Agentic payments. Bounded authority for software that moves money.
An AI agent that can instruct a payment is a delegate, and A8 Core™ treats it as one. Its authority is a grant on an account with a scope, thresholds, count limits, approval modes, a term, and an evidenced decision behind every action — enforced in the account system, not in the application that hosts the agent.
Permissions
Authority vocabulary collapses into three tiers, and an agent is granted only the middle one.
Know
Read balances, positions, history, and status for the client, account, or single position the grant names — nothing outside it.
Act
Submit and execute within the grant: create a quote, register a destination for approval, instruct a transfer or payment under the limits below.
Decide — never delegable
Approve, sign, grant, and authorise belong to the grantor or a named human approver. An agent cannot approve its own instruction, and cannot be granted the power to.
Limits
Value and count, per instruction and per window.
Per-instruction threshold
The maximum any single instruction may carry before it requires a human decision.
Aggregate windows
Rolling value caps over a day, a week, or a month, so many small instructions are bounded as one.
Transaction-count caps
A limit on the number of instructions per window, independent of value.
Destination scope
Known destinations only, or a registered allowlist. A new destination is an approval event regardless of amount.
Approvals
Four modes, chosen per grant and per action.
Auto below threshold
Low-value, known-destination instructions proceed after policy evaluation and are recorded as such.
Approval above threshold
The instruction waits as a pending request with the full decision context attached.
Dual and m-of-n approval
Two named approvers, or m of n, before release — the multi-party primitive B5 Secure™ carries on its roadmap.
Always on a new destination
The first instruction to any destination requires a human decision, with an optional cooling period before it may be used unattended.
Revocation, and the instruction already moving
The question most agent-payment designs leave unanswered.
Revocation is effective the moment it is made. An instruction already in flight fails by default; it completes only if the grantor elected AllowCompletion in advance, and that election is part of the grant, not a runtime flag. Suspension is available as a lighter action: the agent’s grants pause, nothing new is accepted, and nothing is lost. The full model is on the Investor Passport page.
Evidence
Every action leaves a decision record, not a log line.
The decision record
Agent identity, the grant it acted under, action, amount, source, destination, the stated reason, the policy evaluation result, and the approval outcome, with timestamps.
Per-action signature
Each instruction is signed by the agent’s own key at the moment of action, and the signature is captured in the record.
Portable format
Decisions are recorded in the open B5 Secure™ authority evidence shape, so an examiner reads them without this platform.
Linked movement
Every record links to the quote, directive, or transaction it produced, so the money and the decision are never separated.
Interface
What the specification commits the API to.
Quote, then execute
Two-phase instructions: a locked quote with fees and rates, then an execution that references it.
Idempotency keys
Every mutating call carries an idempotency key. A retry is the same instruction.
Signed events
Status changes are delivered as signed webhooks with retry and duplicate suppression by delivery identifier.
Status. Agentic payments is a product specification built on the Investor Passport delegation model. Nothing on this page is generally available today, and availability for any institution depends on that institution’s own authorities, contracts, integrations, and policies. See product status for how A8 Core™ labels capability maturity.
Build on the account operating system.
See how bounded agent authority fits your institution’s account, custody, and operations model.