AI Action Orchestration — Agentic Execution Pipeline
End-to-end action loop: user input → intent parsing → planning → tool selection → guardrails → execution → validation → response · <800ms E2E
1 User input & intent understanding
Input channels
Chat (natural language)
Voice (transcribed)
Quick actions (tap)
Deep link (intent URL)
Widget command
Intent classification
LLM classifier (Claude)
Primary intent
Sub-intent
Confidence score
Ambiguity flag
Intent accuracy: 96%
Entity extraction
Amount ($100)
Recipient (mom, John)
Currency (USD, NGN)
Date/time (tomorrow)
Account reference
Intent types
send_money
32%
check_balance
18%
transaction_status
15%
card_control
12%
savings_action
8%
Clarification
Missing entity → ask
Ambiguous intent → confirm
Multiple matches → select
Slot filling loop
Max 2 clarifications
2 Context retrieval & action planning
User context
User profile (Redis)
Verification level
Transaction limits
Saved recipients
Preferences
Financial context
Balance check
Pending transactions
Recent activity
Credit available
Linked accounts
Conversation context
Session memory
Prior messages (5)
Unresolved intents
Pending confirmations
Error recovery state
Action planning
Plan generator (LLM)
Single-step action
Multi-step workflow
Dependency graph
Rollback points
Planning: <50ms
Tool selection
send_remittance()
get_balance()
lock_card()
transfer_savings()
get_fx_rate()
3 Pre-execution guardrails & validation
Permission checks
User authenticated ✓
Action allowed (RBAC)
Verification level OK
Feature flag enabled
Rate limit check
Financial guardrails
Amount < daily limit
Sufficient balance
Recipient validated
Corridor enabled
No duplicate txn
Compliance checks
Risk score (ML)
Sanctions screening
PEP check
Velocity rules
Amount threshold
Block rate: 0.3%
Confirmation flow
Low risk → auto-proceed
Med risk → confirm
High risk → 2FA + confirm
Show summary card
Await user approval
4 Action execution & API orchestration
Execution engine
Action runner
Idempotency key
Transaction boundary
Timeout handling
Circuit breaker
API calls
Internal APIs
Wallet service
Remittance service
Card service
Partner APIs
State management
Create ledger entry
Update balances
Record audit trail
Emit events (Kafka)
Update session
Error handling
Retry (3x backoff)
Fallback provider
Partial rollback
Queue for later
Escalate to human
Async actions
Job queued (Bull)
Progress tracking
Push notification
Status polling
Webhook callback
Exec time: <200ms
5 Response generation & feedback loop
Result validation
Action succeeded
Expected state verified
Side effects confirmed
Consistency check
Receipt generated
Response generation
LLM response (Claude)
Natural language
Action confirmation
Next steps suggested
Rich cards (UI)
User feedback
Thumbs up/down
Correction capture
Follow-up question
Implicit signal (proceed)
Explicit rating
Learning pipeline
Log to data lake
Intent fine-tuning
Entity model update
Success rate tracking
A/B test outcomes
Metrics
E2E latency
680ms
Success rate
94.2%
1st try resolve
87%
CSAT
4.6
Actions/day
48K
Available tools & capabilities
Money movement
send_remittance
internal_transfer
schedule_payment
cancel_pending
Account
get_balance
get_transactions
update_profile
add_recipient
Cards
lock_card
unlock_card
set_spending_limit
get_card_txns
Savings
create_goal
deposit_savings
withdraw_savings
get_savings_balance
Credit
get_credit_score
request_loan
make_payment
get_credit_history
Info & Support
get_fx_rate
get_fees
faq_lookup
escalate_to_human