AI system
Talent insights
Last updated
Talent insights is a feature of Bizisy that helps HR notice when a conversation with an employee may help: about their career, pay review, workload, a new manager or a contract that ends. It follows fixed, published rules over job data and shows an advisory band with the reasons behind it. A person always looks at it and decides what, if anything, to do.
It is off until an organization switches it on and completes a seven-step setup that includes telling its people. It is an AI system under the EU AI Act (Regulation (EU) 2024/1689), used in employment, which the Act treats as high-risk; we build it to those rules.
1. What it does
Every hour, for each employee of an organization that switched it on, it adds up points from up to twelve reasons (below) and shows a band: Low (0–2 points), Elevated (3–5) or High (6 or more). Where fewer than six reasons have data, the band reads Limited data. HR sees the band word, up to three reasons with the record and date each came from, and a suggested conversation. The points total is never shown in the app and never sorted by.
It is not a prediction of whether someone will leave, and not a judgement of the person. Nothing happens automatically: no email, notification or change of any kind follows from a band.
2. What it uses
Model attrition-v1. Each reason reads aggregates from the modules the organization uses; a module that is off simply adds nothing.
| Id | Reason | Rule | Points | Data from |
|---|---|---|---|---|
| F1 | In their first year | Time in the company under 12 months | 1 | People records |
| F2 | No change of role or level in 2 years | At least 24 months in the company and no job, level or title change in 24 months | 2 | People records |
| F3 | No pay increase in 18 months | At least 18 months in the company and no pay increase in 18 months (the date only, never amounts) | 2 | People records (pay timing) |
| F4 | Manager changed recently | One manager change in 12 months: 1; two or more: 2 | 1–2 | People records |
| F5 | More short-notice absences than usual | Absence days booked less than 48 hours ahead, of the leave types the employer chose, in 90 days: at least 3 and at least twice the person’s own usual rate | 2 | Time off |
| F6 | Unexplained gaps in clocked time | At least 3 scheduled days in 60 days with no clock record and no leave of any type | 1 | Time clock |
| F7 | Sustained overtime | Overtime above 10 % of scheduled hours in 8 of the last 12 weeks | 2 | Time clock |
| F8 | Heavier shift pattern than their team | Night or weekend shifts at least 1.5 times the team median over 8 weeks, or 3 or more acknowledged rest warnings | 1 | Shifts |
| F9 | Lower rating than last cycle | Final rating down at least one level between the last two released review cycles | 1 | Performance |
| F10 | Most goals at risk | More than half of active goals at risk or overdue | 1 | Performance |
| F11 | A pay or role request was declined or is waiting | A pay change request open over 30 days or declined in 90 days: 2; a title request declined in 90 days: 1 | 1–2 | People records |
| F12 | Fixed-term contract ends soon | A fixed-term or temporary contract ends within 90 days with no renewal recorded | 2 | People records |
3. What it never uses
Date of birth or age, sex, pronouns, nationality, birthplace, name, photo, language, address or commute, marital or family status, dependants, emergency contacts, health, disability, union membership, religion; sick leave and any health-related leave, parental, maternity, paternity, carer and family leave, and the “other” leave type; any leave type until the employer marks it as counting (health-related types can never be marked); part-time status or working hours share; pay amounts or pay level; any free text (notes, 1:1s, review comments, messages, documents); clock-in location; survey answers; candidate data.
Some reasons can correlate with protected characteristics even so: time in the company and fixed-term contracts with age, shift patterns with sex and caring duties, overtime. That is why they are tested (section 5).
4. Human oversight
- Only the people the employer names see bands (by default its HR roles). The employer may also show bands and job-related reasons to each person’s managers, after a warning; pay-related reasons then stay with HR.
- Each new Elevated or High band starts unreviewed. A named oversight person acknowledges it (optionally planning a conversation), dismisses it with a reason, or changes the band with a reason; the model’s band stays visible beside theirs.
- The panel always says what the band is not, and the instructions forbid using it alone for pay, promotion, discipline or dismissal, telling a person they have a band, or asking about health.
- Every band worked out, every view, review and setting change is recorded and kept for 12 months; the employer can read and download that record.
- A stop switch deletes every current band at once.
5. The people it is about
Before any band is worked out, the employer tells its people, in Bizisy or another way, and consults workers’ representatives where national law requires it. Each person can, in their own workspace:
- object: they are excluded at once and any current band is deleted;
- ask for an explanation of any band about them, which HR answers there;
- download their data, which includes their band if one exists.
Nobody is shown their own band in the app.
6. Testing
We test each model version on a synthetic population of 10,000 made-up people, with protected characteristics assigned at random and with deliberate links between them and the job data (for example part-time work and sex, time in the company and age, night shifts and sex), so that indirect effects show. For each group we compare the rate of Elevated and High bands with everyone else and flag any ratio outside 0.8–1.25. These are synthetic results, not a measure of accuracy in a real workforce.
Before switching on, each employer runs the same fairness check on its own people with a file of groups used for testing only: it is processed in memory, only results for groups of 10 or more people are kept, and the file is never stored. Organizations with fewer than 50 employees, whose groups are too small to compare, may rely on the synthetic test instead.
7. Instructions for employers
The employer is the deployer of the system and responsible for using it according to its instructions. In short:
- How it works: read the instructions and the five-minute guide for the people who review bands.
- What counts: choose which leave types count for short-notice absences. Health, parental, family and “other” leave are locked out.
- Who may see: name oversight persons with the competence, training and authority to review; decide whether managers see bands.
- Tell your people: send the ready notice (in six languages) or record how you informed them.
- Impact assessment: complete the data protection impact assessment template, filled in with your settings.
- Fairness check: run it on your own people (under 50 employees, you may rely on the synthetic test).
- Switch on: the first bands appear within the hour.
Not validated for teams under 10 or organizations under 50 people. Not for candidates. Never rank people by it. For the full instructions for use, write to us at hello@bizisy.com.
8. Provider
Wisely Lda, NIPC 516845128, Amares, Portugal, which operates Bizisy. Contact: hello@bizisy.com. Employers report problems or incidents from Settings → Talent insights → Report a problem. How Bizisy handles personal data is in our privacy policy.
Questions? Write to privacy@bizisy.com about privacy, or hello@bizisy.com about anything else.