[ LEAVEPILOT · LEGAL ]
Courtesy translation
This translation is provided to make the document easier to read. The text that binds the parties is the Italian one: in the event of any discrepancy the original prevails, and it is its fingerprint that is recorded when the agreement is accepted.
- Version
- ai-instructions-2026-08-12-v1
- In force since
- 12 August 2026
- SHA-256 fingerprint of the text
- 3a37a4ca0ce7418fba88fb8d0653b9a507eae3823d4369a381f0b7f2813cd432
- Verifiable against a copy of the text with shasum -a 256 content/legal/ai-instructions-2026-08-12-v1.en.md. If the fingerprint matches, the text has not been altered since publication.
LeavePilot: instructions for use
Document supplied by the provider to the deploying company Version: 1.0 of 11 August 2026 System: LeavePilot, summary report generation function Provider: MAXYMIZE BUSINESS DI GIURASTANTE ROMANO MAXIMILIAN, Via Delle Valli 57, 66010 Canosa Sannita (CH), Italy, VAT no. IT02747200695, info@maxymizebusiness.com
This document describes what the automatic report commentary function does, what it does not do, what limits it has and who must check its output. It follows the model of article 13 of Regulation (EU) 2024/1689, which imposes such a document on high-risk systems: LeavePilot does not fall within that category, for the reasons given in section 3, and these instructions are supplied all the same, because the distinction depends on how the system is used, and whoever uses it must be in a position to know where the boundary lies.
The Italian version prevails / La versione italiana fa fede.
1. What the system does
LeavePilot manages requests for holidays, hourly leave and remote working. Balance calculations, the application of approval workflows, payroll exports and all tables and charts in reports are produced by deterministic code: given the same input data, the result is always the same and can be verified.
Only one part of the product uses an artificial intelligence model: the descriptive commentary that appears at the end of the report. It receives a set of already-calculated aggregate figures and describes them in a few sentences.
2. What the system does not do
The artificial intelligence model:
- receives no data that identifies a person. The text sent to it contains counts, averages, distributions by utilisation band and departmental aggregates. It contains no first names, surnames, email addresses or internal identifiers, no rows relating to an individual employee, and no individual minimum or maximum values. Departments of fewer than three people do not appear as a separate entry, because a group that small would identify its members even without naming them;
- does not approve and does not reject requests. Decisions on absence requests remain entirely with the company's own people;
- does not calculate balances, remaining days or accruals. Those figures come from deterministic code and can be verified;
- does not assess, rank or score people;
- does not form hypotheses about wellbeing, motivation, fatigue or workplace climate, and does not attribute causes to the differences it observes;
- does not recommend actions concerning people, groups or departments.
These exclusions are not entrusted to an instruction given to the model: the first two are guaranteed by the fact that the data never reaches it, the others by an automatic check that discards the text before it enters the report if it contains a name present in the data, a term referring to personal states, or a verb of recommendation. Discarded text is neither corrected nor regenerated: the report is issued without commentary.
3. Intended purpose and limits of use
Intended purpose: administrative management of absences. The descriptive commentary serves to make a set of figures readable, not to support decisions about people.
Non-conforming use. The report, and in particular the named table showing days per employee, must not be used to assess, rank, compare or monitor the performance or conduct of individual workers, nor as a basis for decisions on promotions, allocation of tasks, financial terms or termination of employment.
The reason is not a formal precaution. Annex III to Regulation (EU) 2024/1689, area 4 point (b), classifies as high-risk those systems intended to monitor and evaluate the performance and behaviour of persons in work-related relationships. LeavePilot remains outside that category because its purpose is administrative and the generated text evaluates no one. If the deploying company uses it for that purpose, under article 25(1)(c) of the same regulation the company itself becomes the provider of the high-risk system, with all the obligations that follow: risk management, technical documentation, event logging, designed human oversight, conformity assessment and CE marking.
4. An obligation that concerns the employer
This point does not concern the provider and is flagged so that the deploying company can assess it with its own advisers.
In Italy, article 4 of the Workers' Statute makes instruments from which the possibility of remote monitoring of workers' activity arises subject to a trade union agreement or to authorisation by the Labour Inspectorate. The second paragraph exempts from that obligation the instruments used to perform the work and those recording access and attendance, a category into which the management of absence requests reasonably falls.
Processing that produces per-employee measures, such as the individual utilisation percentage in the report, is however an operation beyond simple recording. If the company intends to use it, it is advisable to verify its own position with regard to that provision and to the information obligations towards workers and trade union representatives laid down by article 1-bis of Italian legislative decree 152/1997.
5. Accuracy and known limits
The commentary faithfully describes the data it receives, including data that is wrong. If an input figure is inconsistent, the commentary flags it as an anomaly but does not correct it. The report contains a data quality block that counts recurring anomalous cases: people with days recorded beyond their available allowance, people with no allowance set, requests with no type. Those counts indicate anomalies in the recorded data, not people's behaviour.
The commentary may not appear at all. This happens in three cases, all foreseen: the model provider does not respond, the text produced is discarded by the automatic check, or the function is switched off. In all three the report is issued complete with tables and charts, lacking only the commentary section. The absence of the commentary does not indicate a problem with the data.
The commentary is generated automatically and is not reviewed by a person before it is published in the report. It is identified as such within the report itself, by a visible label and by technical marking in the document metadata, in accordance with article 50(2) of Regulation (EU) 2024/1689.
6. Human oversight
The system does not require oversight in order to function, because it makes no decisions. It does however require the deploying company to:
- designate a person responsible for reading the reports before they are used as the basis for any communication or decision;
- check the anomalous data flagged in the data quality block instead of treating it as marginal;
- report to the provider, at the address given at the head of this document, any content of the commentary that turns out to be wrong, inappropriate or directed at individuals.
7. Traceability
Every report containing a generated commentary records, in its own metadata, the model that produced it and the version of the instructions used. A report without a commentary records nothing, because there is nothing to attribute. This information makes it possible to reconstruct how a disputed text was produced, even long afterwards.
8. Model provider
The commentary text is produced by a model from Anthropic, called directly and without intermediaries. No other provider may serve that call: if the named provider is unavailable, the report is issued without commentary. Anthropic is named as a sub-processor in the data processing agreement, which also states the processing region and the retention periods declared.
9. Changes
The instructions used by the system are versioned. Any change that alters its behaviour entails an update to this document and notification to the deploying company.
Next scheduled review: 11 August 2027.
All published versions
Every version stays at its own address, with its own text and its own fingerprint, even once superseded. It is there for whoever is bound by an earlier version and needs to read back exactly that one.
- ai-instructions-2026-08-12-v112 August 2026in force