Internal assistant, not public chatbot
The pilot has an authenticated user, a narrow operational job and a defined permission boundary.
All builds
CONTROLLED AGENT ACTIONS
Agentforce Operations Pilot
An internal Ops Concierge that grounds answers in approved knowledge, confirms writes and hands sensitive or uncertain work to a person with context intact.
AgentforceKnowledgeFlow actionsApex actionsEvaluation scenarios
Interactive demo
Launch the demo Sample build using synthetic data in a Developer Edition org. Created to demonstrate architecture and delivery approach; not presented as client work.
The scenario
A designed problem, not a client claim.
DemoCo operators spend time finding guidance and reconstructing request context, but an unconstrained agent would introduce unacceptable access and action risk.
Before · Messy process
- Guidance is scattered across documents and team knowledge.
- Users repeat context when a request becomes a Case.
- Sensitive intents do not have one consistent escalation rule.
- An agent action could write before the user understands the change.
After · Reliable system
- The agent answers only from approved synthetic Knowledge.
- Writes require explicit user confirmation.
- Case creation reuses the same Flow and Apex process.
- Security, legal, regulated and uncertain requests escalate by design.
Interactive system demo
Watch the system make the decision.
Choose a scenario, play the full path or inspect any decision directly. This is an explainable frontend model, not a live Salesforce connection.
Choose a path4 deterministic scenarios
Live system mapGUARDED ACTION
Interactive frontend modelStep 1 / 4No live org connection
Architecture
Clear responsibilities. Explicit contracts.
01
Topic boundary
Ops Concierge handles guidance, confirmed Case creation and Case escalation only.
02
Grounding
Access Guide, Billing Policy, Integration Runbook and Security Escalation Standard are the approved sources.
03
Actions
Narrow Flow and Apex actions expose only required inputs and predictable outputs.
04
Confirmation
The user sees a concise change summary before any write action.
05
Human handoff
Sensitive or uncertain intent creates an escalation summary without autonomous resolution.
Design decisions
Why the system is shaped this way.
Reuse the automation contract
The agent creates a normal Case; Flow and Apex remain the system of execution and routing.
Escalation is a successful outcome
For sensitive or uncertain work, a context-rich human handoff is safer than forced autonomy.
Production hardening
What changes before real users depend on it.
- Evaluate confirmation, incomplete input, sensitive intent and prompt injection attempts.
- Enforce least privilege and customer-record access in the underlying action layer.
- Version agent instructions, knowledge and evaluation scenarios together.
- Review transcripts with synthetic data only and redact action diagnostics.