Turnover at the factory: what pplPortal detects before departure

In 2023, the average voluntary turnover in Portugal was 10.6% — and 52% of companies admitted difficulty in retaining talent, according to Mercer. In a garment factory with 120 operators, this translates into 12 to 13 departures per year. What no HR director can answer with precision is: how many of them were foreseeable three months in advance? The uncomfortable answer: most of them. The signs were there. No one cross-referenced them in time.

This article defends a thesis that contradicts what most companies do: the problem of industrial turnover is not a lack of data — it is an excess of uncross-referenced data. Attendance is at the control point. Output is at the production terminal. Shift swap requests are in the line supervisor's notebook. Separately, they are noise. Cross-referenced in a model configured for the factory context, they become an alert six to ten weeks in advance. That is enough time to act. The article details how pplPortal performs this cross-referencing — and, more importantly, how to configure the response process so that the alert produces action and not merely a report that no one reads.

Why factory turnover has a different signature

The problem with generic retention models

Most predictive turnover models were developed for office workers with computer access, formal performance reviews and structured feedback cycles. In a textile factory in the Vale do Ave or on a footwear assembly line in Felgueiras, that data simply does not exist in the same way. The operator does not fill in satisfaction surveys. They have no recorded monthly one-on-ones. They do not use the HR portal to request leave — they do so on paper or through the line supervisor.

What does exist, and what is capturable, is operational behaviour: attendance patterns, shift variations, records of refused overtime, requests to change workstation, minor disciplinary occurrences, and — frequently ignored — the temporal sequence of these events. An employee who over six months goes from zero absences to two a month, refuses overtime for the third time and requests a transfer to another section is signalling departure months in advance. The line supervisor knows this instinctively. The problem is that this knowledge enters no system.

Industrial turnover does not announce itself — it accumulates in micro-signals that the line supervisor sees but does not record, and that the HR director never gets to see.

The real intervention window

In Portuguese manufacturing, the window between the first detectable behavioural signal and the actual termination is, typically, six to twelve weeks. That is enough time for a conversation, a reassignment proposal or a salary review — if the signal is captured in time. Without a system, that period passes blank. With a well-configured retention model, the HR team can prioritise the highest-risk cases and act before the decision is made.

There is an operational detail that the manuals do not mention: in factories with two shifts, departure signals tend to appear first in the afternoon shift. Workers on the morning shift have more direct contact with management and more internal visibility — which creates a social brake on the deterioration of behaviour. On the afternoon shift, the line supervisor has less presence and the micro-signals accumulate without anyone noticing. If the predictive model does not segment by shift, it will detect morning-shift cases earlier and systematically under-represent the risk on the afternoon shift. This asymmetry is real, it is common, and it is rarely configured by default.

The signals pplPortal cross-references

Attendance: it is not the number of absences, it is the pattern

An isolated absence says nothing. Three absences on Mondays over an eight-week period, combined with two early departures and a record of lateness exceeding 15 minutes, form a pattern. The pplCore module of pplPortal records each attendance event with a timestamp and declared reason. The analysis engine aggregates these events by employee and calculates deviations against the individual historical average — not against the factory average, which would be statistically useless for operators with very distinct attendance profiles.

The article Attendance and shifts in the factory: what pplPortal automates straight away details the technical configuration of this module. What matters here is the cross-referencing: degraded attendance in isolation is noise; degraded attendance plus another signal is an alert.

Overtime refusal and shift swaps

This is the most underestimated signal in Portuguese factories. A motivated operator accepts overtime regularly — especially during end-of-season production peaks in the clothing sector or in footwear campaigns ahead of international trade fairs. When that pattern reverses, it is because something has changed. pplAdvanced records shift swap requests, overtime refusals and workstation change requests. On their own, they are noise. Cross-referenced with attendance and with the evaluation history, they become a signal.

Variations in production records

In a factory with shop-floor terminals connected to KORA Productivity, each operator records output by production order. A sustained drop in individual output — not explained by a change of article or machine — is a behavioural indicator, not merely an operational one. Degraded individual OEE, when correlated with HR data, has predictive value that the production manager rarely uses for retention. pplPortal can receive that feed and integrate it into the risk profile.

Beware of a frequent configuration error: a drop in output is only a risk signal if it is controlled for a change of article. An operator who moves from shirts to coats will have lower output in the first few weeks — that is learning, not disengagement. If the model does not filter by type of article in production, it will generate false positives in bulk whenever there is a collection change. In clothing factories in the North, this happens twice a year across the entire line.

Training and progression requests

The inverse signal also exists. An employee who for two years requested training and received no response — and who has since stopped requesting it — is in silent disengagement mode. pplTalent records development requests, competency assessments and career plans. The absence of activity in a historically active profile is as valid an alert as an absence.

Silent disengagement is not the absence of problems — it is the phase in which the employee has already decided to leave but has not yet said so.

Communications via the whistleblowing channel

Law 93/2021 requires companies with 50 or more employees to maintain a whistleblowing channel. That channel, when integrated with the HR system, generates data on organisational climate that no satisfaction survey captures with the same honesty. pplPortal manages this channel in compliance with the GDPR and with Law 58/2019 (national implementation of the GDPR), guaranteeing anonymity to the whistleblower and traceability to the process — without exposing personal data to those not authorised.

How the predictive model is built

Inputs, weights and alert threshold

The pplPortal predictive AI engine — integrated into the pplEvolution module — does not use a single model for all factories. The initial configuration defines which signals are available (depending on the active modules and the integrations with production systems), the relative weights of each signal, and the threshold from which an employee enters a departure-risk alert.

What the model does NOT do — and why that matters

The model does not produce a departure probability as an exact percentage. It produces a risk classification (low / medium / high / critical) and a list of signals that contributed to that classification. This distinction is deliberate: a figure such as "78% probability of departure" generates false precision and can lead to disproportionate interventions. A classification with identified signals allows the HR manager or the line supervisor to assess the context before acting.

This approach is also aligned with the AI Act (EU Regulation 2024/1689), which classifies AI systems that influence decisions about workers as high-risk, requiring transparency, human oversight and model documentation. pplPortal keeps a log of each alert generated, with the signals that make it up, auditable by the company's DPO.

Update frequency and latency

The model recalculates each employee's risk profile at a configurable frequency — typically weekly for factories with a high historical turnover, fortnightly for more stable operations. The latency between the recording of an event (for example, an overtime refusal) and its incorporation into the risk profile is less than 24 hours in standard configurations. For factories with real-time integration via KORA Productivity, production data is ingested in near-real-time.

Technical trade-offs by operation size

What works in practice

Pattern 1: the alert as a trigger for a conversation, not for a process

The temptation is to create an automatic workflow: alert generated → email to HR → meeting scheduled → action recorded. In Portuguese factories, this bureaucratic process fails because the line supervisor — who is the one having the real conversation — does not use the HR portal regularly. What works is a simple alert, sent by SMS or notification on the pplPortal mobile application, to the direct line supervisor, with two or three signals in operational language: "João Silva: 3 absences in the last 6 weeks, refused overtime twice, output below usual." The line supervisor decides whether to have the conversation or to escalate it to HR. The system records the decision.

In a garment factory with around 100 employees, this model reduces the response time to risk signals from weeks to days — not because the system is faster, but because the alert reaches the right person, in language they understand, without forcing them into a portal they do not use day to day.

Pattern 2: segmentation by critical role, not by seniority

Most companies treat all turnover alerts with the same priority. A sewing operator with two years at the company is replaceable within six weeks — at a cost, but replaceable. An embroidery machine setter with twelve years of experience on a specific article may take six months to replace, if it is possible at all. Configure pplPortal to classify employees by the criticality of their role before activating the predictive model. High-risk alerts in critical roles have absolute priority. The rest enter a weekly triage queue.

This cross-referencing between departure risk and role criticality is what transforms pplPortal from an HR tool into an operational continuity tool.

Retaining everyone with the same urgency is retaining no one effectively. Role criticality has to enter the model before the risk score.

Pattern 3: using the model to calibrate the recruitment process

pplPortal, through the pplTalent module, records the entry profile of each employee: recruitment source, time to first absence, time to first disciplinary occurrence, length of stay. By cross-referencing this data with the departure alerts generated, it is possible to identify which entry profiles have the highest 18-month retention rate. This analysis, detailed in the article Digital recruitment and succession: what pplPortal records and decides, closes the loop: the departure predictive model feeds the entry selection model.

How to implement: operational sequence

1. Audit the existing data. Before activating any predictive model, check which attendance, disciplinary and production events are actually being recorded in the system — and with what quality. Incomplete or inconsistent data produces useless alerts. This audit should cover the last 24 months and identify gaps by section or shift.
2. Define the role criticality taxonomy. Classify all roles into three levels: critical (replacement >3 months), intermediate (1-3 months), operational (<1 month). This classification is manual and should involve the production director, not just HR.
3. Configure the signals and thresholds in pplPortal. Start with three to four signals — attendance, overtime, production output (if available) — and conservative thresholds. Fine-tune over the first 90 days based on the false positives identified.
4. Define the response flow by risk level. Critical risk: alert to the line supervisor and to HR within 24 hours. High risk: alert to the line supervisor within 48 hours. Medium risk: weekly review in an HR team meeting. Low risk: passive monitoring. Document this flow — it is required by the AI Act for AI systems with an impact on decisions about workers.
5. Integrate with KORA Productivity or with the production ERP. Without individual output data, the model operates with half the available signals. This integration is the step that most increases predictive precision and that is most frequently postponed — do not postpone it.
6. Review the model quarterly. Compare the alerts generated with the actual departures. Calculate the detection rate (how many departures were preceded by an alert) and the false positive rate (how many alerts did not result in a departure). Adjust weights and thresholds based on this review.

Success metrics after implementation

What to measure and how often

The question the controller will ask three months after implementation is: "was it worth it?" To answer with data, define the metrics before you start.

• Early detection rate: percentage of voluntary departures preceded by an alert more than 30 days in advance. A reasonable target for the first year: 50-60%.
• Alert-to-action conversion rate: percentage of alerts that resulted in a documented intervention (conversation, proposal, reassignment). Below 40% means the response process is failing, not the model.
• Average time between alert and intervention: should be less than five working days for high- or critical-risk alerts.
• Variation in the voluntary turnover rate: the final indicator. Do not expect results before 12 months — the model needs time to accumulate history and the interventions take time to produce an effect.
• Cost avoided per prevented departure: estimate the replacement cost per role (recruitment, training, loss of productivity) and multiply by the number of departures prevented. This figure is the argument for the CFO.

What Qlik Sense adds

The data generated by pplPortal gains another dimension when visualised in Qlik Sense dashboards. The operations director can cross-reference the turnover risk map with the production plan — identifying weeks in which the concentration of high-risk employees coincides with load peaks. This analysis, detailed in the article People analytics in the factory: HR data the operations director uses, transforms talent retention into an operational planning problem, not merely an HR one.

Compliance and data protection

GDPR and the processing of behavioural data

Cross-referencing attendance, production and behaviour data for predictive purposes constitutes processing of personal data under the GDPR and Law 58/2019. The most common legal basis is the employer's legitimate interest — but it requires a documented proportionality analysis. pplPortal maintains a record of processing activities (Article 30 of the GDPR), allows retention periods to be configured by data category, and supports employees' right of access and portability.

The associated payroll processing — including the Monthly Remuneration Declaration (DMR) to the Tax Authority and Social Security — is integrated into pplCore and documented in the article Payroll and salary closing in the factory: what to automate in pplPortal.

AI Act: what changes in 2025 and 2026

The AI Act classifies systems that evaluate employees based on behaviour as high-risk AI systems (Annex III, category 4). The obligations include: technical documentation of the model, a conformity assessment before being placed in service, mandatory human oversight of automated decisions, and registration in the EU database for high-risk systems. The entry into force of the obligations for high-risk systems is phased — check the current timetable with your DPO. pplPortal was designed to support this documentation, but the responsibility for compliance lies with the end user (the employer), not with the software vendor.

The mistake that 35 years of Portuguese factories teach

There is a pattern that repeats itself. The company implements the predictive module, configures the alerts, and in the first two months receives a list of 15 high-risk employees. The HR director convenes meetings. The line supervisor has conversations. Three leave anyway. Twelve stay. Internal conclusion: "the system works."

Six months later, the alert response process has been abandoned because "it is too much work" and "most don't leave anyway." The model keeps generating alerts. No one reads them. The turnover rate returns to its previous level.

The problem is not the software. It is that the company treated the implementation as an IT project and not as a process change. pplPortal detects the signals. The decision to act — and the discipline to maintain the response process — is human. Without that discipline, the best factory employee retention software is just a dashboard that no one consults.

Absenteeism forecasting follows the same logic — and the same mistakes. The article Absenteeism forecasting with AI: what pplPortal calculates before the line supervisor details how to prevent the same pattern from repeating itself in that module.

Sources

• Mercer, Global Talent Trends 2023 — average voluntary turnover in Portugal (10.6%) and retention difficulty (52% of companies).
• ManpowerGroup, Talent Shortage Survey 2024 — 65% of employers in Portugal have difficulty finding professionals with the required profile.
• Ordem dos Psicólogos Portugueses, Report on Psychological Health at Work, 2022 — estimated cost of stress and psychological health problems for companies in Portugal (up to €5.3 billion/year).
• Regulation (EU) 2024/1689 of the European Parliament and of the Council (AI Act) — classification of high-risk AI systems, Annex III.
• Regulation (EU) 2016/679 (GDPR) and Law no. 58/2019 (national implementation) — processing of workers' personal data.
• Law no. 93/2021 — general regime for the protection of whistleblowers, requirement for an internal channel for organisations with 50 or more workers.
• Tax and Customs Authority / Social Security — Monthly Remuneration Declaration (DMR), payroll processing obligations in Portugal.

Frequently asked questions

What is factory turnover and why is it different from other industries?

Factory turnover has a different signature because operators do not fill in satisfaction surveys nor have formal evaluations as in offices. Departure signals appear in operational behaviour: attendance patterns, overtime refusals, shift swap requests and output variations. These micro-signals accumulate without being recorded, making prediction difficult without an integrated system.

How much time do I have to act after detecting a risk signal?

In Portuguese manufacturing, the window between the first detectable behavioural signal and the actual termination is, typically, six to twelve weeks. This period is enough for a conversation, a reassignment proposal or a salary review — if the signal is captured in time and the response process is configured to act quickly.

What is the most underestimated signal of departure risk in a factory?

Overtime refusal and shift swap requests are the most underestimated signals. A motivated operator accepts overtime regularly, especially during production peaks. When that pattern reverses, it indicates that something has changed. In isolation they are noise, but cross-referenced with attendance and evaluation history they become a significant alert.

Why is attendance important in turnover prediction?

An isolated absence says nothing. But three absences on Mondays over an eight-week period, combined with early departures and lateness, form a pattern. What matters is calculating deviations against the employee's individual historical average, not against the factory average, since operators have very distinct attendance profiles.

How can a drop in output indicate departure risk?

A sustained drop in individual output, not explained by a change of article or machine, is a behavioural indicator of disengagement. Degraded individual OEE, when correlated with HR data, has predictive value. However, it should be filtered by type of article in production to avoid false positives caused by learning on new articles.

Is there a difference in risk signals between the morning and afternoon shift?

Yes. In factories with two shifts, departure signals tend to appear first in the afternoon shift. Morning-shift workers have more direct contact with management and more internal visibility, creating a social brake. On the afternoon shift, with less presence from the line supervisor, the micro-signals accumulate without being noticed. The predictive model should segment by shift.

What is the main error in configuring factory turnover models?

The main error is using generic models developed for office workers. These models are based on data that does not exist in a factory: satisfaction surveys, formal evaluations and structured feedback cycles. The correct approach is to cross-reference the available operational data — attendance, overtime refusals, shift swap requests and output — configured for the specific factory context.