Questions worth answering before an agent can spend.

Agentic commerce is a purchasing journey in which AI agents can discover options, make selections, and initiate or complete defined actions for a buyer. The buyer remains the source of authority.

No. Conversation may capture intent, but agentic commerce also involves identity, structured actions, merchant communication, authorization, payment infrastructure, and evidence.

It can operate within pre-authorized boundaries if the operator intentionally grants that scope. Requests outside those boundaries should stop for review or be declined.

The accountable person or business cannot delegate responsibility merely by using an agent. A production system should preserve who granted authority, which rules applied, and who resolved exceptions.

Mandate evaluates simulated agent purchase requests against deterministic financial rules and returns APPROVED, APPROVAL_REQUIRED, or DECLINED with machine-readable reasons.

No. The current MVP stores authorization intent, simulated requests, decisions, approvals, and audit events. It does not move money or store card or bank credentials.

A spending mandate is a versioned set of rules defining what one agent may spend, where, when, and under which review conditions.

Least privilege requires permissions to belong to a specific identity and purpose. A travel agent should not automatically inherit the authority of a procurement agent.

The request passes every hard policy check and no policy or risk condition requires human review.

The request remains eligible but crosses a review condition, such as an approval threshold, new merchant rule, or elevated risk factor.

A hard boundary fails—for example, a revoked agent, expired mandate, exhausted budget, blocked merchant or category, disallowed country, or amount above a hard maximum.

Not through the approval queue. A decline represents a hard policy failure. The operator must intentionally change the governing policy or request facts and submit a new request.

No. The initial APPROVAL_REQUIRED result remains part of the evidence. Human approval adds a separate resolution record.

No. An LLM may convert natural-language instructions into a proposed structured mandate. A human reviews that proposal, and deterministic code evaluates transactions.

Natural-language instructions are treated as untrusted data. The model has no payment tools, approval authority, database access, or credentials; its structured output is schema-validated before human review.

It is a 0–100 indicator derived from recorded factors such as amount anomaly, merchant novelty, category mismatch, velocity, and geography. It supports review rather than replacing policy.

No. A hard policy failure remains DECLINED regardless of the risk score.

It lets an integration identify one agent when submitting structured authorization requests. The key cannot expand its own permissions or approve requests.

The credential is scoped to that agent’s identity. This prevents one integration from silently acting with another agent’s budget or policy.

Architecturally yes: an MCP tool could call Mandate’s authorization API before a purchase action. The MVP does not claim a packaged MCP connector or external production integration.

A customer backend, agent runtime, MCP server, or simulator submits merchant, amount, category, country, and idempotency information. Mandate is the system of record for the resulting authorization evidence.

No. Card numbers, CVVs, bank credentials, and real payment-provider secrets are outside the MVP’s scope.

It is hash-linked and designed to make changes evident, but the portfolio MVP does not claim certified immutability. Production maturity requires independent exports, retention controls, and monitored integrity checks.

No. Mandate is currently a simulated authorization and risk-control application, not a bank, money transmitter, card network, or payment processor.