Responsible AI

AI is built into DialPhone across calls, messages, meetings, and the contact center. Our responsibility is to use AI in ways that improve customer outcomes without compromising privacy, fairness, or customer control.

Principles

Customer data privacy

Customer conversations are never used to train shared foundation models. Opt-in fine-tuning keeps data isolated per workspace.

Human in the loop

AI recommendations are reviewable by humans. Agents approve auto-drafted SMS. Clinicians approve AI-summarized notes. No irreversible AI-only decisions in critical workflows.

PII redaction before processing

PHI, PII, and PCI tokens are detected and redacted or tokenized before model processing. Raw sensitive data never crosses the model boundary for shared features.

Transparency

Model cards disclose capabilities, limitations, training data scope, and known failure modes for every AI feature. Bias-testing results published.

Audit logs per AI decision

Every AI action (drafted SMS, transfer decision, summarization) is logged with model version, inputs (tokenized), and the human who approved or overrode.

Right to opt out

Any customer can disable any AI feature. No AI is forced for billing or service eligibility. Opt-out does not reduce service SLAs.

AI features covered by this policy

→ AI live captions and transcription
→ AI meeting summaries and action items
→ AI SMS drafting (Advanced+)
→ Smart Virtual Concierge (AI Receptionist)
→ Proactive Workflow Automation
→ Real-time conversation intelligence (Ultra, Contact Center)
→ Agent Assist and supervisor AI (Contact Center tiers)
→ Predictive CSAT and sentiment analysis

Responsible AI FAQ

Do you train AI on my calls and messages?

Shared foundation models: no. Customer opt-in fine-tunes: only with explicit workspace consent, and data is isolated — never blended with other customers. The default posture is "customer data stays customer data."

What AI providers do you use?

We operate a mix of first-party (DialPhone-trained) and third-party models (Anthropic Claude, OpenAI GPT-family, Google Gemini) behind a safety-and-privacy proxy that strips PII and enforces output filters. Subprocessor list is public.

Can I see what the AI is doing with my data?

Yes. Every AI call is logged with input tokens (sensitive fields tokenized), model version, output, and the human reviewer. Logs are customer-accessible via audit log export.

How do you handle AI hallucinations?

Three mitigations: (1) retrieval-grounded prompts for factual queries, (2) human review for high-stakes outputs (SMS to customers, AI-drafted clinical notes), (3) ongoing eval pipelines that catch regressions before release.

What about bias?

Bias testing is a release gate for every AI feature. Results published in model cards. Continuous monitoring of real-world decisions for demographic disparities. Known issues tracked publicly with remediation timelines.

Can I run DialPhone without any AI?

Yes. AI features are opt-in at the workspace and user level. Core Business Phone and basic Contact Center operate with AI disabled. Billing does not change.