01Six principles
The principles below are binding on the design, development and operation of every AI feature. If a feature does not meet one of them, it is not released.
A human makes the decision
Decisions that affect an employee directly — rejecting an application, disciplinary action, a payroll deduction, cancelling leave — are made only by an authorised user. The model produces a suggestion; approval belongs to a person.
It does not run invisibly
Every screen where AI is in play says so plainly. The user can see on that same screen which inputs produced a score or a suggestion.
Protected characteristics never enter the model
Gender, age, nationality, marital status, disability and similar characteristics are not given to assessment models as inputs. Even where those fields are held in the employee record, the model does not see them.
Models are not trained on customer data
One customer’s data is not used to improve another customer’s results. Model development is done with anonymous, synthetic or open-source datasets.
There is a route to object
An employee or candidate adversely affected by an AI-assisted assessment can ask for the decision to be reviewed again by a person. This right corresponds to the right to object under Article 11 of Law No. 6698.
It is audited regularly
Assessment models go through periodic bias and accuracy testing. Findings are recorded, corrections are made, and a summary report is shared with customers who ask for it.
02What each feature does, and what it does not
Every AI feature is described by the same two questions: how it works, and where it stops. Without the second, the first is not a commitment.
CV match score (ATS)
The candidate’s CV is compared against the requirements in the posting through text analysis, producing a match percentage. The matching and missing headings behind the score are listed.
Candidates are ranked, not eliminated. The decision to reject is made only by a recruiter. Demographic characteristics are not inputs to the model.
Leave conflict warning
Approved leave is compared against the shift plan to flag in advance which day will leave which unit short-staffed.
The warning is shown to the manager; it does not reject a leave request automatically. Leave accrual and balance are calculated by a rules engine under the legislation, not by AI.
Performance trend analysis
Development areas and trends are drawn out of past goal completion rates and feedback records.
It is presented to the employee and manager as information. It cannot be tied automatically to a promotion, bonus or exit decision.
Payroll anomaly detection
Unusual deviations in monthly payroll items — an unexpected deduction, a repeated payment, implausible overtime — are flagged onto a review list.
The system only flags. It never changes, deletes or approves a payroll item on its own.
Company assistant
It answers an employee’s question from the company’s own procedures and guidelines, and cites the document the answer rests on.
Answers are for information and do not replace a formal HR decision. The assistant reaches only the documents the user is authorised to see.
03Model training and use of data
AI features run on customer data, but they are not trained on it. The difference is this: the model reads your data to produce an answer; it does not learn that data permanently and use it for another customer.
- Only anonymous, synthetic or open-source datasets are used in model development.
- One customer’s data does not affect another customer’s results; data isolation is enforced at infrastructure level.
- AI features reach only the records the user is already authorised to see; the assistant cannot cite a document outside those permissions.
- Data sent in prompts to an external provider is minimised, and a data processing agreement is signed with that provider.
For data categories, retention periods and transfer rules, see the data protection and security page.
04Bias testing and audit
Models used in recruitment and assessment go through regular bias testing. The test measures whether the same competency profile produces a different result when protected characteristics are changed. Where a meaningful deviation is found, the feature is stopped, corrected and retested.
How we share this today
05Employee and candidate rights
A person affected by an AI-assisted assessment can ask: whether AI was used in the assessment, which inputs produced the result, and for the decision to be reviewed again by a person.
These requests go to the organisation you work for; the employer holds the data controller role. When the employer refers the request to us, we carry out the technical review and respond in writing.
06Updates and contact
This policy is reviewed at least once a year; when a new AI feature is released, it is updated with that feature’s entry. Material changes are notified to customers by email in advance.
You can send questions about the policy, and objections relating to AI features, to info@hrandtomorrow.com.