- Please describe how AI is used within your product.
- Does the intended use of AI in your tool include marketing, underwriting, rating, claims, and/or fraud detection applications?
- Will the AI-driven outputs in the tool result in a decision made for the user, or merely inform a decision-making process?
- Does Variphy have a formal AI policy?
- Do Variphy’s AI tools use generative AI?
- Do you retain or use customer prompts or data for training or any other purpose?
- Does your application address the risk of GenAI “hallucinations”? How is that risk addressed?
- Do you conduct model testing of the tool?
Please describe how AI is used within your product.
Colleen AI Assistant is an optional, licensed natural-language helper that currently supports two functions: (a) Call History search, which turns a plain-English request into structured call detail record (CDR) search criteria; and (b) Reports, which builds and configures call analytics reports. The model only emits structured search or report criteria, which Variphy Insight validates and executes against the customer’s real data. The model does not read raw CDRs or compute analytics itself.
Does the intended use of AI in your tool include marketing, underwriting, rating, claims, and/or fraud detection applications?
None of these. Variphy Insight is a unified communications CDR reporting and provisioning tool. Colleen is confined to Call History search and report building; there is no underwriting, rating, claims, fraud, or marketing decisioning anywhere in the product.
Will the AI-driven outputs in the tool result in a decision made for the user, or merely inform a decision-making process?
Strictly informs and assists. Colleen produces a search or report that the user reviews and runs; a human stays in the loop. The model has no authority to act, change provisioning, or make automated decisions. Its tool calls only build query and report objects that Insight executes and displays. When Colleen builds a report, it saves the report configuration, and the user still owns review and scheduling.
Does Variphy have a formal AI policy?
Yes, a Responsible AI Policy exists today at Variphy’s Trust Portal.
The Colleen AI Assistant whitepaper is also available here.
Do Variphy’s AI tools use generative AI?
Yes. Colleen uses a large language model (LLM) via an OpenAI-compatible chat completions API, with tool calling and structured outputs. The model is DeepSeek V4.1 Flash, served by Fireworks AI (serverless inference) through the Variphy Cloud AI gateway.
Do you retain or use customer prompts or data for training or any other purpose?
No. Customer data is not used to train any AI models.
Conversation history is stored in the customer’s own configuration database, not on Variphy’s side. By default, each conversation holds up to 10 messages and each user can keep up to 10 conversations. The oldest conversation is pruned when a new one is created. Users can delete conversations. Conversation history is not stored in the cloud, except for Variphy Cloud customers, whose conversation history is stored in their hosted Variphy environment. Conversations are not deleted on a timer; only the caps above apply.
What leaves the install: the current prompt, recent conversation context, system instructions, and tool exchanges are sent to the gateway and model to answer each request. Raw CDR result sets are not sent through the search tools, but the user’s typed message is sent verbatim and is not redacted. Anything a user types, including names and numbers, goes to the model.
Other purpose: The conversation session ID and user ID are sent only to group spend logs for billing and usage tracking, not for content reuse.
Does your application address the risk of GenAI “hallucinations”? How is that risk addressed?
Yes, through design and grounding rather than trusting free-form model output:
Structured criteria: the model must emit structured criteria through tool calls, and Insight validates and executes the real query or report. The model never supplies the actual call data or numbers.
Strict schemas: tool schemas are grammar-constrained, and invalid search types are rejected.
Clarifying questions: the model can ask the user a clarifying question instead of guessing.
Runtime limits: the model runs at a low temperature (0.2), and a limit on tool calls per request, along with a repeat guard, stops runaway loops. When Colleen is not licensed, the feature is disabled.
Do you conduct model testing of the tool?
Yes. Colleen’s behavior is tuned through manual testing, and the underlying system is covered by extensive deterministic unit tests of schemas, criteria parsing and conversion, routing and authorization, gateway error handling, and tool-loop limits. Automated performance testing and a scored evaluation suite against the live model are not currently used.
