-kept candidate archive, with consents and expiry dates tracked, may be enough as a talent pool, without a separate CRM.

Pipeline and workflow

The pipeline is the sequence of statuses of an application (examples in section 2). The workflow is the set of rules that governs it: who moves a candidate, which communication is sent, what must exist before proceeding. An example: without the evaluations completed by whoever conducted the interviews, the offer cannot be opened.

When you need an ATS

The number of hires per year is a poor indicator: a few hires for critical roles or across several sites can justify an ATS before many simple hires would. What counts are the signals.

• Nobody can say quickly who has already spoken to a candidate.
• The line manager evaluates using different criteria in every search.
• The channel through which the hired people arrived is not known.
• The new hire's data is copied by hand into the employee records.
• Answering a candidate's data access request takes hours of searching.

If none of these signs is present, a shared spreadsheet with written rules and an owner may be enough. To evaluate products, see How to choose ATS software.

The ten minimum functions

EcosAgile working list, not a standard. AI is not among the ten; for the rules see AI in recruiting and Privacy, GDPR and the AI Act in recruiting.

• Hiring request (requisition) with approval statuses and history.
• Position profile with requirements and criteria saved with the search.
• Publication on the website and external channels, with the source recorded on every application.
• Application form with a privacy notice.
• Single candidate archive, with search, duplicate checking and a review date for profiles.
• Pipeline with configurable statuses and a mandatory reason for rejection.
• Interviews with diary, shared evaluation form and manager access limited to their own searches.
• Communications with the candidate recorded in their file.
• Offer with approval and transfer of data to the employee records.
• Reports on sources, transitions between stages and timings, with exportable data.

4. Hiring need, job requisition and job profile: how a position is defined

Hiring need, job requisition and job profile answer three different questions. The hiring need says whether and why a person is required; the job requisition authorises filling a specific position, with location, contract and budget; the job profile describes the role and the criteria for choosing. They are approved in this order, before the job ad: whatever is not decided beforehand gets decided by looking at CVs.

Three objects to keep separate

EcosAgile working model, not a standard.

Collecting the hiring need and approving it

The hiring need arrives on a fixed cycle, usually with the budget, or on an event basis, when someone resigns: in the second case you need an approver who responds within days. For headcount and scenarios, see the guide on workforce planning and hiring needs. Only a request that says where the position sits in the organisation chart and which budget or which departing person covers it is approved.

The job profile: checklist

Checklist: EcosAgile working model, to be adapted to the company's roles.

• Mission of the role in two lines
• Main responsibilities
• Position in the organisation chart
• Location, way of working, contract
• Starting salary or band and the related criteria
• Essential requirements (must-have), verifiable
• Preferred requirements (nice-to-have)
• Skills and expected level
• Stages of the process and who evaluates what

Must-have and nice-to-have

A requirement is essential if, faced with a person who does not have it, you would not schedule the interview. If you would schedule it, it is preferred. The test removes requirements inherited from the old job ad. The criteria are written before applications are opened.

An example with made-up data: a payroll specialist for a group with two companies. Must-have: three years of payslip processing under the national collective labour agreement (CCNL) for the commerce sector; management of attendance and absences. Nice-to-have: experience in groups with several companies; reading-level English. The degree from the old job ad is gone: no activity in the role requires it.

Inclusive language and a neutral job ad

A neutral job ad has a gender-neutral title ("Payroll Specialist" or "Payroll Officer"), job-related requirements and no wording that indicates a sex, such as "dynamic young man". Article 27 of Legislative Decree 198/2006 (D.Lgs. 198/2006) prohibits gender discrimination in access to employment, including job ads and selection criteria, with limited exceptions. Article 5 of Legislative Decree 96/2026 (D.Lgs. 96/2026) requires candidates to be given the starting salary or band, with objective and neutral criteria, and notices and calls to use neutral criteria, including for the professional qualifications required: timing and channel are covered in the guide to pay transparency.

Generative AI for drafting job ads and profiles

A generative AI writes tidy drafts and makes things up with the same ease: from the title alone it adds plausible requirements that nobody chose and that become grounds for rejection. Hence: the approved job profile is the only source, the text is compared with the profile line by line, no candidate data goes into the prompts, and publication is done by a person who answers for it.

Approvals and traceability

For every search, keep the approved job profile, who approved it and when, the salary band, and every change to the criteria once the search is open. It serves to answer, months later, "why did we reject this person?".

In EcosAgile: the hiring need is optional: you can start directly from the search in progress, which is pre-filled from the hiring need. The hiring need request has a reason and approval statuses. In the profile and sought-role fields, an icon generates a text proposed by the AI, which the user reviews and supplements; requirements can use the skills catalogue of the HR suite (talent management).

5. Career page, job posting, job boards and referrals: where to find candidates and how to measure it

A job ad reaches candidates in four ways: the career page, job boards and social media, agencies and headhunters, and employee referrals. Start by looking after the career page, the only channel the company controls all the way to the application form, and for every other channel decide in advance what to measure. Without the source recorded on every application, at the end of the year you know how many CVs arrived, not where the hires come from.

"Work with us" career page

It must let people apply in a few minutes, even from a phone. It is worth testing with a phone in hand:

EcosAgile working model.

• Open searches readable without registering
• Every job ad states location, way of working and contract; for salary, see section 4
• Short form: the CV and a few details, the rest is drawn from the CV (section 6)
• Link to the privacy notice in the form (section 13)
• Spontaneous applications are kept separate from searches
• Confirmation to the candidate, with response times
• The source channel is recorded

Job posting and multiposting

Job posting is the advertisement of a search; multiposting is its publication on several channels. The advantage lies in a single approved version of the text and in applications that return to a single archive, with the source recorded. If each portal has its own hand-edited copy, after a month there are three versions of the same job ad.

Agencies and headhunters

They are useful for rare roles, volume peaks or when the company does not want to appear. The written mandate states the approved profile and criteria, timings and data exchanged. Article 5 of Legislative Decree 96/2026 prohibits asking for current or previous salary, including through whoever handles the selection: the instruction must be written into the mandate.

Employee Referral Program

A referral programme holds up if the rules are written and the same for everyone. Minimum policy:

1. who can refer and for which searches;
2. a single, tracked channel, not an email to the manager;
3. same criteria and same process as the other candidates: the referral leads to evaluation, not to a fast lane;
4. conflicts of interest: whoever refers does not evaluate; family ties and close relationships are declared;
5. reward: if the company provides one, the policy writes down its conditions and timing (no amount is suggested here);
6. outcome: it concerns the candidate, and how much to tell the referrer is decided beforehand;
7. data of the referred person: not collected from them, a case that Article 14 GDPR regulates with a notice within a reasonable period and at most one month, or at first contact; to be checked with whoever looks after privacy (section 13).

Measuring sources

For each channel, what counts is the people who arrive, those who move on to the next stage and those hired, not the number of CVs. Formulas and definitions are in the section Recruiting analytics and KPIs.

Sourcing channels

EcosAgile working model.

In EcosAgile: the Career page is published on the company website with ready-made forms (link, iframe, popup) or through an API, with language and company parameters; the form has the checkbox authorising data processing and the link to the privacy notice. For LinkedIn an XML feed is generated, which requires a LinkedIn Recruiter subscription; for other publication channels there are configurable channels and feeds, to be checked case by case.

In EcosAgile: searches can be made visible to employees, who put forward candidates; the CV source can record the referrer and the file records relationships with internal people.

6. Candidate database, CV parsing and CV search

CV parsing reads a free-form document, a PDF or a Word file, and extracts fields from it: contact details, experience, education, skills, languages. With the data in fields, candidates can be searched and compared; as long as they remain free text, they are read one by one. The result depends on the quality of the reading and on the checking done by whoever uses it: an extracted data point is a draft until a person has looked at it.

OCR, parsing and search: three different things

OCR turns an image into text, for example the scan of a paper CV; if the file already contains text it is not needed. Parsing recognises the parts of the CV in the text (this is a role, this a date, this a language) and assigns them to fields. Search then queries the fields or the text. If scans arrive, ask the vendor how they are handled.

Searching is done in three ways:

• Boolean: combines terms with AND, OR, exclusions, brackets and exact phrases. It is precise and repeatable, but finds only what is written as you wrote it: "project manager" does not find "responsabile di progetto".
• Keyword: looks for the terms in the text and ranks by occurrence. It is more tolerant and produces more noise.
• By meaning (also called semantic): compares concepts and not just words, so it finds synonyms and related roles. It costs you in predictability: you need to understand why a CV appears in the results.

It is best to start from Boolean for non-negotiable requirements, because it can be documented, and to widen by meaning when the results are few. For a natural-language search the same questions apply: who can query the archive, which fields they see, where the data is processed. And the criteria remain tied to professional aptitude: Article 8 of the Workers' Statute (Law 300/1970), referred to by Article 113 of the Italian Privacy Code (Codice privacy), prohibits enquiries into political, religious and trade-union opinions and into facts that are not relevant.

CV parsing: input, extracted data, human check

EcosAgile working model, not a standard.

Errors and data quality

Errors come from CVs laid out in columns or with graphics, from dates in different formats, from people with the same name, and from the same candidate being present twice. Before relying on it, test it on a sample of the company's real CVs, read by hand; then check from time to time how many corrections users make. The archive remains useful if the minimum fields are defined and duplicates are merged.

Talent pool and retention period

A talent pool is an archive of people who were not hired but can be contacted again. It works if it has entry criteria, an owner, a review date and exit rules. For CVs the Italian Data Protection Authority (Garante per la protezione dei dati personali) does not give an official period: the storage limitation principle applies (Article 5(1)(e) GDPR), with criteria chosen and documented by the company. It is best to set the period by purpose, search in progress and archive for re-contacting, and not a single one for everything. The privacy notice and the impact assessment are covered in the section Privacy, GDPR and the AI Act in recruiting.

In EcosAgile: the CV-Parser is an optional module: it reads uploaded CVs and fills in the fields of the candidate record, and the data read must be checked and corrected by the user before saving. An alert to HR flags CVs registered for more than a configurable number of days.

In EcosAgile: the attachment search accepts Boolean operators (AND, OR, exclusion, brackets, exact phrase, wildcard) and searches the candidate's information and the CV file. Natural-language search with AI is accessible via MCP (the protocol by which an AI assistant queries a tool), internal and external, that is, also from external AI assistants; the privacy precautions described in section 7 apply.

7. AI in recruiting: where it can support the recruiter

Artificial intelligence helps the recruiter where the work is repetitive and the result can be checked: the first draft of a profile, the summary of a CV, data extraction, a search of the archive, replies to candidates' recurring questions. It is not there to pass or reject an application in place of a person. AI supports the recruiter; the decision remains human and governed, with written criteria and a named person responsible.

AI and human oversight box: AI proposes, the person decides. Every output is a suggestion that the recruiter accepts, corrects or discards; the criterion used and the outcome remain recorded; whoever approves has the time and the authority to overturn the suggestion. The rules and their dates are in Privacy, GDPR and the AI Act in recruiting.

How much it is used, and for what

According to the HR Innovation Observatory (Osservatorio HR Innovation) of the Politecnico di Milano (2025-26 research on 155 HR departments and a panel of 1,500 workers, Ipsos Doxa; data cited by Agenda Digitale on 30 June 2026), 60% of HR departments and 48% of those who looked for work in recent months use AI in selection. For companies, CV screening is cited at 30% and tests at 10%; for candidates, CV optimisation at 21%, writing at 18%, automated applications at 8%. The article does not say on which base the individual items are calculated. Our reading: if a CV can be polished by an AI, its form says less about the person; screening must look at verifiable content.

Matrix of use case, benefit, risk, human control

The matrix is an EcosAgile working model, not a published standard. The judgements are qualitative and must be redone on the company's own processes.

Operational support and decision

Operational support is whatever prepares the work: drafts, summaries, extractions, standard replies, searches. A decision is whatever changes a person's position in the process: who advances, who drops out, who receives the offer. The first group can be delegated with spot checks, the second cannot, not even when the score looks clear-cut.

The following remain with a person: the choice of knock-out criteria, the rejection of an application, the judgement on personal traits inferred from the CV, and the verification of an anonymised document. We also advise against uploading CVs into an external AI assistant that has not been assessed for privacy: they are personal data, to be handled in accordance with the GDPR.

Going deeper with the expert: in the webinar Recruiting with AI and personnel selection (in Italian) an EcosAgile speaker shows how AI helps with hundreds or thousands of unstructured CVs: summary, data extraction, comparison with searches. Recorded in 2025, it also describes functions that were then under development: for those available today, ask EcosAgile.

In EcosAgile: a dedicated icon proposes the texts of the profile and of the sought role, which the user reviews. AI Match calculates the similarity between candidate and position. AI Insight offers a profile inferred from the CV, as an indicative estimate, without confirmation from the candidate. A CV summary with AI and an assistant for the candidate are also present. CV Blind generates a version of the CV without direct identifying references, which a person must always check before use or hand-over. According to the terms of use (Article 6.6), the outputs are suggestions, not automated decisions: the final decision remains with the customer.

8. Screening, matching and skills-based selection

Screening reduces the number of applications by applying requirements relevant to the position; matching compares each candidate's profile with those requirements and produces a score or a reading order. These are preparation steps: a system can flag, rank and propose, but who advances and who drops out is determined by a person who knows the criteria and answers for them. The score describes a CV against a grid, not the person.

Automated screening and decision

Score and decision sit in two different columns. A 72 out of 100 orders the reading, it does not say "suitable". A threshold that closes the application without a re-read is already a decision. If the score becomes the determining factor, an issue under Article 22 GDPR may arise; whether it applies to candidate ranking is an open question of interpretation (Privacy, GDPR and the AI Act in recruiting).

Knock-out criteria, preferred criteria and skills

A knock-out criterion is a requirement without which the role cannot be assigned; a preferred one weighs on the ranking but does not exclude; a skill is what the person can do, at an expected level. Mixing them up is a frequent mistake: a "preferable" with a high weight is a hidden knock-out.

The table is an EcosAgile working model, not a published standard. The examples are made-up data.

False negatives and false positives

A false negative is the suitable candidate who was rejected or ended up at the bottom; a false positive is the unsuitable one who rises to the top. The second emerges at the interview; the first does not emerge if nobody looks. In practice, for each search a person rereads a random sample of below-threshold CVs (twenty, for example), and the cases in which the recruiter overturns the proposed order are recorded, with the reason. If the overturned cases are concentrated on one criterion, that criterion or its weight is reviewed. The method makes errors visible, it does not eliminate them.

Explainable shortlist

For every name on the shortlist, the criteria met, the point in the CV that documents them and the open doubts are stated. The test: the recruiter explains to the line manager why one candidate is in and another out, using only the criteria written before reading the CVs. If you have to say "the system put them at the top", it is not explainable.

Skills-based selection

By skills-based selection we mean giving more weight to what a person can do for the role than to qualifications and career paths. Here we do not present a shared definition or efficacy data: we treat it as a working method. You define the skills of the role, the expected level and the evidence that is enough (described experience, a test, a structured interview). A list of keywords is not evidence. Skills in the rest of the company are the subject of Talent management.

Human oversight in practice

The recruiter proposes, the line manager confirms whom to meet. For each search, the criteria and weights (with date), the score, the outcome, who assigned it and the reason for rejection, from a list of reasons, are recorded. Without these lines, eight months from now nobody can explain an outcome. Bias checks are in Bias, fairness and human oversight, indicators in Recruiting analytics and KPIs.

In EcosAgile: AI Match shows weight bars for education, roles, skills, sectors, languages and managerial level, which the user can modify; results are sorted by score. The outputs are suggestions: the final decision remains with the customer.

9. Bias, fairness and human oversight

Entrusting screening to an algorithm does not eliminate prejudice: it moves it into the data, the criteria and the rules that someone has chosen. A system can make the application of criteria more consistent, but it can reproduce the preferences in historical data, and a criterion that looks neutral can act as a proxy for a characteristic that could not be used. You need relevant, written criteria, traceability, periodic checks, and a person who answers for the outcome.

Human bias and algorithmic bias

Human bias takes familiar forms: the fatigue of someone reading CVs at the end of the day, the preference for people who resemble the selector. Algorithmic bias arises elsewhere: skewed historical data, criteria that replicate the past, variables that indirectly stand in for sensitive data. You sometimes read that AI evaluates skills alone and therefore removes prejudice. That is too strong a claim: AI applies criteria and data chosen by people and inherits their limits.

Sensitive variables and proxies

Gender, age, origin, religion, trade-union opinions, health: as a rule these are not relevant selection criteria, and exceptions are checked with whoever looks after the legal side. It is not enough to remove them from the form: the variables that let them be guessed remain, such as photo, date of birth, address, career breaks. In terms of the rules, Article 8 of the Workers' Statute prohibits the employer from enquiring, at the hiring stage, into political, religious and trade-union opinions and into facts not relevant to professional aptitude. Article 27 of Legislative Decree 198/2006 prohibits gender discrimination in access to employment, including in selection criteria, job ads and pre-selection.

Transparency, criteria and traceability

Every outcome must be reconstructable: which criteria, in which version, who decided, for what reason. The candidate must be informed of the use of automated systems in the cases and within the times required by the rules, which are set out in Privacy, GDPR and the AI Act in recruiting. In brief on the AI Act: selection systems may be high-risk, with obligations, including human oversight, from 2 December 2027; the conditions are in section 13.

A simple check on conversion by stage

The method: for each stage (application, screening, interview, offer) calculate the share of candidates who advance and compare it between relevant groups. A marked difference does not prove discrimination; it indicates where to reread criteria and reasons. Which groups to observe, and with which data, is decided beforehand with whoever looks after privacy: collecting certain information is already processing that has to be justified. The check does not find every distortion; it makes some of them visible.

An example with made-up data: 12 of 60 applications from group A (20%) and 3 of 40 from group B (7.5%) go through to interview. The sample is small and proves nothing, but it is a reason to reread the criteria of that stage.

The following table is an EcosAgile working model, not a published standard. The examples are made-up data.

In EcosAgile: CV Blind uses AI to generate a version of the CV without direct identifying references: the name is replaced by a candidate code and contact details are masked. A person must always check it before use or hand-over, looking for direct or indirect identifying data, irrelevant information and incorrect content. It reduces identifying references; it does not make the evaluation neutral.

10. Interviews, tests and structured evaluation

A structured interview puts the same core questions to all candidates for the same role, scores the answers on a scale written before the meeting, and notes the evidence for each judgement. It serves to make comparable people seen by different interviewers in different weeks. Software helps with the diary, involving managers and collecting evaluations; the criteria remain work for HR and managers, to be done before the first interview.

Scheduling and collaboration between HR and managers

The most concrete constraint is the line manager's calendar. Agree with them, once the search is open, how many hours a week they will devote to interviews, and a matrix assigns to each interviewer the skills they explore: nobody repeats a colleague's questions, and the candidate does not tell the same experience three times. HR looks after the method; the manager brings the technical content. Two agreements are enough: a deadline for the evaluation (the day after the interview) and a quarter of an hour of calibration before the first meeting to read the scales together.

Structured interview and scorecard

The scorecard starts from the job profile (section 4): a few skills that distinguish whoever does well in that role and, for each, a question about past behaviour, a scale with descriptors and a space for the evidence. An existing skills catalogue, for example in talent management, is the starting point. Each interviewer fills in their own form before talking to the others, and the final score comes out of the discussion of the evidence, not from the average of the numbers.

The scorecard is an EcosAgile working model, not a published standard.

An example with made-up data: purchasing manager, scale from 1 to 4, one interviewer's evidence.

Tests and questionnaires

A test makes sense when it measures what the role requires: a spreadsheet exercise for a controller, a writing exercise for an editor. Personality questionnaires call for more caution: in its decision no. 302 of 21 July 2011, the Garante found questionnaires on sex life, health, pregnancy and criminal records to be irrelevant, and the processing unlawful. The Italian Privacy Code (Codice privacy) preserves, in Article 113, the prohibition in the Workers' Statute on enquiring into facts not relevant to professional aptitude. Every question must be tied to a skill of the role, and the questionnaire should be reviewed by whoever looks after privacy and legal matters before it is adopted.

Feedback to the candidate: descriptive or a score

A score serves internal comparison and tells the candidate little. Descriptive feedback says what was observed against a skill ("explained how they renegotiate, did not bring a case with dates") and holds up even if the candidate challenges it. A number with no evidence passes from one stage to the next without anyone remembering where it comes from. It is best to write the notes as if the candidate could read them: on candidates' data access requests there is already a decision by the Garante (no. 737 of 4 December 2025). How to communicate the outcome is the subject of section 11.

Unvalidated signals

In a video interview the temptation is to give weight to voice, gaze, expressions. A tone of voice is not a skill of the role: a score built on those signals must be shown to be valid for the job before it enters a decision, and until it is, the judgement concerns the content of the answers. The AI Act has prohibited, since 2 February 2025, the inference of emotions in the workplace, except for medical or safety purposes (Article 5(1)(f)); whether this also applies to selection is not clarified, as section 13 explains (Privacy, GDPR and the AI Act).

In EcosAgile: interviews are organised by interview type, with diary and calendar; the manager sees their own interviews in the Manager Self Service. For the types that provide for it, tests are assigned, with score and result. The interview evaluation is structured by areas and by characteristics, and managers enter their evaluations in the Manager Self Service. Video interviewing with templates is available.

11. Candidate experience and employer branding

Candidate experience is what a person goes through from reading the first job ad to the outcome, positive or not: clarity of the job ad, effort to apply, communications, timing, interviews, final reply. It is improved by removing friction and silence, and checked with a few numbers: where people drop out, how long they wait, who replies. Employer branding is the reputation that results; technology helps if it reduces waiting without making the person disappear from the process.

Candidate journey

The journey is designed from the candidate's side, one touchpoint at a time: I know what they are asking of me, I know whether they have received it, I know what happens now.

The table is an EcosAgile working model, not a published standard. It needs to be adapted to the company's real touchpoints.

Timing and communications

Each stage has an internal maximum time and the candidate knows it: the first email says what happens and when (example with made-up data: outcome within ten working days). If the time slips, you write anyway. The candidate's status is updated as soon as it changes and messages sit in a single view, so that two people do not reply twice, or nobody does.

Career page and brand

The career page is the point where the promise (culture, growth, flexibility) meets the real process: if it talks about caring for people and the process does not deliver, the difference shows. Better to say less and stay consistent. The length of the form is also a brand choice: according to the speaker in the EcosAgile webinar (in Italian) (recorded in 2025), fifty fields lose above all those who are not urgently looking for work.

Assistants and self-service

An assistant for candidates handles repetitive questions (status, documents, timing) and leaves the others to a person. As good practice, it states that it is an AI system and always gives a way to reach a recruiter; the outcome stays with a person. Article 50 of the AI Act, applicable from 2 August 2026, requires providers of these systems to design them so that people know they are interacting with an AI, unless this is obvious from the context. The rest is in Privacy, GDPR and the AI Act.

Feedback and rejection

For every hire there is a number of people who receive a no, or no reply, and it is the moment that weighs most on reputation. The rejection goes out after a person has made the decision, says at what point the application stopped and, if there is a reason that can be communicated, gives it by category. If the profile is of interest for future searches, the candidate is told; how long to keep it is decided and documented (Privacy, GDPR and the AI Act). On interview feedback, section 10 applies.

Application drop-off

Application drop-off measures how many applications that were started are not completed, as an indicator of the quality of the application experience (paraphrased from Wikipedia, entry "Recruiting metrics", CC BY-SA 4.0, accessed on 29 October 2026). The same ratio can be read as abandonment or as completion: state which. It is not the same as candidates leaving between one stage and the next. Formulas in Recruiting analytics and KPIs.

Candidates already use AI

According to the HR Innovation Observatory (Osservatorio HR Innovation) of the Politecnico di Milano (2025-26 research, 155 HR departments and a panel of 1,500 workers, Ipsos Doxa; cited by Agenda Digitale, 30 June 2026), AI in selection is used by 48% of those who looked for work in recent months; by type of use, CV optimisation 21%, writing 18%, automated applications 8%. These are sample data: the article does not give the base of the individual items, and they are not a measure of effectiveness. Our reading: if many applications are born with an assistant, a well-made CV distinguishes less and specific questions and a structured interview weigh more (section 10).

In EcosAgile: the Communication Hub is the single view, in the Recruiting module, of communications to and from the candidate, linked to the stages of the process, on email and WhatsApp channels. It has personal views, filters (unread, with attachments, archive), threads and statuses such as "To reply" and "To do". The standard automatic emails cover confirmation of receipt, rejection and interview reminder. An assistant is also available for the candidate.

12. Offer, electronic signature and onboarding

The offer fixes the terms before acceptance, the electronic signature makes the agreement documentable, and onboarding prepares the person's arrival. The point to watch is the handover: the data collected during selection must reach personnel management without being re-keyed, with an owner for every step and a record of who approved what.

Placement proposal and letter of intent

The proposal gathers the terms of the relationship: role, company and location, type of employment, start date, probation period, pay, benefits and equipment. The letter of intent is the document addressed to the candidate by which the company expresses its intention to hire on those terms; what legal effects it has depends on the text and the type of employment, and the legal office or the labour consultant assesses this. The proposal goes through internal approval before reaching the candidate and is linked to the search and, if there is one, to the hiring need it arises from (section 4), so that the post and the terms remain traceable.

Electronic signature

The offer and the letter are signed electronically when the candidate is far away or the start date is near. The type of signature (simple, advanced or qualified) depends on the document and on the risk the company wants to cover. The signing process is described in Document & Signature, and for what changes with eIDAS 2 there is the news item on eIDAS 2 and digital signature (in Italian). In recruiting, one safeguard is enough: the signed document enters the candidate's file and from there the employee's, without going through email.

From candidate to employee

What is needed for the employment relationship passes through, not the selection file. These must arrive without rewriting:

EcosAgile working list, to be adapted to the data the company requires.

• personal details and contact details, verified against documents;
• role, company, location and type of employment;
• start date and probation period;
• agreed pay, benefits and equipment;
• signed proposal and letter;
• link to the search, the hiring need and any person being replaced.

Interview notes, test results and scores remain in the selection file, with the retention periods the company has chosen and documented (Article 5(1)(e) GDPR); passing only the necessary data to the employee records is minimisation (Article 5(1)(c)). Both choices are agreed with whoever looks after privacy. Payroll data goes to the system or the consultant who processes it: the handover is designed together and not redone by hand. The data framework is in Privacy, GDPR and the AI Act.

Onboarding with tasks, documents and notifications

Onboarding starts before the first day: documents to collect, access and equipment to prepare, people to notify. It works when every activity has an owner and a deadline tied to the start date (example with made-up data: five days before for the laptop, the first day for the badge) and when whoever has to act receives a notification instead of finding out mid-morning. Templates by role or by site avoid rebuilding the list for every hire. The overview with the other HR processes is in the guide to the digitisation of HR processes.

In EcosAgile: in the Recruiting module you fill in the placement proposal and attach the letter of intent for the candidate. Electronic signature of the offer is available. In the proposal you select "Create employee record from candidate", with the CV data proposed as initial values, and a flag notifies Personnel Administration when the offer status is "Approved". The onboarding part for new hires is also available, which we do not describe in detail here.

13. Privacy, GDPR and the AI Act in recruiting

Recruiting involves processing CVs, interview results and evaluations of people not yet in the company. The GDPR requires clear purposes, a privacy notice, relevant data and retention based on a criterion. The AI Act adds, for systems that filter and evaluate applications, obligations from 2 December 2027; prohibitions and general obligations already apply. For CVs the Garante does not appear to have set a retention period: whoever processes the data chooses it and documents it.

This section describes the rules and is not legal advice: they must be checked against the text in force before being applied.

What the rules say, and from when

Sources consulted on 29 October 2026. Texts prior to July 2026 still give 2 August 2026 for high risk: that date has been superseded. Regulation (EU) 2026/1744 (OJEU 24 July 2026, in force from 27 July 2026) has moved the obligations of Annex III, selection included, to 2 December 2027.

The sources of the rules and decisions are listed in the section Sources. The step from the rule to recruiting is an interpretation, declared as such in the following paragraphs.

Candidate data and minimisation

In the application form every mandatory field must serve to assess aptitude for the role, as must every questionnaire or test. The Garante found personality questions on irrelevant aspects to be unlawful (no. 302 of 21 July 2011); the Workers' Statute prohibits enquiries into opinions and facts that are not relevant.

Privacy notice to candidates

For job ads and forms Article 13 applies, for CVs that come from third parties Article 14, for the unsolicited CV Article 111-bis. Interpretation: Article 111-bis does not mention job ads and active searches, where the cautious reading of Articles 13 and 14 applies.

Retention of CVs

It does not appear that the Garante has indicated an official period: a number read in a guide or in a contract is the choice of whoever wrote it, not an indication from the Garante. The principle of Article 5(1)(e) remains. Implementation: write down the criterion (for example the closing of the search), put it in the privacy notice, and assign a person to the periodic check of the archive.

Automated decisions and profiling

In the SCHUFA case, which concerned credit, the Court treated as a decision under Article 22 even the score on which a third party relies in a determining way. Whether this applies to a score that ranks candidates, when formally a person makes the choice, is an open question of interpretation that this section does not settle: whoever uses a ranking as a de facto filter should put the case to a lawyer.

High-risk AI in recruiting

What matters is what the system does: targeted job ads, filtering and evaluation of applications fall under Annex III, point 4(a). The derogation under Article 6(3) requires the absence of significant risk and at least one condition (narrow procedural task, improvement of a human activity, detection of patterns without influencing the assessment, preparatory task); with profiling it does not apply, and the provider that applies the derogation documents the assessment before placing on the market and registers. Whether a system that ranks candidates deserves it is assessed case by case. Interpretation: whoever adopts an ATS is normally a deployer, a qualification to be verified. For oversight in practice, section 9.

Information for candidates and workers

Law 132/2025 speaks of the "worker": whether it applies to someone who is still a candidate is a question of interpretation; the scope and sanctions of Article 1-bis are not covered here. Cautious implementation: say in the privacy notice whether and at which stage AI intervenes. On pay transparency (Legislative Decree 96/2026) there is the guide to pay transparency.

Governance and documentation

If AI evaluates candidates, the question of the DPIA (Article 35(3)(a)) must be asked and the answer written down. Decision no. 737 of 4 December 2025 found a breach of Articles 12(3) and 15 GDPR for the late reply to a candidate's access request.

The governance table is an EcosAgile working model, not a published standard.

In EcosAgile: the application form for the company website includes a checkbox authorising data processing (EU Reg. 2016/679) and the link to the privacy notice, the text of which belongs to the customer. What to declare and which legal basis to indicate is decided by the controller. Access profiles separate HR, managers and employees. An alert flags to HR the CVs registered for more than a configurable number of days; what to do with them is for the company to decide. AI outputs are suggestions; under the terms of use (Article 6.6) final decision and human oversight remain with the customer.

14. Recruiting analytics and KPIs: what to measure to see the funnel

Recruiting KPIs serve to see where the process slows down, loses candidates or produces hires that do not last; the number of CVs received answers none of these questions. A useful set measures the time to fill the position, the time from application to acceptance, conversion between stages, the effectiveness of sources, application abandonment, offer acceptance and the quality of hire. Each must be defined in writing, with a start point, an end point and a unit of measurement.

Time to fill and time to hire are not synonyms

Time to fill runs from approval of the requisition to acceptance of the offer: it includes the time before the job ad is online and the time needed to receive the first applications. Time to hire runs from the application of the person who is then hired to the same acceptance: it is the process time as seen by the candidate and is calculated only on those hired. Industry sources do not agree on start and end; for example, some close time to fill at the pre-employment checks (paraphrased from Wikipedia, entry "Recruiting metrics", CC BY-SA 4.0, accessed on 29 October 2026). No formal standard appears to exist for these two indicators. Choose a definition and write it down.

Industry sources diverge on the start point. With the one chosen here, the difference between the two values is the time that passes before the application of the person who is then hired arrives: with a high time to fill and a low time to hire, the search stands still before collecting useful applications (approvals, publication, channels); if instead it is time to hire that stretches, the delay lies in the interviews or in the decision.

An example with made-up data, in calendar days: requisition approved on 3 March, application of the person hired on 17 March, offer accepted on 14 April. Time to fill 42 days, time to hire 28. The 14 days of difference are the two weeks before the application arrived: that is the first thing to look at, before the interviews.

The KPI table

The table is an EcosAgile working model, not a published standard: frequencies and owners are proposals to be adapted to the organisation.

For cost per hire there is a standard: ANSI approved in June 2012 a standard sponsored by SHRM, for the US market. According to a third-party source the formula is external costs plus internal costs, divided by the hires in the period; the official text of the standard is not reproduced here. For quality of hire, SHRM experts indicate components such as performance, retention of new hires, time to full productivity, the manager's judgement and fit with the role; the company chooses which to use and states them.

An example with made-up data: 300 applications started and 200 completed, abandonment 33%; the following steps are in the table. Candidates put forward per hire: 20 for 3, that is 6.7.

Segment, avoid vanity metrics, define in writing

A company-wide average hides more than it shows: the time to fill of a warehouse operative and that of a department head do not fit in a single number. Segment at least by role and by channel, and do not read as a trend a value calculated on three hires.

Vanity metrics are CVs received, job ad views, the number of job ads published: they become useful only when you relate them to a later stage. For abandonment in applications see Candidate experience, for channels Career page and job posting, for transition rates as a check on the outcomes of an AI system Bias and human oversight.

Each KPI has a card: name, formula, start, end, unit, calendar or working days, what is included, data source, owner, frequency. If the definition changes, the history is recalculated, otherwise the comparison between periods is not valid. Here you will not find market values for times and costs: the numbers in circulation are second-hand, often American and from different years. The comparison you need is with your own previous periods, by role and channel.

In EcosAgile: the Recruiting module has a Dashboard panel (Received CV Analysis, Search Analysis, Interview Analysis) and a Report panel (Search Analysis, Application Analysis, Internal Application Analysis, Characteristic Evaluation Analysis, CV Parsing). Which KPIs to derive from them and with which definitions remains a choice for the company.

15. How to choose ATS software

An ATS is chosen starting from your own process: how many searches are open at once, how many companies and sites, who works in the system every day and who opens it once a month. You set essential and preferable criteria, ask written questions on AI and data exit, try the product on a real case and calculate the cost over several years, not the licence alone.

The author is an HR software vendor: read the criteria with that in mind.

Volumes, complexity, essential and preferable

Count parallel searches, channels, companies and sites, and above all the managers who log in to the system. Then fix the list before the demos. Essential is whatever the process cannot do without or whose absence would breach a rule (traceability of stages, privacy notice, roles and permissions, export); preferable is whatever can be worked around. The minimum functions are in What an ATS is.

AI and governance: questions for the vendor

Four questions in writing; rules and dates are in Privacy, GDPR and the AI Act.

• What output does the AI produce (text, score, ranking) and who sees it?
• Can it exclude a candidate without human intervention? Who decides and where is it recorded?
• Which logs does it keep, for how long, and can they be exported?
• What information does it provide for the impact assessment and the privacy notice, and how are the outcomes checked (Bias)?

Checklist of 20 criteria

The checklist and matrix are an EcosAgile working model, not a published standard.

• Hiring need with request and approval statuses
• Career page, channels and referrals in the same system
• Short application, with privacy notice, from a phone
• CV parsing with data verification by the user
• CV search, text-based and by skills
• Configurable pipeline and statuses by type of search
• Interviews with diary and structured evaluation
• Communications with candidates tracked
• Offer, electronic signature, handover to the employee records
• Reports on funnel, sources and timings with documented definitions
• AI declared: suggested outputs, human decision
• AI: logs, testing and monitoring documented
• Privacy: configurable retention, data subjects' requests
• Roles and permissions for recruiters, managers and companies
• Security: data location, authentication, tracked access
• Documented APIs, integration with employee records and payroll
• Multiple companies and multiple languages
• Total cost over several years, services included
• Implementation, training, change management
• Exit: full export of candidates, history, consents

SME or multi-company: what weighs most

Demo script

EcosAgile working list, to be adapted to the company's process.

1. Bring a real search and anonymised CVs, not sample data.
2. Take a hiring need request all the way through to the career page.
3. Apply as a candidate from a phone and count the steps.
4. Search the CVs and correct a data point read by the parsing.
5. Have a manager with a limited profile enter an evaluation.
6. Have them show you what the AI did, who decided and where a record remains.
7. Export a closed search and generate the timing report (KPIs).

Exit, red flags, stakeholders

Ask in writing for the format, timing and conditions for exporting candidates, attachments, stage history and evaluations, and what happens to the data at the end of the contract.

Red flags: absolute promises on AI and bias; no written answer on logs and data exit; demos only on sample data; no API documentation. Involve HR and recruiters, IT, privacy and legal, hiring managers, finance and everyday users from the start.

In EcosAgile: Recruiting manages multiple companies, with the company code in the links and APIs of the Career page, which is delivered in the language indicated by the parameter among those that are active. For the company website there are ready-made forms and REST APIs. It uses data from the HR suite (organisation, skills, employee records). See the Recruiting software page or book the demo.

16. Frequently asked questions