Converse — AI Chat Interface

Overview

Converse is Prizm’s AI-powered conversational interface that allows users to manage data assets, trigger workflows, generate recommendations, and explore the data catalog — all through natural language. Powered by Claude (Anthropic) and Azure Open AI(where user can selected the required model), Converse uses MCP (Model Context Protocol) tool integrations to autonomously invoke the right Prizm capabilities based on user intent.

How Converse Works

User Input (natural language)
        │
        ▼
Intent Detection
(What does the user want to do?)
        │
        ▼
Tool / MCP Selection
(Which Prizm capability handles this?)
        │
        ▼
Execution + Context Gathering
(Searches catalog, lineage, profiling, org context)
        │
        ▼
Recommendation Cards or Action Results
(Displayed as structured, actionable UI components)
        │
        ▼
User Reviews → Approve / Reject / Create

No function calls. No SQL. No API invocations. Just natural language.

What You Can Do in Converse

Data Discovery

"Show me all tables in the Finance domain that haven't been updated in 7 days"
"What are the most critical assets downstream of the customer_transactions table?"
"Find all assets tagged PII in the Marketing product"

Metadata Generation

"Generate descriptions for all tables in the sales schema"
"Update attribute descriptions for the orders table"
"Recommend business terms for the customer domain"

Quality Management

"Recommend quality metrics for the orders table"
"What data quality issues are open in the Finance domain?"
"Run a profile on the new_customers table"

Glossary & Governance

"Generate a glossary for the Customer360 domain"
"Create a business term for 'Customer Lifetime Value'"
"Recommend domains for the recently discovered assets"

Execution & Creation

Converse can directly trigger actions — not just recommend them:

Action	Example Prompt
Create domain	”Create a new domain called Logistics”
Create glossary terms	”Add the recommended terms to the Finance glossary”
Run profile job	”Profile the top 10 tables by usage in the Data Warehouse”
Run quality job	”Run all completeness metrics on the orders table now”
Apply asset mappings	”Apply the recommended domain mappings for the new assets”

AI Model Configuration

Converse routes requests to the appropriate model based on task complexity:

Task Type	Default Model	Rationale
Asset descriptions	Lightweight (Claude Haiku)	High volume, simple generation
Metric descriptions	Lightweight (Claude Haiku)	Templated output
Business Metric Recommendation	Advanced (Claude Sonnet/Opus)	Requires deep reasoning
Governance Recommendation	Advanced (Claude Sonnet/Opus)	Complex domain analysis
Glossary Generation	Advanced (Claude Sonnet/Opus)	Organization context synthesis
Conversational Q&A	Claude Sonnet	Balanced cost/capability

Model selection is configurable by platform Admins under Organization → Platform → Configuration. Prizm supports both Azure OpenAI and Anthropic (Claude) models; which provider/model is assigned to each task type by default varies by release, so check that configuration screen for the current mapping rather than treating any single model name in this page as fixed.

Contextual Awareness

Converse uses the following sources to enrich its responses:

Existing Domains — Current domain structure and ownership
Existing Glossary — Business terms already defined in the catalog
Asset Metadata — Descriptions, tags, profiling statistics, quality scores
Organizational Summary — Company context and business vocabulary
Lineage — Upstream and downstream asset relationships
Usage Insights — Who uses what data and how frequently

Recommendation Cards

When Converse generates recommendations, they appear as structured cards in the chat:

┌─────────────────────────────────────────────────────┐
│ 🤖 Quality Metric Recommendation                     │
│                                                     │
│ Asset: sales.orders                                  │
│ Metric: Completeness — customer_id                   │
│ Confidence: 97%                                      │
│ Rationale: customer_id is used in 23 downstream     │
│ models and has historically maintained 99.8%+       │
│ completeness. A drop below 99% would impact         │
│ revenue reporting.                                  │
│                                                     │
│  ✅ Approve    ❌ Reject    ⚙️ Customize             │
└─────────────────────────────────────────────────────┘

Users can approve, reject, or customize individual recommendations. Approved recommendations are automatically applied without leaving the Converse interface.

Permissions

All Converse actions respect:

Module licensing — Features available based on your Prizm license tier
Mode configuration — Guided vs. autonomous execution modes
User permissions — Role-based access controls (RBAC/ABAC)

A Member-role user will not be able to trigger metric creation even through natural language — permissions are always enforced at the execution layer.

Guided vs. Autonomous Mode

Mode	Behavior
Guided	Converse presents recommendations and waits for user approval before any action
Autonomous	Converse executes approved action types automatically based on configured trust levels

Autonomous mode is configurable per workflow type and requires explicit Admin enablement.

title: “Converse — AI Chat Interface” description: “Prizm’s natural-language interface for querying, configuring, and acting on your data through conversation”

Overview

Converse is Prizm’s conversational AI interface — the surface where the Multi-Agent Architecture becomes something you can simply talk to. Instead of navigating menus or writing SQL, you describe what you want in plain language, and Converse identifies intent, selects the right tools, and either answers directly or executes a workflow on your behalf.

How Converse works

At its core, Converse is powered by the Prizm Agent (master conversational controller) and a Q&A Summarize Agent that turns raw tool output into clean, readable answers — both described in detail in Multi-Agent Architecture. Converse operates in three modes depending on what you’re asking for:

Conversational
Workflow
Hybrid

Natural dialogue for explanations, insights, and clarifications — no tool execution required.“Why did this table’s freshness score drop last week?”

Structured, multi-step execution: Converse extracts parameters, validates them, runs the appropriate tool, and returns a result.“Connect a new Snowflake source called sales_prod.”

Workflow execution examples

Generate SQL query — you describe what you need; the SQL Query Generator constructs an optimized, context-aware query and returns it for execution
Recommend DQ rules & generate measures — the DQ Rule Recommendation Engine suggests rules and auto-generates measures based on SQL logic
Recommend a semantic term for an attribute — the Semantic Term Recommendation Engine analyzes metadata and existing ontologies to suggest a contextually relevant business term

What you can ask Converse to do

Converse supports a wide library of natural-language prompts across metadata, governance, profiling, and quality workflows — scoped to an asset, attribute, domain, product, application, or tag:

Descriptions

“Update Asset Description for this asset”
“Update all Asset Descriptions” — optionally scoped to a domain, product, application, or tag
“Update Attribute Description for this attribute”
“Update all Attribute Descriptions” — optionally scoped to a domain, product, application, or tag

Semantics & glossary

“Recommend Term for this asset / attribute / domain / product / application / tag”
“Generate glossary” — using organization context to build categories and business terms
“Generate glossary for this category / domain”

Profiling & quality

“Recommend profile for this asset”
“Run profile for this asset”
“Recommend quality metric for this asset”
“Run quality metric for this asset”

GPT-generated recommendations only populate metadata fields that are empty, unless explicitly configured otherwise — existing, manually curated metadata is never silently overwritten. Every generated recommendation moves into a review state (e.g. Ready for Review) so a human can accept or reject it before it’s applied.

Tool-driven, not code-driven

Converse doesn’t require you — or Prizm’s own engineers — to write a function or API call for every capability. It uses the same MCP (Model Context Protocol) tool ecosystem described in Multi-Agent Architecture: Converse automatically identifies and invokes the appropriate tool based on your intent, then chains multiple tools together for multi-step reasoning. This tool-based design means Converse can:

Search available sources, assets, and governance structures
Analyze metadata and organizational context
Generate contextual recommendations, with reasoning explaining why each one was made
Present recommendations as interactive cards you can review, select, approve, or reject
Trigger creation and execution directly from the conversation — creating domains, generating glossary terms, applying mappings, running profile or quality jobs

Converse supports both guided/manual workflows (you drive each step) and autonomous workflows (Converse chains tools on its own), depending on your selected mode — the same Human / AI Guidance / AI Assisted / AI Completed spectrum described in AI Stewardship.

Model support

Converse is model-agnostic by design, with model selection tuned to task complexity:

Model tier	Used for
Lightweight / lower-cost models	Asset description, attribute description, metric description, metric purpose and rationale
Higher-capability / reasoning models	Business metric recommendation, governance recommendation, domain/glossary recommendation, advanced quality recommendations, complex AI insights

Model routing is configurable at the platform or workflow level, and all AI-driven actions executed through Converse respect module licensing, mode-level configuration, and user permissions.

Security, privacy, and data handling

Converse’s underlying LLM infrastructure is built on Azure OpenAI and Anthropic with the following guarantees:

Guarantee	Detail
Encryption	All data to and from the model service is encrypted in transit and at rest
No training on customer data	Prompts and completions are not used to improve any Microsoft or third-party models; Prizm generates its own synthetic data for internal model development, with no access to customer data
Tenant isolation	Data is logically segmented — never available to other customers or to the model provider itself
Dedicated API keys	Each customer can use their own Azure OpenAI subscription, or one dedicated to them and provided by Prizm
Compliance	Operates within Microsoft’s compliance framework — ISO 27001, HIPAA, GDPR
Model lifecycle	Models are versioned through a centralized registry; admins have full control over model selection from the Prizm portal
Auditing	Model performance and prompt management are logged for audit; automated monitoring alerts on anomalies or security incidents

Common questions

Are there AI-specific clauses in the contract? Not currently.Is there added risk from using Converse? Prizm gives customers full flexibility to bring their own Azure OpenAI subscription, or use one dedicated exclusively to them.Does anything get pushed to my data warehouse? Converse uses the LLM to generate SQL, which is executed as a push-down operation into your warehouse (e.g. Snowflake) through your existing connector. All AI reasoning happens within Prizm — nothing is pushed to the warehouse beyond the resulting query.Is my data used to train the model? No. Prizm does not use customer data for model training, and generates synthetic data internally for that purpose instead.

File attachments and governance writes — current limitations

File uploads are accepted but not ingested. The Converse chat interface allows you to attach a file (for example, a PDF), but Converse does not currently read or extract content from attachments. If you attach a document and ask Converse to act on the information inside it — for example, “add the glossary terms in this PDF to the source” — Converse will not parse the file. Depending on the request, it may create an empty object (such as a glossary with no terms) and then ask you to supply the terms directly. Until file ingestion is supported, provide the specific values you want acted on in your prompt text rather than relying on an attachment. Broad governance writes can affect more than expected, with no built-in undo. A single confirmed Converse action that links a domain (or similar governance object) to a source “and all related assets” can apply more broadly than the request implies — extending to attributes, and to assets that were deliberately left out of that domain. Converse does not currently show a full preview of everything a write will change before you confirm it, and there is no one-step undo for a governance write once it’s applied. Before confirming a broad or source-scoped governance instruction:

Prefer narrower prompts that name specific assets rather than “all related” or source-wide phrasing.
After the action completes, spot-check a sample of the affected assets and attributes rather than relying only on the summary count Converse reports.
If a broad write needs to be reversed, plan to manually revert the affected domain/attribute assignments — there is no single undo action for this today.

Multi-Agent Architecture

The agent and tool infrastructure that powers Converse

AI Stewardship

How actions triggered through Converse are governed and reviewed

Get Started

Core Concepts

Platform

Architecture

Data Sources

Integrations

Security & Access Control

Deployment

Help

Settings

MCP

Release Notes

API Reference

Converse — AI Chat Interface

Overview

How Converse Works

What You Can Do in Converse

Data Discovery

Metadata Generation

Quality Management

Glossary & Governance

Execution & Creation

AI Model Configuration

Contextual Awareness

Recommendation Cards

Permissions

Guided vs. Autonomous Mode

title: “Converse — AI Chat Interface” description: “Prizm’s natural-language interface for querying, configuring, and acting on your data through conversation”

Overview

How Converse works

Workflow execution examples

What you can ask Converse to do

Tool-driven, not code-driven

Model support

Security, privacy, and data handling

File attachments and governance writes — current limitations

Multi-Agent Architecture

AI Stewardship

​Overview

​How Converse Works

​What You Can Do in Converse

​Data Discovery

​Metadata Generation

​Quality Management

​Glossary & Governance

​Execution & Creation

​AI Model Configuration

​Contextual Awareness

​Recommendation Cards

​Permissions

​Guided vs. Autonomous Mode

​title: “Converse — AI Chat Interface” description: “Prizm’s natural-language interface for querying, configuring, and acting on your data through conversation”

​Overview

​How Converse works

​Workflow execution examples

​What you can ask Converse to do

​Tool-driven, not code-driven

​Model support

​Security, privacy, and data handling

​File attachments and governance writes — current limitations

​Related pages

Multi-Agent Architecture

AI Stewardship

Overview

How Converse Works

What You Can Do in Converse

Data Discovery

Metadata Generation

Quality Management

Glossary & Governance

Execution & Creation

AI Model Configuration

Contextual Awareness

Recommendation Cards

Permissions

Guided vs. Autonomous Mode

title: “Converse — AI Chat Interface” description: “Prizm’s natural-language interface for querying, configuring, and acting on your data through conversation”

Overview

How Converse works

Workflow execution examples

What you can ask Converse to do

Tool-driven, not code-driven

Model support

Security, privacy, and data handling

File attachments and governance writes — current limitations

Related pages