MuleSoft AI Capabilities (Course #ai-capabilities)

Programme every line is something you do

8 modules · 20 topics · in course-flow order

1. Module 1

   Chart the AI landscape

   Objective. Position each MuleSoft AI offering against pilot criteria and produce a scored shortlist.

   • Map Einstein/Agentforce touchpoints, Anypoint AI features and AI connectors into one capability map
   • Score each offering for a claims-assistant pilot on maturity signals, data residency and exit options
2. Module 2

   Call LLMs with discipline

   Objective. Parameterize LLM calls from Mule apps and benchmark cost and latency per prompt pattern.

   • Call OpenAI, Anthropic or Azure OpenAI from flows, setting temperature, token ceilings and timeouts per use case
   • Externalize prompts as versioned assets and diff outputs across regression suites
   • Consume streaming chunks with bounded buffers (flush every 256 KB) to cut time-to-first-token
3. Module 3

   Actions agents can trust

   Objective. Expose curated Mule APIs as agent actions and bound what each action may do.

   • Register Mule APIs as Agentforce actions with typed inputs, outputs and idempotency guarantees
   • Scope each action read-only vs transactional and dry-run risky ones in a sandbox first
4. Module 4

   Data Cloud & Copilot hooks

   Objective. Assess Data Cloud and Einstein Copilot triggers for governed, auditable flow invocations.

   • Bind flows to Salesforce Data Cloud data spaces, aligning identity-resolution keys with API calls
   • Trigger Einstein Copilot actions from flow events, logging every invocation for audit trails
5. Module 5

   AI pair-programmer on trial

   Objective. Benchmark in-Anypoint AI assistance against coding standards before any team-wide adoption.

   • Grade generated code and documentation on a sample project against your review checklist
   • Measure suggestion acceptance without edits (target metric: ≥ 60%) across two sprints
   • Set guardrails: no secrets in prompts, mandatory human review on generated DataWeave
6. Module 6

   Grounded answers, cited proof

   Objective. Prototype retrieval-augmented answering over enterprise content served through Mule APIs.

   • Chain vector DB connectors — embeddings plus similarity search — behind process APIs to ground replies
   • Quantify groundedness with citation checks (target: ≥ 95% of answers citing retrieved chunks)
7. Module 7

   Guardrails before scale

   Objective. Harden the AI surface: injection defenses, privacy masking and gateway quotas.

   • Red-team prompts for injection and jailbreak vectors, logging findings in a threat register
   • Mask PII before model calls and pin processing to approved regions to satisfy residency rules
   • Enforce token quotas and content filtering at Flex Gateway (alerts at 80% and 100% of monthly budget)
8. Module 8

   Labs: the pilot stack

   Objective. Deliver the demo stack — LLM-backed API, agent action, gateway guardrails — with metrics attached.

   • Build an LLM-backed API with retries, bounded streaming and structured error responses
   • Expose it as an Agentforce action and drive a scripted conversation end to end
   • Secure the endpoint with Flex Gateway policies and capture a metrics snapshot for the pilot verdict

References & Official Documentation every claim above has a source

Module 1

• MuleSoft Agent Fabric OverviewMuleSoft Docs

Module 2

• Configuring Text Generation OperationsMuleSoft Docs
• MuleSoft Inference ConnectorMuleSoft Docs

Module 3

• Enabling an API Project for Topics and Agent ActionsMuleSoft Docs

Module 4

• Salesforce Data Cloud ConnectorMuleSoft Docs

Module 5

• Developing Integrations Using MuleSoft VibesMuleSoft Docs

Module 6

• MuleSoft Vectors ConnectorMuleSoft Docs

Module 7

• LLM Token Based Rate Limit PolicyMuleSoft Docs

More in Architecture