From Email Deflection to Verified Resolution
An Agentforce Knowledge-to-Action Experience
A CX strategy and Salesforce Agentforce implementation blueprint, presented as an interactive simulation. It is not a production deployment, and no results shown here were achieved in a live environment.
The executive challenge
Email volume is rising. “Deflection” metrics look great. Customers aren't always helped.
Support inboxes serve both external customers and internal employees. Automated replies can push deflection numbers up while repeat contacts, reopened cases and frustration quietly grow. Knowledge gaps go unnoticed because nobody connects “no good article” back to the people who write articles.
Philosophy
“Deflection does not equal delight; verified resolution matters.”
Interactive step-through
Three inbound emails, three journeys
Step 1 of 5
Inbound email arrives
A Case is created automatically from the email. Nothing has been answered yet.
Honest outcome labels
Four ways a case can end — counted separately.
AI sent a grounded response. Not yet proven to have solved anything.
AI summarised, retrieved or drafted; a human decided and sent.
Only when verified — customer confirms and no reopen or repeat contact within the window.
A person resolved the issue, with or without AI assistance.
Proposed KPIs
What I would measure
Verified AI resolution rate
AI Resolved ÷ AI-handled, confirmed and not reopened in 7 days
Repeat contact rate
Same customer + intent within 14 days
Escalation quality
% of escalations where the specialist didn't need to re-ask
Knowledge gap cycle time
Gap detected → reviewed article published
Audience-safety incidents
Target: zero internal content in external replies
CSAT by outcome label
Compare AI Resolved vs Human Resolved experiences
Proposed metrics only — no targets have been achieved or claimed.
Human-in-the-loop safeguards
Where people stay in charge
- Audience is set before retrieval; internal articles are excluded from external answers by rule.
- Below-threshold confidence never auto-replies — it triages or routes to a person.
- Frustration, repeat contact or a request for a human always escalates, with full history.
- AI-drafted articles stay DRAFT until technical review and editorial publication.
- Credits, refunds and policy exceptions require human approval.
- Every AI reply is disclosed, grounded in a cited article, and auditable on the Case.
Optional
Illustrative capacity calculator
1,600
verified AI resolutions / month
320
agent hours redirected to complex work
2.3
FTE-equivalent capacity (140 h/month)
Counts only verified resolutions — not answers sent. Hypothetical inputs; adjust to explore.
My approach
Leading with people, designing for trust.
CX leadership
Define what “resolved” means with the business first, so AI success is measured by customer outcomes rather than ticket avoidance.
Human-centered service design
Map moments of frustration and decide where automation helps, where it should step aside, and how hand-offs feel to the customer.
Operational governance
Set confidence thresholds, audience rules, knowledge ownership and review cadences so the system stays accurate as it scales.
This blueprint reflects how I would approach the strategy and operating model. It is a concept, not a description of a system I deployed.
Want to go deeper?
An optional, more detailed demo (also a simulation) opens in a new tab.