How Vertiqa's AI stays safe
The governance model behind Vertiqa's agents — human approval gates, no autonomous external sends, per-org budgets and rate limits — plus how models are handled and the honest answer on training.
AI-safety questions are guaranteed in any 2026 evaluation. The good news: Vertiqa's answer is a design philosophy, not a defensive caveat.
The thesis: AI drafts and chases, humans approve
The one line that carries the whole conversation: AI drafts and chases, humans approve. Agents surface work and prepare it; a person decides. That's not a limitation bolted on — it's how the system is built.
What stops the AI from doing something wrong on a client's behalf
Concretely, and all shipped:
- Human approval gates on agent actions. A governed action waits on an explicit human approval before it executes. Nothing consequential happens unattended.
- No autonomous external sends. Agents do not send email to a customer on their own. Drafting and proposing, yes; sending without a human, no.
- A full audit trail. Agent actions are recorded in the same typed compliance audit log that covers the rest of the system (see Security and tenant isolation).
- Per-organization budgets and rate limits. Each org has AI spend budgets and rate limits, so an agent can't run away — with cost or with volume.
This is the strongest part of the story. Deliver it as the product thesis, not an apology.
Which models, and what happens on an outage
Vertiqa uses a multi-provider architecture — OpenAI (via Azure OpenAI) and Anthropic (Claude) are both in use today, with provider resilience so a single provider's outage doesn't take the feature down. Provider costs are Vertiqa's to manage; a client's pricing doesn't fluctuate with model prices.
"Is our data used to train models?"
The client-facing answer is no — Vertiqa does not use customer data to train models, and the model providers are used through their commercial business APIs, whose standard terms do not train foundation models on data submitted through them.
Deliver that clearly and without hesitation. When a client wants it in writing — a specific data-processing or no-training clause — route that through Vertiqa rather than improvising contractual language on a call. The posture is firm; the paperwork is a Vertiqa conversation.