Index/ CSM/ Productivity to Agentic
Customer Success · the arc

Productivity today. Agentic next.

The platform shipped the productivity layer — AI that drafts. The next layer is AI that acts. This is the arc between them.

Today the human is the integration, the action, and the guardrail — all at once. Remove the human, and all three become code. That’s the job.

ShippedAI drafts,
human acts
AI drafts, human acts — Q Business ingests four systems and drafts follow-ups, health checks, renewals, email. Every output is a draft a human reviews and sends. Jan 2024 → Oct 2024 · 30+ custom Q Apps · 18,300+ hrs/year reported.
The gapThe human is
the seam
The human is the seam — it retrieves and generates. It doesn’t execute, decide, or safely fail — the human does all three. Last mile is manual · no guardrail but a human eyeball · measured in hours, not outcomes.
The visionAI acts,
human judges
AI acts, human judges — agents do the routine work end-to-end. The human moves above the loop — gate, exception, relationship. Orchestration + MCP · HITL guardrails + eval · outcome, not hours.

Not “remove the human.” Move the human up.

01

The productivity layer — what shipped

The
architecture
A Chat UI on Amazon Q Business, four sources into one context layer. S3 · Redshift · SharePoint · Totango · 30+ custom Q Apps. AWS case study (co-authored)
The
use cases
All drafting and summarizing — none of it autonomous. Post-meeting follow-ups (summary + action items) · Health Check analysis · Renewal justifications · Personalized email drafts. AWS case study
Reported
results
Self-reported, unaudited. Say “reported,” never “audited.” 18,300+ hrs/year · 40% admin gain · 12 hrs/week/consultant baseline.
The stated
roadmap
The next move is named in the case study — and it’s not built yet. Salesforce, Confluence, Smartsheet connectors · Agentic orchestration (Totango + Orchestrator). agentic = planned · AWS case study
02

Where it stops — the bottlenecks

The last mile
is manual
A Q App drafts the renewal justification; a consultant hand-copies it into the CRM. Generation is automated. The action isn’t.
Data answers,
can’t act
Four systems in, Salesforce still unconnected. Worse: cross-system identity unresolved — is the Totango account the Salesforce one? Unsafe to act until IDs reconcile.
Q Business
retrieves
Q Business retrieves, doesn’t execute — RAG answers and drafts. It has no action layer — hence Quick Automate + UnifyApps + MCP.
The human is
the only guardrail
A draft gets an eyeball before it ships, so eval and drift-monitoring are implicit. Remove the human and you need guardrails, eval, rollback, audit — none built yet.
Decentralized
= quality long tail
The Program Champion network (150+ workflows, 50+ agents) drives adoption, not shared eval or security. Someone has to fold it into a governed layer.
Hours,
not outcomes
Everything reported is hours saved. Gartner: productivity alone can’t fund agentic AI. The missing metric is revenue — renewal, retention, upsell.
03

The work — what the role actually does

1Orchestration
Turn Q Business answers into governed executions across Salesforce, Totango, Clari, Gong via Quick Automate + UnifyApps.
2Identity +
data quality
Make the five-plus systems share one account ID so the agent acts on the right record.
3Guardrails +
eval
Where the human sits, approval gates, rollback, drift detection, grounding checks.
4MCP breadth
Expand the tool surface the agent can reach.
5Outcome
measurement
Revenue and retention attribution, not another hours-saved dashboard.
6Rationalize
the long tail
Fold 150+ ad-hoc workflows into a governed orchestration layer.

The edge. The CSM Health & Renewals Agent already encodes #3 and #6. Deterministic engine decides, model narrates, human gates, outcome tracked. The pattern exists — not guessed at.

04

The process — the agentic loop

Signals are ingested on one account ID, scored by code, drafted by the model, and gated by a human above the loop. Low-risk work auto-flies; high-risk work waits on approval. Every action is tracked by outcome, and an eval loop feeds the whole chain.

Data Decides Narrates Human / outcome
Ingestone account ID Scorecode decides Proposeagent narrates Gatehuman above loop Auto-flylow-risk Human approveshigh-risk Trackoutcome, not hours signals flag draft low-risk high-risk execute approve eval loop

Through-line. Risk math is code. The model narrates. The human gates. That’s the guardrail pattern Q Business never built — and the exact shape of the CSM Health & Renewals Agent.

05

Sources

Sources
AWS case study · Amazon Quick (Q Business)