Outcome, budget and expiry
Agentic finance · Payment controls
Before the Agent Pays.
A four-transaction simulation for teams deciding when an AI agent may pay, when a human must step in and when the system must stop.
Interactive prototype
01
The performance gap
Most agent payment flows offer two controls.
Approval on every purchase stalls the task. Broad account access turns one bad decision into a financial incident.
Define the authority before the agent reaches checkout.
87%
concerned about agent accuracy
81%
concerned about security and privacy
17%
report improved team collaboration
Evidence base: Stack Overflow Developer Survey 2025 · Forrester Agentic Payments Research 2026
Audience
Engineering managers, AI product owners and operations leads.
Performance goal
Given a transaction and a temporary mandate, choose allow, escalate or deny.
Design constraint
Safety cannot depend on constant human attention or model self-restraint.
02
How the system works
Every payment passes through the mandate.
The decision gate checks purpose, merchant, amount and evidence. It then executes, escalates or blocks the transaction.
bounded brief
Merchant, category and limits
evaluate
Context + deterministic checks
route
Inside mandate
Material exception
Policy breach
EVIDENCE LOOP
ReceiptWhat moved
→Audit logWhy it moved
→ReviewWhat to change
→Earned autonomyLimits change after review
03
The decision model
MANDATE
The agent receives temporary authority tied to one outcome, budget and expiry.
Mission
What outcome is authorised?
Authority
What may the agent do?
Narrow constraints
Which limits are deterministic?
Decision thresholds
When must a human decide?
Audit evidence
What proves the action?
Termination
How is access revoked?
Earned autonomy
How does trust increase?
04
Interactive prototype
Make four payment decisions under one live mandate.
The incident changes after every choice. Use keys 1–3 or select a command.
→ mandate inspect --agent ops-07
MISSIONRestore checkout within 30 minutes
AUTHORITYPurchase incident infrastructure
BUDGETEUR 300 / incident
ESCALATENew merchant or spend > EUR 120
05
Decision architecture
The framework appears after the first decision.
The first screen asks for a payment decision. Feedback exposes the cost, then the team applies MANDATE to a new transaction.
Cold decision
Choose before seeing the framework.
Consequence
See what the choice costs or exposes.
MANDATE
Identify the missing rule.
Second attempt
Apply the rule to a new transaction.
Transfer
Write a mandate for an agent at work.
06
Decision standard
A valid mandate must hold under a new scenario.
A mandate passes when it handles a new scenario and produces the correct payment decisions.
01Task completion
02Financial exposure
03Control efficiency
04Auditability
05Recovery design
Critical-failure rule
Permanent credentials, undefined budgets or payments without proof of intent fail the safety review, regardless of the average score.
07
Pilot measurement
The pilot tracks decisions, spend and review time.
Fewerlow-value human approvals
Higherpolicy-aligned execution rate
Lowervariance between budget and spend
Fasterreconstruction of financial decisions
My contribution
I turned an agent-payment risk into a testable decision system.
- Research synthesis and performance-gap analysis
- Task model and scenario architecture
- Interaction design and functional prototype
- Decision rubric and pilot measurement plan