Technical Enablement Manager — Assignment

MCP Integrations for Solution Architects

A hands-on lab designed around the 5 things that actually go wrong

The thesis

Working back from support tickets to SA enablement.
SAs don't need to learn MCP. They need to learn what breaks in MCP deployments and how to handle it in front of a customer.

What I built

A 90-minute lab where SAs experience the 5 most common MCP failure modes with their own hands — so the first time they see each one is in a safe room, not in front of a customer.

The 5 MCP Failure Modes

Every section of the lab exists to make the SA experience, diagnose, or explain one of these.

Failure Mode 1
The tool never gets called. Vague description — silent failure — nothing in the logs. Customer says "it's not working" and there's nothing to look at.
Failure Mode 2
The response blows up. No result size control — context window overflows — LLM hallucinates or truncates. Customer sees garbage.
Failure Mode 3 — the dangerous one
Everyone can see everything. Admin token in production. No error. The demo looks perfect. The permissions are completely wrong. This is a security incident, not a bug.
Failure Mode 4
It works in Claude but not in Cursor. Spec drift between hosts. Transport framing varies. Customer uses a host you haven't tested.
Failure Mode 5
"It worked yesterday." API format change — server parses old format — silently returns wrong results. No crash. No error. Just bad answers.

Lab Flow — 90 Minutes

Hook
3 min
Build It
35 min
Connect + Break
20 min
Secure It
15 min
Sell It
10 min
Hook (3 min)

"Explain MCP to a VP in two sentences. No 'protocol.'" Then name the 5 failure modes they'll experience.

Build It (35 min) — Failure Modes 1 & 2

SAs build a working MCP server from a starter template. They write the tool definition and handler. Anatomy learned through building. Encounter: vague description (silent failure) and context overflow (no size limit).

Connect + Break It (20 min) — Failure Modes 3, 4, 5

The aha moment: it works. Then deliberately break it three ways. Admin token — restricted docs appear. Host compatibility — same server, different behavior. API drift — silently empty results.

Secure It (15 min) — The money section

Grounded in the failure they just experienced. Four CISO questions, memorized by the end. Exercise: write a 3-sentence answer to "Is this secure enough for our regulated environment?"

Sell It (10 min)

One scenario done well: technical buyer evaluating MCP vs custom RAG. Peer feedback on 4 criteria. Close: "Name the 5 failure modes."

Live Demo

Working MCP server wrapping a search API. 80 lines, Node.js, zero dependencies.

Architecture
User asks question ──► MCP Host ──► MCP Server ──► Search API ──► Permissioned results ──► AI answers
Demo StepWhat HappensFailure Mode
1. The WinSearch "remote work policy" → 3 clean results, permission-aware
2. Admin TokenSwitch to admin token → CONFIDENTIAL docs appearMode 3
3. FixSwitch back to user token → restricted docs gone
4. API DriftChange API format → "No results found" (silently broken)Mode 5

→ Switch to MCP Inspector for live demo

From Support Tickets to SA Readiness

Failure LayerExampleSelf-Resolvable?
1. Tool DescriptionLLM ignores the tool — vague descriptionYes
2. Response SizeContext overflow — no pageSize limitYes
3. AuthenticationAdmin token in production — permission leakYes
4. Host CompatibilityWorks in Claude, breaks in CursorYes
5. PermissionsSource-system ACL mismatchEscalate

If SAs can self-resolve layers 1–4, engineering only sees layer 5. The exact reduction depends on internal support data — but the structural logic is clear: most MCP support tickets are field-resolvable with the right diagnostic training.

Same Knowledge, Different Surfaces

Three audiences, three adaptations of the same lab.

SectionSACSM / AIOMSupport
BuildHands-on coding"Map It" exercise — scope MCP for customer scenariosSame build — must understand code to troubleshoot
Break3 failure modesWatch demo, explain to a partner5 failure layers — trace failed requests
Secure4 CISO questionsWrite 3-sentence CISO emailPermission Escalation Lab
SellTechnical buyer pitchQBR expansion + renewal framingEscalation Decision Lab — 5 tickets, self-resolve vs escalate
ArtifactSA Readiness GuideCustomer Conversations CardDiagnostic Decision Tree Poster

Module 5 of 12

The MCP lab is one module in a complete SA onboarding curriculum. Same pattern replicates across every topic.

Weeks 1–2: Foundations
M1. Platform & Search
M2. Enterprise Graph
M3. Connectors
M4. Permissions & Security
Weeks 3–4: Extend
M5. MCP Integrations ◄
M6. Agent Builder
M7. Agent Governance
Weeks 5–8: Deploy
M8. Customer Deployment
M9. Troubleshooting & Escalation
Weeks 9–12: Expert
M10. Competitive Positioning
M11. Enterprise Architecture
M12. Product Roadmap

Each module follows the same structure

Artifact Tiers
4
micro · meso · macro · agentic
Role Variants
3
SA · CSM · Support
Modes
2
learning · performance support
Feedback
3
activity · behavior · outcome

Four KPIs This Program Moves

SA Ramp Time
Faster to first customer demo
Escalation Rate
Self-resolution patterns
30-Day Milestone
Implementation readiness
Feature Adoption Lag
Days from ship to field fluency

Measurement

LevelWhatSignal
ActivityDid they complete the lab?Completion rate
BehaviorAre SAs demoing MCP in customer calls within 30 days?CRM notes / Slack poll
OutcomeAre customers requesting MCP integrations after SA conversations?Pipeline attribution

If behavior is flat, the lab needs redesign — not more training.

What This Enables

When technical enablement is built as infrastructure, not content:

New team members ramp faster — the readiness guide tells them exactly what to know, what to demo, and what to watch for.

The demo prototype is annotated line-by-line — anyone can understand it, modify it, and present it without the original builder in the room.

The presentation framework is reusable across product areas — swap the content, keep the structure. Builder of builders.

The product keeps getting deeper, so the field has to keep getting sharper. Enablement systems should be built from the moments where the burden is highest — and scale from there.