Start Your Free Trial

AI Safety Policy

AI Safety Policy

Effective Date: March 1, 2025
Last Updated: May 16, 2026

ReplixAI uses artificial intelligence to automate customer support responses on behalf of businesses. This AI Safety Policy describes how we design, constrain, and monitor that AI to ensure it is safe, accurate, ethical, and trustworthy for both businesses and their customers.

1. What Our AI Does

ReplixAI‘s AI engine performs a single, well-defined function: it reads inbound customer messages on Instagram and Messenger and generates helpful, accurate responses based on the business’s training content.

The AI:

  • Operates only in response to messages customers send first (inbound only)

  • Generates responses based on the business’s uploaded FAQs, product information, and policies

  • Does not access external websites, databases, or knowledge sources beyond what the business provides

  • Does not make decisions — it provides information and escalates when needed

The AI does not:

  • Initiate contact with users who have not sent a message first

  • Send bulk, scheduled, or broadcast messages

  • Conduct social media marketing campaigns

  • Collect or process personal data beyond what is needed to generate a response

  • Make consequential decisions autonomously (e.g., issuing refunds, placing orders) without business-defined rules

2. Our AI Technology

ReplixAI uses the following AI components:

Retrieval-Augmented Generation (RAG) Before generating a response, our system searches the business’s knowledge base for the most relevant information. The AI only answers questions that are covered in the training content. This constrains hallucination and keeps responses grounded in facts the business has approved.

GPT-4o mini (OpenAI) Our language generation layer uses OpenAI’s GPT-4o mini model, one of the most capable and safety-tested AI models available. OpenAI maintains its own Usage Policies that apply to all API-based usage.

Confidence Thresholds If the AI cannot find a sufficiently relevant answer in the knowledge base, it will not attempt to guess or fabricate an answer. Instead, it delivers a graceful fallback message and optionally triggers a human handoff notification.

3. Accuracy and Hallucination Prevention

We take the risk of AI hallucination seriously. Our mitigation measures include:

  • Source-grounded responses: All answers are drawn from the business’s training content, not from general internet knowledge

  • Confidence thresholds: Responses below a confidence threshold result in fallback messages, not guesses

  • Human handoff: Any conversation the AI cannot handle with confidence is escalated to a human agent

  • Business control: Businesses review and maintain their training content and are responsible for its accuracy

Businesses are advised to keep their training content accurate, current, and comprehensive to maximize AI response quality.

4. Human Oversight and Control

AI automation at ReplixAI never operates in a fully autonomous, unmonitored way. Businesses retain full oversight and control:

  • Human handoff: Can be triggered automatically or manually at any time

  • Response review: Businesses can review all AI responses in their dashboard

  • Training content control: Businesses can update, correct, or delete training content at any time, taking effect immediately

  • Deactivation: Businesses can pause or disable the AI at any time from their dashboard or by contacting us

We believe AI should augment human support, not replace human judgment in complex or sensitive situations.

5. Prohibited AI Uses

ReplixAI‘s AI features must not be used for:

  • Deceiving users by claiming the AI is a human agent when directly and sincerely asked

  • Generating misleading, fraudulent, or harmful content

  • Harassing, threatening, or manipulating users

  • Circumventing Meta platform policies

  • Any activity prohibited under our Acceptable Use Policy

Businesses are responsible for ensuring their training content does not instruct the AI to generate prohibited content. Violations may result in account suspension.

6. Transparency With End Users

We encourage businesses using ReplixAI to be transparent with their customers that automated AI responses may be used. Where applicable, we recommend businesses:

  • Include a disclosure in their social media bio or automated response (e.g., “Our responses may be AI-assisted. For complex queries, type ‘Talk to a person’ to reach our team.”)

  • Enable human handoff so customers can always reach a real person

7. Data Minimization and Privacy

Our AI is designed to collect and process the minimum data necessary:

  • The AI reads only the message content needed to generate a response

  • It does not build persistent profiles of end users

  • Customer conversation data is retained for 90 days then deleted automatically

  • No conversation data is used to train shared AI models

See our Privacy Policy for full details.

8. Ongoing Safety Monitoring

We monitor AI performance and safety on an ongoing basis, including:

  • Review of flagged conversations reported by businesses

  • Automated detection of anomalous response patterns

  • Regular review of human handoff triggers to identify AI gaps

  • Updates to model configurations and content filters as needed

9. Reporting AI Safety Concerns

If you believe our AI has generated a response that is harmful, inaccurate, deceptive, or inappropriate, please report it to:

Email: support@replixai.com Subject: “AI Safety Report”

Please include: the business name, a description of the conversation, and the specific response of concern. We will investigate and respond within 5 business days.

Contact

Email: support@replixai.com
Phone: +92 346 3557128