April 2026 Workforce Onsite

Fullstory Data, DOM Decoration, MCP,
and the Agentic Future.

Behavioral Mining & Auto-Doc Gen for Gusto's Automation Roadmap

FullStory Strategic Solutions Team for Gusto  ·  April 2026 Workforce Onsite

A sovereign architecture for closing the loop between specialist behavior and automation logic — leveraging Gusto's existing internal MCP and the Silver Schema sitting idle in Snowflake today.

Agenda

The Roadmap to ROI

01
The Unlock
Connecting the Silver Schema for Outcome Velocity.
02
The Design Pattern
Integrating Lexicon into Gusto's MCP.
03
The Stack
Semantic Decoration → Cortex-Readable Behavioral Data.
04
Cortex Mining
Turning Warehouse Data into Agentic Action.
05
Process Audits
Python CLI for Hippo/Panda Waste Detection.
06
Next Steps
Connecting the Silver Schema to fuel Gumloop.
The Goal

Outcome Velocity

Specialists bounce between Hippo, Panda, and the Wiki. Context switching is the primary source of inflated AHT in Carrier Submissions.

30–60m
AHT in Carrier Submissions
43-day
Variance in State Data Exchanges
The Problem: Fullstory Anywhere Warehouse isn't currently connected, so behavioral data analysis must exclusively be done in Fullstory Analytics. Gusto's semantic layer lives in Snowflake.
The Shift

From Manual Shadowing → Agentic Audit

The 43-day variance in State Data Exchanges isn't random. It's a workflow pattern — and it's already captured in Fullstory's behavioral data. We don't need to go find it. We need to query it.

  1. 1
    Activate the Silver Schema
    Connect the Silver Schema via low-lift ETL. 6 years of behavioral events become queryable in minutes.
  2. 2
    Surface Friction Patterns via Cortex
    Use SNOWFLAKE.CORTEX.COMPLETE to identify state-specific bottlenecks at scale.
  3. 3
    Feed Intent into Gumloop
    Structured "process intent" data becomes the trigger logic for automated workflows.
Reference Architecture

Lexicon MCP: A Blueprint for Gusto

Gusto has already built an internal MCP — that's the foundation we're building on, not replacing. github.com/fullstorydev/lexicon is the reference MCP architecture: an open pattern that teaches developers how to expose Fullstory APIs as MCP tools. Separately, Fullstory's upcoming Analytics MCP will enable direct querying of analytic data living in Fullstory today — a complementary capability, not the same thing.

The Distinction: These two capabilities are complementary, not redundant. Use both.
Warehouse
Historical mass mining. Cohort analysis. "What happened across 10,000 sessions last quarter?"
MCP (Live)
Real-time triage. "The Why." "Why is this specific specialist hitting the GA SUI wiki 4x more than VT?"
What This Means for Gusto
  • Instrumentation is the foundation. The fs-skills library makes your AI coding assistants experts in Fullstory semantic decoration — ensuring every interaction Fullstory captures carries semantic meaning that Cortex can reason over.
  • Gusto's MCP becomes the orchestrator. FullStory tools become callable modules — getSessionInsights, summarizeFriction — that Gusto's agents call when they need behavioral context.
  • The behavioral layer completes the stack. Snowflake has the schema. Cortex has the inference. FullStory has the raw behavioral signal. The MCP integration connects all three.
  • Two MCPs, two jobs. Lexicon teaches you how to call Fullstory behavioral APIs as MCP tools. The upcoming Analytics MCP lets you query Fullstory analytic data directly. Adopt both — they compose cleanly with Gusto's existing stack.
Semantic Instrumentation Layer

The Fullstory Skills Repo

github.com/fullstorydev/fs-skills

A library of AI Agent Skills — Markdown skill files loaded into AI coding assistants (Cursor, Claude Code, etc.) that make them experts in Fullstory implementation and semantic DOM decoration. When your developers use these skills, every interactive element in the app gets machine-readable semantic attributes.

The causal chain: Semantic decoration → Fullstory full capture → every behavioral event in Snowflake carries semantic context → Cortex understands natural language queries against your behavioral data.
What the Skills Cover
  • Stable Selector Decoration — teaches AI assistants to add data-component, data-id, data-section to every interactive element — web and mobile.
  • Event and Identity APIs — correctly implement FS.identify, custom events, page properties with proper schema discipline.
  • Privacy Controls — auto-apply HIPAA, PCI, GDPR decoration patterns. Industry skills for banking, healthcare, SaaS, and more.
  • Framework Patterns — React, Angular, iOS, Android, Flutter — design system auto-decoration strategies.
Why this matters for Cortex: A Cortex agent querying Snowflake can answer "why do VT specialists hit the SUI wiki 4× more?" — because the behavioral data in Snowflake isn't raw DOM events. It's semantically labeled interactions. fs-skills is what puts that label on every click.
The Design Pattern

Lexicon: Tool Definitions for Your MCP

github.com/fullstorydev/lexicon

Lexicon is a reference MCP architecture — an open pattern that teaches developers how to expose Fullstory's behavioral APIs as MCP tool definitions. Not a standalone server. A blueprint you implement in your own stack, connecting Gusto's existing MCP to Fullstory's behavioral data layer.

Strategic note: Gusto already has an internal MCP. Lexicon shows you exactly how to add Fullstory behavioral APIs as native tools — sessions, events, users, session summaries — so your agents can call them like any other tool.
Fullstory APIs → MCP Tool Definitions
GETSessions
GETEvents Stream
GETUser Properties
POSTSession Summary
POSTCustom Events
GETPage Navigation
GETSegments
GETReplay URL
MCP Tool Definition Pattern — JSON Schema JSON
{
  "name": "getSessionSummary",
  "description": "Summarize a Fullstory session via AI profile",
  "parameters": {
    "session_id": "string",
    "profile_id": "string"
  }
}
The Power of Connection

Querying the Silver Schema

The Silver Schema is the prerequisite — a low-lift ETL that lands Fullstory's behavioral events into Gusto's Snowflake environment. Once connected, 6 years of behavioral history is queryable at scale. Cortex is one path; Gusto's existing Cube semantic layer is another.

The question you can ask once connected: "Why are specialists navigating to the GA SUI wiki 4× more often than the VT wiki?" — answered in seconds, not sprints.
Note: Gusto already has a Cube semantic layer. You can expose that directly via MCP — Cortex isn't strictly required. Mark Triassi will demo this in the next presentation.
Implementation Pattern
Snowflake Cortex — Natural Language Query SQL
-- Ask Cortex why specialists are detouring to GA wiki
SELECT SNOWFLAKE.CORTEX.COMPLETE(
  'mixtral-8x7b',
  'Analyze this session cohort. Why are specialists
  navigating to the GA SUI wiki 4x more often than
  the VT wiki? Identify the trigger pages.',
  {
    'context_data': (
      SELECT page_sequence, session_duration,
             carrier_id, state_code
      FROM   fullstory.silver.sessions
      WHERE  carrier_id = 'GA-SUI'
      AND    session_date >= DATEADD('day', -30, CURRENT_DATE)
    )
  }
);
Output: Structured JSON response identifying the exact page transitions and error states that precede wiki navigation — ready to feed directly into Gumloop automation logic.
Architectural Blueprint

Plugging FullStory Intelligence into Gusto's Sovereign Stack

Source Layer
Specialist Browser
Hippo · Panda · Wiki — all specialist tools instrumented via DOM Decoration
DOM Decoration
Event Enrichment
Semantic labels applied to carrier-specific UI elements at capture time
FullStory Capture
Cloud Processing
Session recording → behavioral event normalization → Silver Schema
Low-Lift ETL
FullStory → Gusto Snowflake VPC
Silver Schema: normalized behavioral events, page sequences, friction signals — 6 years of history
Snowflake Cortex
LLM Inference
SNOWFLAKE.CORTEX.COMPLETE queries behavioral data in natural language
Gusto's Internal MCP
Orchestration Layer
Already built — now extended with FullStory Tool Definitions via Lexicon pattern
fs-skills / Lexicon
FullStory Tools
getSessionInsights · summarizeFriction · discoverGroups
Output
Gumloop / PI Reports
Process intent data → automation trigger logic → Markdown audit reports
Design Pattern

Python CLI Pattern: Automating Process Audits

A design pattern — not a pre-built tool — for how Gusto could implement batch behavioral auditing against the Silver Schema without requiring a live agent session for every query.

The Pattern: A lightweight CLI wraps the Lexicon/Skills approach for overnight batch runs — pull from Silver, classify waste indicators, output structured Markdown for PI leads. Adopt and adapt to fit Gusto's stack.
  1. 1
    Pull logs for a specific carrier cohort from Silver
    Filter by carrier_id, state_code, date range, session outcome.
  2. 2
    Apply "Waste Indicators" via Lexicon/Skills
    Classify sessions: excessive wiki navigation, repeated field errors, multi-tool context switching.
  3. 3
    Output: Markdown reports for PI leads
    Structured output ready for PI leads to validate Gumloop automation logic.
Pattern Reference
Waste Audit CLI Python
# Run a waste audit for GA Carrier Submissions, last 30d
python audit.py \
  --carrier "GA-SUI" \
  --date-range "2026-03-01:2026-03-31" \
  --profile "waste-indicators-v2" \
  --output ./reports/GA-SUI-march.md

# Output structure:
# ─ session_count: 1,847
# ─ waste_rate: 38.4%
# ─ top_friction: "Hippo → Wiki redirect at Step 4"
# ─ gumloop_candidates: 712 sessions
Markdown Report Output MD
## GA-SUI Carrier Submissions — March 2026
### Waste Indicators Detected
1. Specialists navigated to wiki at Panda step 4
   in 38% of sessions — median +14m to session time.
2. Field error loop (EIN validation) triggered
   fallback in 22% of sessions. No automation present.
3. Context switches per session: avg 4.2 (target: 1)
Closing the Loop

Prototype Journey: From Silver Schema to Gumloop in 4 Steps

1
Connect the Silver Schema via Low-Lift ETL
Fullstory's Silver Schema lands in Gusto's Snowflake environment via a low-lift ETL. 6 years of behavioral events become immediately queryable — minimal pipeline work, no new infrastructure.
2
Adopt Lexicon/Skills patterns into Gusto's MCP
Implement Lexicon tool definitions — Sessions, Events Stream, Session Summary — into Gusto's existing internal MCP. Fullstory's behavioral APIs become callable tools alongside everything else in the stack.
3
Use Cortex to identify Friction Patterns in Manual Data Exchanges
Run SNOWFLAKE.CORTEX.COMPLETE against the Carrier Submissions cohort. Surface the exact page transitions and error states driving the 43-day variance. Output: ranked friction map by state/carrier pair.
4
Feed "Process Intent" data into Gumloop workflows
Structured friction output becomes the trigger condition for Gumloop. Agents now have the behavioral context to know why a submission is stalling — and can automate the resolution without specialist intervention.
The Business Case

Days to Insight vs. Months of Architecture

Two roads to the same destination. The behavioral intelligence Gusto needs already exists — the only question is whether you build the wrapper or adopt the pattern.

The Slow Road: Build custom data wrappers from scratch. 3–6 months of infra work before the first insight lands.
The Fast Road: Connect the Silver Schema + adopt Lexicon. Days to first insight. Your Cube semantic layer already gives you the query layer — the only work is wiring behavioral data in.
Days
to first insight after Silver Schema connection
Months
saved vs. custom behavioral data wrapper build
ROI Curve
ROI / Insight Velocity Time →
← Custom Wrappers
(3–6 months)
← Lexicon + Silver Schema
(Days)
Today
30d
6mo
Moving Toward Real-Time

Roadmap: StoryAI Opportunities into Gusto's Triggers

The batch audit is the foundation. Real-time is the destination. As Gusto's MCP matures with FullStory tool integration, the same behavioral intelligence that powers nightly reports can power live agent decisions.

Use Case: An agent discovers a new carrier-site error via discoverGroups and automatically updates Gusto's internal automation queues — no human triage required.
Integration Points
  1. 1
    StoryAI Opportunities API → MCP Webhook
    FullStory surfaces emerging friction clusters in real-time. Gusto's MCP receives the signal and evaluates against active Gumloop rules.
  2. 2
    discoverGroups for Carrier Error Detection
    Behavioral clustering identifies when a new carrier portal is generating submission errors before the support queue spikes.
  3. 3
    Auto-Update Automation Queues
    Structured error context from the session summary populates the Gumloop automation queue with the carrier ID, error type, and affected specialist cohort.
  4. 4
    Closed-Loop Validation
    After Gumloop executes, Fullstory confirms the behavioral change — did specialists stop navigating to the wiki at that step?
Future-Proofing

Zero Rework Bridge

The logic you build today is a permanent asset. Because the Silver Schema is the shared source of truth between Gusto's internal MCP and FullStory's Enterprise MCP, the patterns implemented today are 100% compatible with Enterprise MCP features tomorrow.

The guarantee: You are not prototyping a throwaway. You are building the production foundation. Every Lexicon tool definition you implement today is a tool your Enterprise MCP will recognize natively.
Compatibility Map
Today — Gusto's Internal MCP
  • Lexicon tool definitions imported as custom MCP tools
  • fs-skills instrumentation providing semantic event labels
  • Silver Schema in Snowflake, queryable via Cortex or Cube semantic layer
  • Python CLI pattern for batch audit + PI report generation
Tomorrow — FullStory Enterprise MCP
  • Same Lexicon tools — now natively registered
  • Same Silver Schema — same data contract
  • Same skill logic — now managed and versioned by FullStory
  • Same query patterns (Cortex or Cube) — extended with new FullStory primitives
Net result: Zero migration cost. Zero rework. The upgrade path is a version bump, not a rewrite.
Onsite Action Plan

Next Steps: What We Do in the Room

Four concrete actions that move from conversation to committed prototype by end of Day 2. Each one is independently valuable — any single step creates compounding return.

Action 1 — Immediate
Connect the Silver Schema

Stand up the low-lift ETL that lands Fullstory behavioral events into Gusto's Snowflake environment. Minimal engineering work. Once live, 6 years of behavioral history is immediately queryable.

Action 2 — Day 1
Review fs-skills + Lexicon for MCP Integration

Engineering review of github.com/fullstorydev/fs-skills and lexicon. Identify the 3 tool definitions most relevant to Carrier Submissions.

Action 3 — Day 1/2
Query the Silver Schema Against Carrier Submissions

Run the first natural language query against the GA-SUI cohort — via Cortex or Gusto's Cube semantic layer. Validate output structure. Confirm behavioral signal quality for Gumloop consumption.

Action 4 — Day 2
Scope the Waste Audit Pattern for Gusto's Stack

Scope the batch audit design pattern against the March Carrier Submissions cohort. Define waste indicators with PI leads. Identify the top 3 Gumloop automation candidates by session volume.

FullStory Strategic Solutions

Let's activate the Silver Schema
and prototype the
Manual Data Exchange flow now.

Lane Greer · Sr. Manager, Strategic Solutions Specialists

lane@fullstory.com

The behavioral data is already captured. The Silver Schema is ready to share. The Lexicon pattern is documented and open. The only remaining step is the connection.

github.com/fullstorydev/fs-skills  ·  github.com/fullstorydev/lexicon

Technical FAQ — The Hybrid Strategy

Why Adopt the Lexicon Pattern?

Engineering defense points for teams evaluating whether to build vs. adopt.

Don't we already have a behavioral data pipeline? Why add FullStory's layer?
Your pipeline handles structured events — form submissions, API calls, state transitions. Fullstory captures the behavioral layer between those events: the dead clicks, the wiki detours, the repeated field errors that never surface as a structured event but drive 38% of session waste. These are orthogonal data streams, not overlapping ones.
Why not build our own tool definitions instead of adopting Lexicon?
Pre-Built Intelligence: Lexicon's 29 tool definitions represent accumulated prompt engineering and API design for behavioral queries — purpose-built for the exact use cases Gusto needs (friction detection, navigation anomalies, outcome classification). Use them as a template; modify the 20% that's domain-specific. Don't reinvent the 80% that's already proven.
How does this integrate with our existing Snowflake stack?
Warehouse Gravity: Gusto's internal MCP already connects to Snowflake. The low-lift ETL lands the Silver Schema directly into your existing Snowflake environment — same credentials, same warehouse, same query access patterns you already use. The integration point is one new schema, not a new data platform. And if you're already using Cube as your semantic layer, you can query behavioral data through it directly.
What's the PI (Process Improvement) ROI justification?
Outcome-First: The 43-day variance in State Data Exchanges is a measurable baseline. The Waste Audit CLI produces a ranked list of automation candidates by session volume. PI leads validate the list, Gumloop implements the top 3, and Fullstory measures the behavioral change. Variance reduction is the metric — not tool adoption.