Decagon

Decagon states it enforces zero-data retention with its AI providers so that no conversation data is stored by them or used for training. The statement covers the providers; it does not address whether Decagon's own fine-tuned guardrail models are trained on customer conversations.

Logs and transcripts are retained after a conversation ends (they are redacted of PII just after it ends), but no retention period or deletion schedule for those records is published.

Requires written confirmation, see Before you sign ↓

Decagon names OpenAI and Anthropic as examples of the AI providers it uses, framed as 'all AI providers like', which indicates an open-ended set rather than a complete list; no specific models are identified.

Requires written confirmation, see Before you sign ↓

Decagon is headquartered in the United States and may transfer website/representative personal information to the US or other countries with less protective privacy laws. This is scoped to the public-facing Service, not to customer conversation data, whose processing location is not published.

Confirm in writing: Ask the vendor to state this in writing before signing.

Decagon publishes a security.txt naming [email protected] as the reporting contact, a PGP encryption key, and a security policy URL. This is a vulnerability-disclosure channel; no customer-facing breach-notification commitment or timeframe appears in the collected documents.

Confirm in writing: Ask the vendor to state this in writing before signing.

Decagon publishes a coherent control stack for its agents: a pre-deployment testing suite (regression, replay, unit and integration tests), in-conversation bad-actor detection and a supervisor model that checks every response for hallucination, post-conversation QA via Watchtower, customer-defined human escalation rules, and zero-data retention enforced with its AI providers so conversation data is not stored or used for training by them.

This is more operational detail on AI-specific risk controls than most vendors publish, and it directly addresses prompt manipulation, hallucination and provider-side training risk. All of it is vendor-published, so it should be confirmed contractually and, where possible, in a trust-centre evidence pack.

Question for vendor: Can you provide the testing-suite and guardrail descriptions as contractual commitments, and the provider zero-data-retention terms in writing?

The Security page links to a Trust Center, but none of the seven collected documents names SOC 2, ISO/IEC 27001, ISO/IEC 42001, a penetration test or any other third-party attestation. The independent_assurance domain is marked not_evidenced rather than not_applicable: a hosted SaaS AI platform processing customer conversations can and should hold such assurance, and comparable vendors publish it.

Every security and AI-safety statement in this entry rests on Decagon's own word. Without an attestation report the buyer cannot verify that the described controls operate as stated.

Question for vendor: Which certifications or attestations (SOC 2 Type II, ISO/IEC 27001, ISO/IEC 42001) does Decagon hold, what is their scope, and can the reports be shared under NDA via the Trust Center?

The public privacy policy expressly excludes data processed for business customers, deferring to private agreements. The only location statement (US headquarters, transfers to the US or elsewhere) is scoped to website and representative data. Logs and transcripts are retained and PII-redacted after a conversation, but no retention period is given. The zero-data-retention commitment applies to third-party AI providers; whether Decagon trains its own fine-tuned bad-actor and supervisor models on customer conversations is not stated.

A buyer cannot establish from public material where conversation data is processed, how long transcripts are kept, or whether their customers' conversations contribute to Decagon's own model training. These are the questions most likely to arise in a privacy impact assessment.

Question for vendor: For conversation data: where is it processed and stored, what is the retention period for logs and transcripts, and is any customer conversation data used to train or fine-tune Decagon's own models (including guardrail and speech models)?

OpenAI and Anthropic are named as examples ('all AI providers like'), not as a complete list, and no specific models are identified. 'Model redundancy' is listed as a resiliency feature, which implies requests may fail over between models or providers. Google's DLP service is disclosed for PII redaction. No subprocessor register was found; the OpenAI and Databricks items are partnership announcements, not processing disclosures.

Zero-data retention is only as strong as the set of providers it covers. If fallback routes traffic to an unnamed provider or model, the buyer's data-recipient picture is incomplete.

Question for vendor: Please provide the full list of AI model providers and models in the request path, the fallback/routing policy, and a current subprocessor register with regions.

Decagon markets Duet Autopilot as a self-improving agent and has published a post on governance of self-improving agents in production, but only the title was collected. The evidenced change control is customer-side: replay and regression testing to detect unintended behavioural change. Separately, the privacy policy may be modified unilaterally with notice by posting.

An agent that updates its own behaviour in production shifts change management from a release process to a continuous one; buyers need to know what approvals, rollback and monitoring apply before an autonomous change reaches their customers.

Question for vendor: For Duet Autopilot, what approval, rollback and audit controls govern self-initiated changes to agent behaviour, and can customers require human sign-off before changes go live?

Technical evidence (DNS, certificate, or HTTP) shows a dependency on Microsoft as a service provider. This dependency is not identified in the vendor's published materials.

An undisclosed provider dependency is exactly the gap independent verification exists to surface.

Question for vendor: Verification required — Microsoft appears to be involved as a service provider: confirm whether this dependency exists, and whether it processes customer data.