Responsible AI and security
How Alobri uses AI (it proposes, a person decides), agents, AI governance, information security, incidents and continuity.
Version 1.1 · Effective November 16, 2026
Part A. Responsible AI Policy
A.1 Guiding principle
At Alobri, artificial intelligence, algorithms, machine learning and agents analyze and propose. Every decision is made by an authorized person, based on that analysis.
A.2 Commitments
- Human oversight: no decision about people (hiring, rejecting, assessing, promoting, terminating, paying) is executed without the approval of an authorized user.
- Transparency: every AI-generated result is labeled "AI-generated proposal" and shows the main factors considered.
- Review: any assessed person can ask for a human to review a result.
- Non-discrimination: models do not use sex, age, pregnancy, marital status, origin, religion, disability, health, sexual orientation or their proxies (photograph, name, postal code) to score people. Bias testing is carried out before launch and every year.
- Prohibited uses: we do not infer emotions, sexual orientation, political opinions, religion or health from the face, voice or behavior; we do not perform social scoring; we do not use manipulative techniques.
- Data: each Customer's data serves only that Customer; to improve models we use anonymized Aggregated Data; our AI provider does not train on our Customers' data.
- Customer control: the Administrator can turn AI off by module and defines what each agent may do.
- Security: we protect models and agents against malicious instructions hidden in documents, emails or CVs.
- Reporting: anyone can report an error or bias to ia@alobri.com or with the "Report result" button.
A.3 Alobri AI systems
| System | What it does | Who decides | Risk level |
|---|---|---|---|
| TalentIA (assistant) | Answers, drafts, autocompletes and suggests | The user | Medium |
| Candidate analysis | Summarizes profiles, proposes a ranking and flags matches with the vacancy | The recruiter | High |
| Assessments and reports | Scores tests and generates interpretations | Whoever administers the test | High |
| Test proctoring | Detects signs of possible cheating | A person reviews every alert | High |
| Facial recognition in Time | Verifies that the person checking in is the employee (one-to-one) | The employee can use a PIN or QR | High |
| HR and operations predictive analytics | Proposes turnover, performance, scheduling and headcount scenarios | The Customer | High |
| Module agents | Prepare tasks, drafts and calculations | Human approval for irreversible actions | High |
| Commercial recommendations | Suggest plans and modules based on the Customer's configuration | The Customer | Medium |
| Discriminatory language blocking | Flags prohibited text in job postings | The recruiter | Low |
A.4 Providers
Alobri's AI uses Anthropic models. Alobri may change provider or model without reducing these protections.
A.5 Connecting third-party assistants (Claude)
Each person can connect Alobri One to their own Claude account, by Anthropic, through the Alobri connector (MCP protocol). The connection uses the alobri.com sign-in, always shows the permissions screen and only gives access to the companies and data that account can already see. Claude reads and prepares drafts; only if a company administrator allows it can the person accept proposals or post drafts from the chat, and only after confirming it themselves. It never closes periods or makes payments. Every action is logged with the person’s name. What the person queries is sent to their Claude account and is governed by Anthropic’s terms with that person or their company. The connection can be revoked at any time from the “Claude” tab in Alobri One.
Part B. AI Agents Policy
- Per-module permissions: each agent can only read and act within the module it belongs to and with the permissions the Customer enables.
- Actions that always require human approval: validating or stamping payroll; sending mass communications; rejecting, hiring or terminating people; deleting data; paying or charging; changing permissions; publishing job postings.
- Limits: maximum number of actions per hour and credits per day, configurable by the Customer.
- Log: each action records the agent, version, data used, proposal, who approved it and the result. The super administrator can view it in the console.
- Shut-off: a button to stop each agent instantly.
- Untrusted inputs: text from CVs, emails, documents and web pages is treated as data, never as instructions. An agent never executes commands found inside a document.
- Rubber-stamp detection: if a user repeatedly approves every proposal within seconds, the Platform flags it and asks for review.
- Alobri internal agents (engineering, QA, SRE, audit): they are not sold to Customers; they work without real Customer data; no change goes to production without express human approval.
Part C. AI governance
C.1 Inventory and model cards
Each system in table A.3 has a card covering: purpose, input data, prohibited variables, model and version, quality metrics, bias test results, risks, controls and internal owner.
C.2 Impact assessment
Before launching or materially changing a high-risk system, the following is documented: who it affects, what can go wrong, how it is mitigated, how it is monitored and which law applies in each country.
C.3 Bias testing
- Before launch and every year, by country and with test data.
- Selection or scoring rates are compared across groups; if the difference exceeds the defined threshold (reference: the four-fifths rule), the system is not launched until it is corrected.
- For customers in New York, an annual independent audit and publication of the summary.
C.4 Versioning
Each result stores the version of the model, algorithm and rules that produced it.
C.5 AI incidents
A harmful or discriminatory result or an improper action by an agent is treated as an incident: the system is stopped if necessary, corrected, the affected Customer is notified and the incident is recorded.
C.6 Regulatory calendar
- EU: prohibited practices in force since February 2025; high-risk employment obligations from December 2, 2027.
- United States: New York, Illinois, California and Colorado (see the US local terms).
- Mexico: notice of automated processing (LFPDPPP 2025).
- Brazil, Peru and Quebec: right to review of automated decisions.
Part D. Information Security Policy
D.1 Responsibilities
- Security officer: the person designated by Alobri's management, contact security@alobri.com.
- Everyone with access to Alobri systems signs a confidentiality agreement and receives annual training.
D.2 Access control
- Least privilege and role-based access.
- Mandatory two-factor authentication for administrators, cloud accounts, GitHub, Stripe and email.
- Quarterly access review; immediate removal when a relationship ends.
- Cloud accounts and repositories in the company's name, never personal.
- Secrets and keys in a secrets manager, never in the code.
D.3 Data protection
- Encryption in transit (TLS 1.2 or higher) and at rest.
- Logical separation by Customer and tests showing that one Customer cannot see another's data.
- Even more restricted access to biometric and health data.
- Separate production and test environments; tests only with fictitious data.
D.4 Logging and monitoring
- Log of access, permission changes, exports and agent actions, kept for 12 months.
- Alerts for anomalous access.
D.5 Providers
- Security assessment before engaging a provider that processes data.
- Contract with data protection and breach notification clauses.
D.6 Responsible disclosure
Researchers can report vulnerabilities to security@alobri.com. Alobri will not take action against anyone who reports in good faith, without damaging data or services and without disclosing before the fix.
D.7 Information classification
| Level | Examples | Rule |
|---|---|---|
| Public | Website, published prices | Unrestricted |
| Internal | Processes, plans | Alobri staff only |
| Confidential | Contracts, finances, code | Need-to-know access |
| Restricted | Customer data, biometrics, health, payroll | Minimum access, encryption and logging |
Part E. Incident response
| Step | What is done | Target time |
|---|---|---|
| 1. Detect | Automatic alert, report from a Customer or researcher | Immediate |
| 2. Classify | Severity and whether personal data is affected | 2 hours |
| 3. Contain | Isolate, revoke access, stop agents or features | As soon as possible |
| 4. Notify the Customer | If Customer data is affected | Without undue delay; at most 48 hours from confirmation |
| 5. Notify authorities | When the law requires it of the controller (Alobri when it is the controller; the Customer when it is) | Per country (EU: 72 hours) |
| 6. Fix and restore | Fix, restore from backup | Depending on severity |
| 7. Document | Report with cause, impact, actions and lessons learned | 10 business days |
Alobri publishes incidents affecting availability on its status page.
Part F. Business continuity and backups
- Automatic daily database backups, encrypted and stored in a region different from the primary one.
- Restore test at least every quarter.
- Targets: restore the service within 8 hours and lose at most 24 hours of data in a disaster.
- Emergency contact list and backup access under management control.
- Contingency plan if a critical provider fails (cloud, payments, AI, stamping).
Other languages
Legal · United States · Master Cloud Services Agreement · Non-Disclosure Agreement (NDA) · Legal contact: legal@alobri.com · Privacy: privacidad@alobri.com