Use cases / Finance and operations
AI Procurement
Procurement speed with limits enforced while the agent works.
01
Problem
Buyers compare suppliers and contracts by hand. An agent that can approve its own spend, or read unrelated employee data, is not acceptable.
02
AI capability
Read supplier information, analyze contracts, compare pricing and prepare recommendations and draft purchase orders.
03
Data
The agent works from the context you connect, not from the open internet by default.
- Approved supplier master data
- Contracts
- Price lists and past orders
- Budget and approval thresholds
04
Agent
AI Procurement Agent
Prepares purchase recommendations and draft POs; people approve what matters.
05
Workflow
- 1. Purchase request
- 2. Read supplier info
- 3. Analyze contracts
- 4. Compare pricing
- 5. Prepare recommendation
- 6. Create draft PO
- 7. Approval if above threshold
- 8. Outcome logged
06
Governance
Governance runs inside the agent at runtime: policy changes what it can actually do.
Can
- Read approved supplier info
- Analyze contracts
- Compare pricing
- Prepare purchase recommendations
- Create draft POs
Cannot
- Access unrelated employee data
- Approve its own high-value transaction
- Make payments
- Modify restricted records
Requires approval
- Purchases above threshold
- Contractual changes
- Sensitive external comms
Records
Identity, Data accessed, Model used, Output, Tools called, Policy applied, Decision, Approval, Action, Outcome.
Suggested starting autonomy: L1 Assist
Start at L1 (Assist), then move to L3 as evidence builds. See the worked example on the Controlled Autonomy page.
07
Outcome
What you measure, so the agent earns more autonomy on evidence:
- Cycle time from request to approved PO
- Share of recommendations accepted
- Contract issues caught before signature
- Spend decisions with a complete record
Questions
Can the procurement agent pay a supplier?
No. Payments are a restricted action. Purchases above the threshold also require human approval.
How do we increase its authority?
Progressively, through the Controlled Autonomy Model, using its recorded approval history as evidence.