Artificial Intelligence Transparency Statement

On this page

1. What is Fonema?
2. How Fonema's AI Works
3. Limitations and Critical Warnings
4. Voice Data Processing
5. Training Data Policy
6. Client Responsibilities
7. Known Risks
8. Applicable AI Regulations
9. Continuous Improvement
10. Contact for AI Concerns
11. Glossary of Terms
12. Additional Resources

Fonema Inc. Last updated: March 30, 2026

1. What is Fonema?

Fonema is a conversational artificial intelligence platform that automates communications via phone, WhatsApp, and web. Our AI agents can:

Make and receive phone calls
Hold conversations via text message on WhatsApp
Provide support and chat responses on web widgets
Record, transcribe, and analyze conversations
Take actions based on configured business logic

Fonema is designed to be a non-human agent that provides communication automation to businesses and organizations in Latin America.

2. How Fonema's AI Works

2.1 Technical Architecture (High-Level)

The Fonema platform operates using the following flow:

1. VOICE/TEXT INPUT    End-user → Microphone/Keyboard → Captured Audio/Text  2. INPUT PROCESSING    Audio → Speech-to-Text (STT) Transcription    Text → Intent and Context Analysis  3. LLM PROCESSING    Transcription/Text → Large Language Model (Claude, GPT-4, etc.)    LLM processes input and generates response  4. RESPONSE SYNTHESIS    Text Response → Text-to-Speech (TTS) Synthesis    Generated Audio → Transmission to user  5. STORAGE AND ANALYSIS    Full Recording → Encrypted Storage    Transcriptions → Behavior Analysis    Metadata → Reports and Optimization

2.2 Language Models

Fonema uses third-party large language models:

OpenAI (GPT-4, GPT-3.5): General conversational processing
Anthropic (Claude): Alternative processing with a focus on safety

These models are pre-trained on public and private historical data. Our use of these models introduces inherent limitations and risks.

2.3 Custom Configuration

Customers can customize agents using:

System Prompts/Instructions: Directives that define the agent's behavior
Dynamic Context: Information about the end-user, interaction history, customer data
Logic Flows: Conditions and actions executed based on user responses
Knowledge Bases: Documents, policies, or data that the agent can consult

The quality of responses depends directly on the quality of these configurations.

3. Limitations and Critical Warnings

3.1 Hallucinations and Inaccuracies

Fundamental limitation: Language models can generate responses that are:

Factually incorrect
Fabricated or "hallucinated" (information that does not exist)
Inconsistent with instructions
Grammatically disordered or confusing

Example: If you request information about a customer policy, the agent might generate a response that sounds plausible but is completely false.

Mitigation:

Use well-structured context/knowledge bases
Regularly monitor transcripts
Implement manual validation for critical decisions
Use agents in "consultant" mode (requires human approval) for important transactions

3.2 Not a Real Person

Critical requirement: End-users MUST know they are speaking with AI, not a person.

The Fonema agent:

Is NOT an employee, salesperson, or human representative
Has NO personal experiences, emotions, or consciousness
Cannot make independent legal or medical decisions
Does NOT have persistent memory between conversations (unless explicitly configured)
MAY make mistakes a human would avoid

Failure to disclose its AI nature violates emerging regulations and may constitute fraud.

3.3 No Capacity for Professional Advice

Legal Disclaimer: Fonema agents must NEVER:

Providing legal advice
Providing medical advice or diagnosis
Providing financial or investment advice
Providing accounting or tax advice
Providing mental health counseling or therapy

These areas require licensed human professionals. If the user needs professional advice, the agent must immediately escalate to a human or provide professional contacts.

3.4 No Emotions, Consciousness, or Personal Opinions

Conceptual Clarity:

The agent does NOT emotionally "understand" the user's problem
The agent does NOT have personal preferences, political beliefs, or opinions
The agent is NOT capable of genuine empathy (it can simulate emotion recognition)
The agent does NOT have independent intentions or desires

Any response that seems personal is algorithmically generated, not genuine.

3.5 Quality Depends on Configuration

Agent quality depends on:

Quality of the prompt/system instructions
Completeness of the knowledge base
Quality of context data
Selection and tuning of the LLM model
Accuracy of conversation flow logic

An agent with poor configuration will produce poor responses, even when using powerful models.

3.6 No Performance Guarantees

Fonema does NOT guarantee:

Specific conversion rates
Specific accuracy or error rate
That the agent will understand all user requests
That responses will be completely accurate or helpful
That the agent will not offend some users
That the agent will solve all customer problems
Specific business outcomes

Performance varies depending on use case, configuration, and end-users.

4. Voice Data Processing

4.1 Voice Data Flow

When an end-user calls:

Capture: Audio is transmitted encrypted to Fonema servers
Transcription: Audio is converted to text using STT (Speech-to-Text) service
LLM Processing: Text is sent to language model (OpenAI/Anthropic)
Voice Generation: Response is converted to audio using TTS (Text-to-Speech) service
Storage: Full recording, transcription, and metadata are stored

4.2 Recording Storage

Duration: Recordings are retained according to client configuration (typically 30-90 days)
Location: Stored in Amazon S3 in the United States (us-east-1 by default)
Encryption: In transit (TLS) and at rest (AES-256)
Access: Limited access to authorized Fonema personnel for support
Compliance: Complies with retention regulations based on jurisdiction

4.3 Transcriptions and Metadata

Processed and stored as non-audio data
Used for reports, analysis, and agent improvement
Available for client download
Processed by sub-processors (OpenAI, Anthropic, AWS) under DPA

5. Training Data Policy

5.1 Client Data Protection

Clear policy: Fonema does NOT use client data to train AI models without explicit consent.

End-user conversations are NOT used to improve models
Client configuration data is NOT used for training
Personal data is NOT used for training

This is our DEFAULT policy (opt-out).

5.2 Consent for Training

If a client gives EXPLICIT consent to use training data:

Anonymization: Data will be anonymized and personal identifiers removed
Aggregation: Data will be combined with other clients' data
Transparency: Client will receive reports on data usage
Revocation: Consent can be revoked at any time

This is configured in the client's Master Service Agreement (MSA).

5.3 Third-Party Training

The LLM models we use (OpenAI, Anthropic) have their own training policies:

OpenAI: May use data to improve models (check OpenAI Business Terms)
Anthropic: Does not use enterprise client data for training without consent

Fonema can request an opt-out on behalf of the client with OpenAI for enterprise accounts.

6. Client Responsibilities

6.1 Mandatory AI Disclosure

The client MUST:

Disclose that it is AI to end-users before or at the start of interaction
Do so in a clear, accessible, and immediately understandable manner
In the end-user's language
Comply with emerging regulations (EU AI Act, local regulations)

6.2 Behavior Monitoring

The client MUST:

Regularly monitor agent transcripts and outputs
Verify that the agent does not make critical errors
Monitor compliance with regulations (TCPA, recording, consent)
Audit knowledge base for accuracy

6.3 Appropriate Configuration

The client MUST:

Configure the agent appropriately for your use case
Use clear and precise prompts
Provide updated context and knowledge base
Escalate to humans for complex or critical issues

6.4 Regulatory Compliance

The client MUST:

Comply with all applicable laws (TCPA, GDPR, CCPA, LGPD, etc.)
Obtain required end-user consents
Comply with local recording regulations
Comply with AI disclosure requirements

6.5 Prohibition of High-Risk Areas

The client MUST NOT use Fonema for:

Medical diagnosis or treatment
Legal advice
Critical financial decisions
Identity authentication (without additional verification)
Discriminatory decisions regarding employment, credit, or housing

7. Known Risks

7.1 Bias

AI models can reflect biases in training data:

Gender bias in job recommendations
Racial bias in credit assessments
Age bias in product offerings
Language bias (better performance in English than in less represented languages)

Regularly audit responses for bias; monitor affected users.

7.2 Privacy and Security

Potential risks:

Insertion of sensitive data into prompts (card numbers, SSN)
Prompt injection to bypass safeguards
Interception of unsecured communication
Unauthorized access to recordings

Mitigation: Use encryption; implement access controls; provide security training.

7.3 Toxicity

Models can generate:

Offensive or discriminatory language
Hateful content responses if provoked
Sexually explicit language

Mitigation: Use guardrails; supervise responses; implement content filters.

7.4 Failed Legal Compliance

Non-compliance risks:

Failure to disclose AI (fraud)
Recording without required consent (TCPA)
Making spam calls to numbers on the DNC list (TCPA)
Non-compliance with emerging AI regulations

Mitigation: Review Acceptable Use Policy; consult local legal counsel.

8. Applicable AI Regulations

8.1 European Union AI Act

Article 50 (Transparency): AI systems must inform users that they are interacting with AI.

Applies if users are in the EU
Fonema complies by disclosing in client settings
Client is responsible for implementing disclosure to end-users

8.2 Emerging Local Regulations

Multiple jurisdictions are developing AI regulations:

Mexico: LFTR includes transparency requirements for automated communications
Brazil: LGPD requires disclosure of automated processing
Colombia, Argentina, Chile: Emerging regulations under development

Recommendation: Consult with local legal counsel regarding compliance.

8.3 Data Privacy Laws

Existing privacy regulations apply:

GDPR (EU): AI processes must comply with GDPR
CCPA (California): Disclosure of consumer data use
LGPD (Brazil): Consent for personal data processing
LFTR (Mexico): Telecommunications and privacy regulations

9. Continuous Improvement

Fonema invests in improving:

Accuracy: Reducing hallucination and error rates
Security: Strengthening against prompt injection and attacks
Fairness: Reducing bias in responses
Compliance: Upholding emerging regulations

This statement will be updated as technology and regulations evolve.

10. Contact for AI Concerns

If you have AI-related concerns or incidents:

Email: ai-safety@fonema.ai Subject: "AI Safety/Compliance Concern"

Provide:

Description of concern
Specific transcripts or examples
Use case context
Date and time of incident

11. Glossary of Terms

Term	Definition
Hallucination	When an LLM generates false information that appears plausible
STT (Speech-to-Text)	Technology that converts audio to text
TTS (Text-to-Speech)	Technology that converts text to speech audio
LLM (Large Language Model)	AI model that generates language based on learned patterns
Prompt	Instruction or indication given to the AI model
Token	Small unit of text processed by an LLM
Bias	Systematic tendencies towards biased outcomes
Prompt Injection	Attack where a user modifies the LLM's behavior through special input

12. Additional Resources

Note on entities: These terms are issued by Fonema Inc. For customers in Mexico contracting through Fonología, S.A. de C.V., the specific conditions set forth in Annex D of the Master Services Agreement apply.

Fonema Inc. Last updated: March 30, 2026