Product
Built for people who need to trust the answer, not just get one.
A model-agnostic Customer Intelligence AI Harness that captures feedback from 60 sources, runs governed multi-agent analysis, and delivers evidence-cited insight. This is how it works — in detail.
Platform architecture
One system. Four layers. Zero black boxes.
Data sources
Survey platforms
Medallia · Qualtrics · SurveyMonkey · Typeform · InMoment · +20 more
25 survey connectors
Review sites
Trustpilot · Google Maps · Amazon · TikTok · TripAdvisor · +many more
35 review sources
Direct upload
CSV · Excel · API · Database export
File & API
Governed data layer
Unified dataset format
One row = one piece of feedback
Automated quality report
Completeness, ranges, anomalies
Knowledge Pack
Metrics · Glossary · Rules · Caveats · Reference docs — LLM drafts, human approves.
Multi-agent analysis pipeline
Each agent: bounded iteration budget · passes findings forward. QA agent challenges findings before delivery.
Outputs
Every output carries provenance: source tier, SQL, freshness.
Layer 1 — Data Sources
60+ connections across survey platforms (direct API), review sites (powered by Apify), and direct file/API upload. Every source is normalized into the same canonical format: one row, one piece of customer feedback.
Layer 2 — Governed Data Layer
Before any analysis runs, datasets receive an automated quality report — completeness, valid ranges, anomalies. The Knowledge Pack sits here: your approved metrics, glossary terms, business rules, caveats, and per-dataset reference docs. The LLM can draft these; humans approve them.
Layer 3 — Multi-Agent Pipeline
Nine specialized agents analyze the data in sequence and parallel, each with a bounded iteration budget. The final QA agent adversarially challenges the findings before they're delivered. Results carry confidence scores and provenance records.
Layer 4 — Outputs
Every analysis result lands in a shared Library. Build dashboards, create narrative stories, export reports, ask follow-up questions conversationally, or access everything programmatically via API.
Capture
60 sources. One format. Zero manual cleanup.
Connect the platforms your customers actually use. InsightNarrator pulls feedback on a schedule or on-demand, normalizes it into a single canonical dataset, and runs automated quality checks — completeness, valid ranges, anomaly detection — before any analysis touches the data.
Survey & feedback platforms (25)
Review & social sources (35)
Direct upload
OAuth2 and API key authentication — no credential storage in plaintext, tokens encrypted at rest
Scheduled sync — pull feedback daily, weekly, or on-demand; analysis always runs on fresh data
Automatic normalization — different schema from different sources → one unified format
Quality report on import — completeness score, anomaly flags, column statistics
Analysis
43 analysis types. Each one an analyst's playbook, made reproducible.
Most platforms give you "sentiment analysis." InsightNarrator gives you 43 analysis types organized by the business question you're actually trying to answer — from "what's driving detractors?" to "where is churn being born?" Each type is an encoded analyst playbook: a senior CX analyst's procedure, made reproducible.
Total: 43 analysis types across 14 categories.
Aspect-Based Sentiment Analysis (ABSA)
What it does. Goes beyond flat "positive/neutral/negative" sentiment. ABSA identifies specific aspects of the customer experience (Pricing, Support Quality, Onboarding, Shipping Speed) and scores sentiment per aspect — so you see that overall sentiment is fine but Shipping → Delivery Delays is generating negative signals this month.
Why it's different. The AI can only assign categories that exist in your governed taxonomy — never free-invented ones. Wave-over-wave comparison stays consistent because "Pricing" always means "Pricing."
| Aspect | Positive | Neutral | Negative |
|---|---|---|---|
| Pricing | 42% | 38% | 20% |
| Support quality | 61% | 27% | 12% |
| Onboarding | 55% | 31% | 14% |
| Shipping → delivery delays | 18% | 24% | 58% |
Complaint-Risk Scoring (C-Risk)
What it does. Scores each piece of feedback for complaint risk — how likely this customer is to churn, escalate, or complain publicly. Risk factors come from the governed taxonomy, crossed with sentiment per category and subcategory.
Why it's different. Instead of a flat "negative sentiment up 12%", you see that Billing → Refund Delay is the specific cell generating severe-risk complaints this month — with severity levels computed from the data, not vibes.
| Category | Subcategory | Severity |
|---|---|---|
| Billing | Refund delay | Severe |
| Billing | Unexpected charge | High |
| Support | Repeat contact | Moderate |
| Product | Feature gap | Low |
Multi-agent pipeline
Nine specialists. One QA gate. Every finding is challenged before it's delivered.
When you run an agentic analysis, nine specialized agents work in sequence and parallel — each with a bounded iteration budget, each passing findings forward. The last agent in the chain doesn't analyze. It challenges everything the others produced.
- →
01
Data
Profile the data quality
- →
02
Domain
Context the domain & expertise
- →
03
Sentiment
Score per aspect & category
- →
04
Themes
Extract the themes & topics
- →
05
Trends
Detect patterns over time
- →
06
Segments
Cluster segments & profile
- →
07
Risk
Score risk factors
- →
08
Synthesis
Combine findings into one report
09
QA
Challenge every finding before delivery
Data
Profiles dataset quality, structure, column statistics
Produces: Data quality report, column semantics, grain definition
Domain
Establishes domain context — what industry, what KPIs matter
Produces: Domain assumptions, relevant metric definitions
Sentiment
Runs sentiment scoring overall, by channel, by topic
Produces: Sentiment distribution, driver analysis
Themes
Extracts themes, topics, and recurring patterns from verbatims
Produces: Theme taxonomy, frequency map, representative quotes
Trends
Detects temporal patterns, wave-over-wave changes
Produces: Trend lines, emerging/declining topics, anomalies
Segments
Clusters feedback into meaningful segments
Produces: Segment profiles, cross-tab comparisons
Risk
Scores complaint risk by category and subcategory
Produces: Risk matrix, severity heatmap, top risk cells
Synthesis
Combines all findings into a coherent narrative
Produces: Executive summary, prioritized findings, contradictions
QA
Adversarially challenges findings, SQL, assumptions
Produces: Confidence scores, evidence citations, corrections
The QA Agent — "Don't self-certify."
The QA agent is the last line of defense. It reviews the SQL the other agents wrote, checks their assumptions against the Knowledge Pack, flags findings that lack sufficient evidence, and adjusts confidence scores accordingly. This is the discipline that independent agentic-analytics research identified as the single largest accuracy improvement — and it's built into every pipeline run.
Knowledge Pack
The semantic layer that makes the AI use YOUR definitions — not its own.
Plugging an LLM into your data without a semantic layer is how you get "analytics slop" — confident, plausible answers that are wrong in ways only a domain expert would catch. The Knowledge Pack is the governed layer between the model and your data.
Without a Knowledge Pack, "revenue" could map to 12 different columns. "Active user" means whatever the LLM thinks it means. With a Knowledge Pack, the agent consults your approved definition first — and raw SQL is the explicitly discouraged fallback, restricted to an allowlist of data tables.
STEP 1
AI drafts
LLM reads the dataset and drafts documentation
STEP 2
Human reviews
Domain expert checks for accuracy
STEP 3
Approved
Entry enters agent context only after approval
Needs review
Schema changed? The doc is re-drafted and flagged. Never silently stale.
Key principle. The model writes the documentation. The human owns the meaning. No Knowledge Pack entry enters the agent's context until a human approves it.
Metrics
Approved business metrics with exact SQL formulas
"Active Customer = status='active' AND last_purchase < 365 days"
Glossary
What terms mean IN YOUR organization
"Detractor = NPS score 0-6, not just anyone who complained"
Rules
Constraints the agent must follow
"Exclude test responses. 'Last month' = last complete calendar month."
Caveats
Documented data traps
"Q3 2026 data excludes migration period — do not trend across it."
Example Queries
Golden question → SQL pairs
"Top 5 churn drivers → SELECT aspect, COUNT(*)..."
Reference Docs
Per-dataset documentation
Business context, entity grain, dimensions, gotchas
Outputs
Insight that doesn't just sit in a report. It lives in your workflow.
Dashboards
Interactive dashboards with charts, KPIs, tables, and text widgets. Build manually or let Scoty generate a Smart Dashboard from your analysis results in one click. Real-time data refresh. Shareable via public links or embedded.
Stories
Narrative stories that combine analysis results, annotations, and commentary. Versioned narrative blocks. Block-level commenting for team collaboration. Not a static slide deck — a living document that updates when the analysis updates.
Reports & Library
Every analysis result is saved in a persistent Library — tagged, searchable, and organized into folders. Export to PDF, HTML, Markdown, JSON, or CSV. Each report carries evidence citations, confidence scores, and per-agent contribution breakdowns.
Scoty AI agent
An agent with 104 tools — not a chatbot with a prompt.
Scoty operates through 10 dedicated tool packs: analysis, datasets, dashboards, stories, library, capture, memory, topics & risks, live agent monitoring, and SQL. It doesn't hallucinate capabilities — it calls real tools that write to real tables.
Analysis
7Run, batch, synthesize, monitor, and retrieve analyses
Stories
28Create, edit, annotate, version, comment on narrative stories
Topics & Risks
22Create taxonomies, queue ABSA jobs, save results, build dashboards
Dashboards
11Create, update, delete dashboards and widgets, share, export
Library
9Folders, move items, update, delete library entries
SQL
9Read-only exploration, temp tables, views, CSV export
Datasets
6Update, delete datasets
Memory
6Create memory entries and links for cross-session learning
Live Agent
3Monitor live observation sessions
Capture
3Create surveys, add questions
Total: 104 static tools + dynamic tools from enabled skills and MCP servers.
User: "What's driving detractors this quarter?"
- 1. Consults Knowledge Pack → approved "detractor"
- 2. Consults Knowledge Pack → approved "this quarter"
- 3. Calls analysis tool → sentiment driver analysis
- 4. Calls SQL tool → verifies against raw data
- 5. Answers tagged: knowledge_pack | high confidence
If the Knowledge Pack doesn't cover the question → raw exploration, tagged raw_exploration | verify before forwarding.
Comparison
How InsightNarrator compares.
InsightNarrator
- Model-agnostic: ✅ 8+ AI providers, switch anytime
- Governed semantic layer: ✅ Knowledge Pack (human-approved)
- Multi-agent pipeline: ✅ 9 agents + QA validation
- Confidence scores & provenance: ✅ On every finding
- ABSA with governed taxonomy: ✅
- Complaint-Risk scoring: ✅ C-Risk by category/subcategory
- Survey platform connectors: ✅ 25 platforms
- Review site connectors: ✅ 30+ sources
- European data sovereignty: ✅ EU & Switzerland-hosted — no US Cloud Act exposure
- Narrative stories & dashboards: ✅ Both
- Agent memory & learning: ✅ Workspace-scoped
- Self-hosted knowledge: ✅ You own definitions
- Credit-based pricing: ✅ Pay for what you use
- Public API: ✅ Scoped keys
Medallia
- Model-agnostic: ❌ Vendor-locked
- Governed semantic layer: ⚠️ Partial
- Multi-agent pipeline: ❌ Single-model
- Confidence scores & provenance: ❌
- ABSA with governed taxonomy: ⚠️ Basic sentiment
- Complaint-Risk scoring: ❌
- Survey platform connectors: Native (own platform)
- Review site connectors: ⚠️ Limited
- European data sovereignty: ❌ US Cloud Act
- Narrative stories & dashboards: ⚠️ Dashboards only
- Agent memory & learning: ❌
- Self-hosted knowledge: ❌
- Credit-based pricing: ❌ Enterprise contracts
- Public API: ⚠️ Limited
Qualtrics
- Model-agnostic: ❌ Vendor-locked
- Governed semantic layer: ⚠️ Partial
- Multi-agent pipeline: ❌ Single-model
- Confidence scores & provenance: ❌
- ABSA with governed taxonomy: ⚠️ Basic sentiment
- Complaint-Risk scoring: ❌
- Survey platform connectors: Native (own platform)
- Review site connectors: ⚠️ Limited
- European data sovereignty: ❌ US Cloud Act
- Narrative stories & dashboards: ✅ Both
- Agent memory & learning: ❌
- Self-hosted knowledge: ❌
- Credit-based pricing: ❌ Enterprise contracts
- Public API: ⚠️ Limited
DIY LLM
- Model-agnostic: ✅ But you manage it
- Governed semantic layer: ❌ None
- Multi-agent pipeline: ❌ Single prompt
- Confidence scores & provenance: ❌
- ABSA with governed taxonomy: ⚠️ Inconsistent
- Complaint-Risk scoring: ❌
- Survey platform connectors: ❌ Manual export
- Review site connectors: ❌
- European data sovereignty: ❌ Depends on provider
- Narrative stories & dashboards: ❌
- Agent memory & learning: ❌
- Self-hosted knowledge: ⚠️ You manage prompts
- Credit-based pricing: ✅ API costs only
- Public API: ❌
Information based on publicly available documentation as of August 2026. Feature comparison is directional, not exhaustive. We respect both Medallia and Qualtrics as category leaders — this table exists to show where InsightNarrator is different, not where they're wrong.
Security & compliance
Security is architecture. Not a certificate on a wall.
Infrastructure
- EU & Switzerland-hosted — no US Cloud Act exposure
- Encryption at rest — all sensitive data encrypted (AES-256)
- Encryption in transit — TLS 1.3 for all connections
- Automated backups — point-in-time recovery
- 24/7 monitoring — uptime and anomaly detection
Data Governance
- GDPR-compliant by design — not bolted on
- No US Cloud Act exposure — EU & Switzerland-hosted
- Right to erasure — full data deletion on request
- Full audit trail — every action logged: who, what, when
- Credential encryption — API keys hashed with bcrypt, OAuth tokens encrypted
Access Control
- Multi-tenant workspaces — strict isolation between tenants
- Role-based access control — Owner, Admin, Member + custom roles
- Scoped API keys — per-key permissions, revocable
- Body limits — 10MB default to prevent DoS
- Rate limiting — configurable per workspace
Integration & API
Connect InsightNarrator to your stack.
Everything in the platform is accessible programmatically. The public API covers datasets, analyses, library items, dashboards, stories, and more — with scoped API keys and full rate limiting.
# Create an analysis job via API
curl -X POST https://api.insightnarrator.com/api/v1/analysis-jobs \
-H "Authorization: Bearer $API_KEY" \
-H "X-Workspace-ID: $WORKSPACE_ID" \
-H "Content-Type: application/json" \
-d '{
"dataset_id": "550e8400-...",
"analysis_types": ["sentiment_driver", "topic_modeling", "absa"],
"config": { "taxonomy_id": "660f9500-..." }
}'Illustrative example — not production documentation.
REST API — full CRUD on all platform entities
Scoped API keys — per-key permissions, bcrypt-hashed at rest
Webhooks — receive notifications on analysis completion, pipeline events
MCP (Model Context Protocol) — connect InsightNarrator as a tool to external AI agents
WebSocket events — real-time job progress, pipeline status
REST API
Full programmatic access to all features
WebSocket
Real-time progress on jobs and pipelines
Webhooks
Push notifications to your systems
MCP Server
External AI agents can use InsightNarrator as a tool source
See it on your data.
Connect a dataset. Run an analysis. Watch the multi-agent pipeline work. In 15 minutes, you'll know if governed agentic analytics is what your CX program has been missing.
No credit card required14-day free trialCancel anytimeEU & Switzerland-hosted — no US Cloud Act exposure