LEGAL · AI ETHICS & SAFETY

AI Ethics and Safety

Last updated: 1 May 2026

PsychFlo applies artificial intelligence to workplace behavioural data. This page sets out the ethical principles, design constraints, and safety mechanisms that govern how our AI systems operate — and what they are explicitly prohibited from doing.

FUNDAMENTAL PRINCIPLE

PsychFlo is a decision-support system. It is not a diagnostic tool, not a clinical instrument, and not an automated decision-maker.

Every AI output is a probability and a prompt for human review — not a verdict about any individual. All consequential decisions remain with qualified HR professionals and managers. Our AI assists; it does not decide.

1. What Our AI Does

PsychFlo uses machine learning to identify patterns in aggregated workplace behavioural metadata. Specifically, our AI:

✓Aggregates passive behavioural signals (communication metadata, calendar patterns, task completion rates) at team level
✓Compares current patterns against each team's own 28-day rolling baseline to detect anomalies
✓Runs signals through a 9-engine ensemble model to produce a Workforce Risk Score
✓Generates natural language summaries of risk patterns and recommended actions using Claude (Anthropic API)
✓Identifies when predefined risk thresholds are crossed and triggers configurable alerts

2. What Our AI Explicitly Does Not Do

✗Diagnose any individual with a mental health condition, burnout, or psychological disorder
✗Make or record employment decisions (hiring, firing, promotion, demotion) — automated or otherwise
✗Read the content of any message, email, document, or audio/video recording
✗Generate individual-level outputs about identifiable employees' psychological states
✗Share any data about individual employees with their manager without HR admin authorisation
✗Apply scoring to teams of fewer than five members (de-anonymisation prevention)
✗Use employee data to train AI models without explicit client consent
✗Produce outputs that claim certainty — all risk scores include confidence ranges

3. AI Regulatory Alignment

PsychFlo is designed in alignment with emerging AI regulatory frameworks globally, including the EU AI Act and applicable guidance on AI and data protection. Specifically:

Safety

AI outputs are designed to reduce harm, not create it. Risk scores prompt human review — they do not trigger automated sanctions.

Transparency

Employees whose data is processed are informed via their employer. Platform users see model confidence levels on all outputs. No black-box scoring.

Fairness

Our model is benchmarked against each team's own baseline — not population norms — reducing the risk of systematic bias against particular groups or demographics.

Accountability

Every AI recommendation requires human approval before being logged as an action. Responsibility for decisions remains with HR professionals.

Contestability

HR admins can dispute, override, and suppress any AI output. Employees can request exclusion from scoring at any time.

4. Automated Decision-Making (GDPR Article 22)

GDPR Article 22 restricts decisions based solely on automated processing that produce legal or similarly significant effects on individuals. PsychFlo complies with this requirement as follows:

✓All AI outputs are team-level aggregates. No individual-level automated decision is produced.
✓Where AI narration suggests actions affecting individuals (e.g. 'consider a 1-1 with a team member'), this is a recommendation to a human decision-maker — not an automated decision.
✓HR professionals must review and approve any action before it is recorded in the platform.
✓Employees have the right to request human review of any risk assessment that influences an HR decision about them.

5. Bias and Fairness

We take the risk of algorithmic bias seriously, particularly in a workplace context where biased outputs could disadvantage protected groups under applicable equality and anti-discrimination legislation.

Personal baseline comparison

Every individual's signals are compared against their own baseline — not a population average — which reduces the risk that atypical work patterns (e.g. part-time, neurodivergent, caring responsibilities) generate false risk flags.

Team-level aggregation

Individual signals are never assessed alone. Team-level patterns are required for any output, meaning single-person variation is absorbed into group context.

Protected characteristic exclusion

PsychFlo does not collect or process protected characteristic data (age, gender, ethnicity, disability, religion, or local equivalents) and does not allow risk scores to be filtered or segmented by such characteristics.

Ongoing bias monitoring

We commit to periodic review of model outputs for demographic parity concerns. Client organisations may request a bias audit report for their deployment.

6. AI Model Transparency

We are committed to explainability. PsychFlo does not operate a black-box scoring system. All risk outputs include:

→A Workforce Risk Score (0–100) with a confidence interval
→The top three signal drivers contributing to the score (e.g. 'Meeting load elevated 42% above baseline', 'Response latency increased 28% week-on-week')
→A plain-language explanation of what the signals indicate and the recommended review actions
→A clear label on all AI-generated text: 'AI-generated decision support — requires human review'

7. Third-Party AI (Claude / Anthropic)

PsychFlo uses the Claude API (Anthropic) to generate natural language summaries of risk patterns and recommended actions. The following safeguards apply to this use:

✓Claude is used only for text generation (narrating risk patterns) — not for risk scoring or decision-making
✓No personal data about individual employees is sent to the Claude API. Only anonymised, aggregated signal summaries are used as prompts
✓All Claude outputs are reviewed and framed by PsychFlo's output layer before display — raw API responses are never shown directly
✓Anthropic processes data under a Data Processing Agreement with standard contractual clauses (SCCs) for GDPR compliance

8. Reporting Concerns

If you believe a PsychFlo AI output has been applied inappropriately, generated a biased result, or been used to make an employment decision without appropriate human review, please contact us immediately:

AI Ethics & Safety Contact

Email: info@psychflo.com

Subject: “AI Ethics Concern” — we will acknowledge within 2 business days and respond within 10.