Human oversight statement

Preamble, nature and effect of this document

Niltech Europe S.r.l., with registered office at Via Calmaggiore 5, 31100 Treviso (TV), Italia, VAT No. IT 05614380268 (“Niltech”), adopts this document in order to make the operator's ability to understand limits, verify sources, correct results, stop the workflow and escalate effective.

This document is a corporate transparency and accountability record. It is not a third-party certification, legal opinion addressed to persons other than the company, absolute security warranty or blanket compliance statement; contractual commitments arise solely from the applicable agreements.

Human oversight is established as the effective and informed authority to understand, verify, correct, reject, stop and escalate. The nominal presence of an operator or a merely formal confirmation is not treated as a sufficient measure.

Personal and material scope

The objective scope includes Andrea AI, its public interfaces and processing strictly connected with the described functions. The commercial website and document library are published on andreaaiagent.com and use Bluehost infrastructure, MySQL and server mail transport. The application is made available through the separate andreaai.net domain on Hetzner infrastructure; Pinecone, Cloudflare, OpenAI and Google Gemini are involved only within the functions and agreements applicable to them.

The relevant operations concern document and image intake and analysis, policy-reading support, estimates and report drafts, case-file organisation, and ESG indicators only where supported by verifiable data. Potential information categories are: contact and prospect data on the website; in the application, case data, documents, images, notes, technical metadata and AI-assisted outputs according to configuration and contract. The actual privacy role, lawful basis and extent of processing depend on the contractual relationship and the lawful instructions of the party determining purposes and essential means.

Definitions and interpretation

• “Service” means the Andrea AI functions made available under the agreement.
• “Customer” means the legal person or professional entering into the agreement with Niltech.
• “Authorised User” means an individual enabled by the Customer to use the Service under its responsibility.
• “Customer Data” means data, documents, images, instructions and other content submitted or generated on the Customer’s behalf.
• “Assisted Output” means a result produced through automated rules or artificial-intelligence components and subject to the stated controls.
• “Further Supplier” means a third party providing Niltech with a technical service relevant to the documented scope.
• “Incident” means an event compromising or capable of compromising confidentiality, integrity, availability, authenticity or resilience.
• “Business Day” means a day other than Saturday, Sunday or an Italian national public holiday.

Specific duties and safeguards

1. Within its assigned role and without prejudice to the Customer’s responsibilities, Niltech shall train the operator on purpose, limits, bias, security and privacy. The applicable file identifies the owner, scope, dependencies, acceptance criterion and evidence; absent those elements, the safeguard is not treated as demonstrated.

2. Within its assigned role and without prejudice to the Customer’s responsibilities, Niltech shall present inputs, provenance and outputs understandably. The applicable file identifies the owner, scope, dependencies, acceptance criterion and evidence; absent those elements, the safeguard is not treated as demonstrated.

3. Within its assigned role and without prejudice to the Customer’s responsibilities, Niltech shall require comparison with sources and context before approval. The applicable file identifies the owner, scope, dependencies, acceptance criterion and evidence; absent those elements, the safeguard is not treated as demonstrated.

4. Within its assigned role and without prejudice to the Customer’s responsibilities, Niltech shall allow editing, rejection, reprocessing and override. The applicable file identifies the owner, scope, dependencies, acceptance criterion and evidence; absent those elements, the safeguard is not treated as demonstrated.

5. Within its assigned role and without prejudice to the Customer’s responsibilities, Niltech shall define stop conditions and available escalation channels. The applicable file identifies the owner, scope, dependencies, acceptance criterion and evidence; absent those elements, the safeguard is not treated as demonstrated.

6. Within its assigned role and without prejudice to the Customer’s responsibilities, Niltech shall record material decisions without using logs for improper evaluation. The applicable file identifies the owner, scope, dependencies, acceptance criterion and evidence; absent those elements, the safeguard is not treated as demonstrated.

Human authority

The authorised operator may accept, correct, ignore or request reprocessing. Professional decisions remain with the competent organisation.

Control points

• case identity and input-suitability check
• output comparison against documents, images and sources
• validation of data, estimates, citations and conclusions
• approval before export or external use

Escalation and suspension

Contradictions, low quality, unexpected sensitive data, possible harm, security anomaly or out-of-scope use require escalation. The user stops the flow and uses the manual process when verification is not possible.

Evidence

Controlled records document training, instruction revisions, exceptions, overrides, incidents and reviews. No automated review percentage is asserted unless measured.

Scope, audience and status of this document

This document is intended for customers, prospects, authorised users, advisers and control functions needing to understand the Andrea AI scope. Its specific objective is to make the operator's ability to understand limits, verify sources, correct results, stop the workflow and escalate effective. It applies to the stated document revision and date and must be read with the applicable agreement, order, DPA, technical specifications and controlled procedures.

The commercial website and document library are published on andreaaiagent.com and use Bluehost infrastructure, MySQL and server mail transport. The application is made available through the separate andreaai.net domain on Hetzner infrastructure; Pinecone, Cloudflare, OpenAI and Google Gemini are involved only within the functions and agreements applicable to them.

Executed agreements and actually approved configurations prevail in case of inconsistency. Public information describes the control programme; it does not turn optional provider capabilities into Niltech controls or automatically attest legal applicability or satisfaction.

Exceptions, non-conformity and escalation

A deviation is not accepted by custom. The owner records the affected requirement, cause, impact, exposed data and persons, compensating measures, approver, expiry and closure criterion. The exception is reviewed if risk changes or a measure does not work as expected.

Incidents, possible unlawful processing, loss of data control, outputs with severe impact, contractual breaches, unapproved suppliers or unreliable evidence must be escalated without delay. Current Legal and functional scope: The documented scope includes the commercial website, the public library and, where expressly stated, processing connected with the application.

• contain risk and suspend the affected phase where needed
• preserve evidence, timing, decisions and communications
• involve privacy, security, product, legal or management owners as appropriate
• resume only after measure verification and documented authorisation

Review, change and improvement

The document is reviewed at least every six months and earlier when purpose, audience, data, GDPR or AI Act role, supplier, model, architecture, location, contractual terms or legal requirements change. Incidents, complaints, failed tests and new vulnerabilities trigger an extraordinary review.

Each review records inputs, participants, decision, changes, superseded evidence, remaining gaps and next date. Material corrections are published without retroactively altering the prior document revision. Contact and requests: info@nil-tech.net.

• check change register and related documents
• retest affected controls
• update manifest, PDF, HTML and hashes
• notify recipients where the change affects their rights or duties

Effectiveness of oversight

A person’s nominal presence in a workflow is not oversight. The reviewer must understand system capabilities and limitations, have access to inputs and sources, sufficient time and competence, autonomy to disagree and authority to correct, disregard, stop or escalate.

For Andrea AI, no assisted output becomes a professional decision merely because it is generated or displayed. The approver makes the decision on the overall file and records the material elements supporting acceptance or departure.

Control points and stop criteria

• before processing: lawfulness, relevance, quality, provenance and excessive data
• on output: completeness, consistency, sources, uncertainty, bias and consequences of use
• before external use: approver identity, recipients, draft status and required warnings
• mandatory stop where a source cannot be verified, essential data are missing, a material conflict exists, risk is unauthorised or the use is excluded

Override, challenge and records

Override does not require proof that the system is certainly wrong; it is enough that output does not meet the reliability criterion for use. The reviewer records correction, source, reasons and any escalation and is not penalised for good-faith disagreement.

Repeated overrides, near misses, complaints and material errors feed into risk review. Records are minimised, protected and retained according to purpose; they are not used for incompatible personnel assessment.

Allocation of responsibility and reliance limitations

Within its sphere of responsibility, the Customer warrants the lawfulness of submitted data and instructions, user authorisation, suitable lawful bases and notices, and professional verification of outputs. Niltech remains responsible for activities directly under its control and does not assume the Customer’s regulatory, professional or decision-making functions.

Outputs from Andrea AI are auxiliary. Unless expressly agreed and subject to mandatory law, they are not legal advice, an expert determination, insurance decision, liability finding, credit assessment or other reserved professional act. The recipient must examine sources, completeness, consistency and consequences before use.

Nothing excludes liability that cannot lawfully be excluded. Outside those cases, attribution, remedies, limitations and quantification principles follow the applicable agreement, taking account of contributory conduct, mitigation duties and foreseeability under the governing law.

Evidence, review, requests and governing law

Every material assertion must be traceable to a contract, approved configuration, register, minutes, test, log or other reliable evidence. Supplier statements and Niltech controls are kept distinct. Absence of incidents is not, by itself, proof that a measure is effective.

Revisions are dated, reasoned and approved. A later revision does not retroactively alter facts or commitments applicable to earlier periods. Published copies are identified by code, date and cryptographic digest; those elements evidence copy integrity, not the substantive effectiveness of described controls.

Reports, clarification requests, rights requests and complaints may be sent to info@nil-tech.net. Niltech verifies identity and authority where necessary, records the request, responds within applicable periods and communicates any reasoned extension or refusal.

Unless mandatory law or a written agreement provides otherwise, Italian law governs interpretation. The Italian text is controlling; the English translation is provided for convenience.

Official sources

• Regulation (EU) 2024/1689 (AI Act)
• Regulation (EU) 2026/1744 — Digital Omnibus on AI
• European Commission — AI literacy following Regulation (EU) 2026/1744