SI CE Practice · Three-Play Motion
Internal · Nov 2026
Customer Engagement Practice

Three plays. One motion.

The SI has a defined CE Assessment. Assessment creates the account relationship and produces the gap map that drives two parallel Phase 2 tracks — Ontology Accelerators and AgentOS / Harness. All three are execution priorities for 2027.

Phase 1 · Entry
CE Assessment
4-week fixed-price diagnostic · 8-pillar scoring · 6 agent classes · CDO findings delivery
↓ CE Maturity Score · Context Gap Map · 90-day Roadmap → Account entry
Phase 2A · Parallel
Ontology Accelerators
Domain context packages · metric views · semantic views · insurance & BFSI
Phase 2B · Parallel
AgentOS · Harness
Multi-agent control plane · decision economics · governance at runtime
Play 1 · Entry CE Assessment

A 4-week fixed-price engagement that tells a CDO exactly which AI agent classes they can deploy today and what is blocking the rest. Week 2 is where the platform scan runs — using platform-native utilities from Databricks and Snowflake, depending on the customer's stack.

Week 1
Discovery
  • Target Class elicitation
  • Business intent alignment
  • Top-20 priority questions
  • Platform inventory
Week 2
8-Pillar Platform Scan
  • Data interviews (3–4 sessions)
  • Platform scan per pillar
  • Evidence collection
  • P8 governance deep-dive
↓ Platform utilities run here
Week 3
Floor Plan Analysis
  • Score all 8 pillars
  • Target Class Matrix
  • Class-gating gaps
  • Quick Win Spec draft
Week 4
Findings Delivery
  • CDO findings presentation
  • CE Maturity Score Report
  • Floor Plan one-pager
  • 90-day CE Roadmap

WEEK 2 · PLATFORM UTILITIES — run whichever matches the customer's stack

Databricks
Genie Ontology Readiness App
databricks-solutions/genie-ontology-readiness
  • Deploys into customer's workspace — read-only, no data access
  • 7-pillar scorecard: UC maturity, metadata, relationships, metric views, Genie Agents, domains, adoption
  • Outputs 0–100 score + maturity stage + prioritised action plan
  • Also runs as a Genie Code skill — no App deployment needed
  • Field-engineering open source, used by Databricks SE team
Snowflake
AI Readiness Score
/ai-readiness-score (CoCo skill · Snowsight)
  • Runs entirely on customer's Snowflake account — read-only
  • 3 dimensions: Demand Coverage · Semantic View Coverage · SV Quality
  • Scores consumption-ready tables against 4 behavioural signals
  • Outputs 0–100 score + self-contained HTML report per schema
  • Results are cached — subsequent runs complete quickly
10.6%
→ 52.1%
Same model · same data · context alone Databricks insurance walkthrough: answer accuracy moves from 10.6% to 52.1% by climbing 9 rungs of ontology context — no model change, no data change.
39
→ 82
One schema · six semantic views Snowflake internal test: schema with strong query demand but zero semantic views scored 39. After building SVs over 6 tables, score moved to 82.

Assessment outputs

CE Maturity Score (0–100) Context Gap Map 6 Agent Class readiness Class-Gating Gap Register OSI Quick Win Spec 90-day CE Roadmap
Full CE Assessment Playbook — 8 pillars, 6 agent classes, floor matrix
↓
The assessment output shapes both Phase 2 tracks
The gap register tells you which domains need ontology work first — metric views, semantic views, the 5-layer context stack. The agent class and harness gap tells you what governance infrastructure to build. Both tracks start from the same 90-day roadmap and run in parallel.
Context gaps → Ontology Accelerators Harness & governance gaps → AgentOS
Play 2A · Parallel
Ontology Accelerators

Domain context packages that make AI agents answer correctly — built in the platform's own format, reusable across engagements.

What gets built
  • Databricks metric view YAML packages (Genie ontology layers)
  • Snowflake semantic view YAML with verified example queries
  • Apache Iceberg as the cross-platform bridge — write-once, reason-anywhere
  • Domain-specific vocabulary: SKOS, OWL 2, SHACL constraints
Starting domains
  • Insurance — loss ratio variants, IFRS 17 metrics, underwriting rules
  • BFSI — risk metrics, regulatory reporting dimensions
  • Reusable per domain, not rebuilt per customer
10.6% → 52.1%
Same model, same data. Accuracy improvement is from context engineering alone — no retraining, no model swap. This is the SI's leverage point.
The 5-layer context stack
  • L1 Taxonomy — labels, synonyms, preferred terms (SKOS)
  • L2 Semantic Layer — business questions mapped to queries
  • L3 Ontology — class hierarchy, inference rules, constraints (OWL 2)
  • L4 Knowledge Graph — persisted facts and relationships (RDF)
  • L5 Context Graph — live runtime state (event streams)
Ontology reference — stack, platforms, insurance examples
Play 2B · Parallel
AgentOS · Harness

Multi-agent control plane that decides which agent acts, at what quality, latency, and cost — governance is built in, not bolted on.

What gets built
  • Agent harness: per-agent governance wrapper — RBAC, audit trail, circuit breakers
  • Orchestration layer: LangGraph / AG2 for multi-agent delegation chains
  • Tool registry and agent identity — who called what and why
  • Eval pipeline: hallucination detection, output validation, regression
Decision economics
  • Not just "does the agent run" — which path meets quality and policy thresholds at lowest cost
  • Budget enforcement at orchestrator level (cost per business outcome)
  • Model routing: pick the most economical model that clears the quality floor
  • Latency and risk trade-offs surfaced as runtime decisions, not config
4-layer AgentOS taxonomy
L1 Framework · L2 Orchestration · L3 Agent Harness (per-agent governance) · L4 Agent Management Platform. The SI builds L2 and L3; L4 is Microsoft / IBM / ServiceNow territory.
Governance requirements addressed
  • OWASP Agentic AI Top 10 — prompt injection, tool poisoning, uncontrolled iteration
  • EU AI Act Art. 9 risk management (Class 5 autonomous agents)
  • NIST AI RMF — Govern / Measure / Manage functions
  • AIUC-1 AI Usage Controls certification alignment
Accountable AI by Design — taxonomy, SI play, platform comparison

How 2A and 2B reinforce each other

The ontology accelerators produce the context packages that agents reason over. The AgentOS harness governs which agent accesses that context, at what cost, within what policy. Play 2A raises answer quality; Play 2B controls the runtime conditions under which that quality is reliably delivered. Neither is complete without the other — which is why both should start from the same gap register, in parallel.