PrymeIntelligence

Banking and fintech·Playbook

How to let an AI agent stage trades for a person to release

Let an AI agent stage trades for a person to release: a stage-only order system role, pre-trade checks attached, and expiry on stale tickets.

The right role for the agent in your order management system creates staged orders and nothing else. It reads positions and mandates, proposes trades, runs the pre-trade checks and stages tickets with the rationale attached. Routing to market stays with a named person. Staged tickets expire if prices move or time passes, and every step is filed.

What does staging a trade mean, and who owns it?

A staged order sits in the order management system with every field filled in, but it has not been sent to a broker or venue. Someone must release it before it can execute. Most order systems support this as a workflow state, and many teams already use it between portfolio managers and dealers.

The portfolio manager or treasury lead owns the decision to trade. The dealing desk owns execution. Compliance owns the restricted list and mandate rules. An agent slots in before all three: it prepares, and they decide. The idea is the same one that governs payments, where agents prepare payments and people release them.

How do you set it up, step by step?

  1. Person: defines the stage-only role. The order system role can create, amend and cancel its own staged orders. It cannot release, route or amend orders created by anyone else.
  2. Agent, depth 1: reads the book. Positions, cash, mandate limits, model weights, the restricted list and current prices.
  3. Agent, depth 2: proposes trades. For a rebalance, a cash flow or a breach, it calculates the orders needed and writes the reason for each.
  4. Agent, depth 1: runs pre-trade checks. Mandate limits, concentration, restricted list, cash availability. Each result is attached to the order.
  5. Agent, depth 2: stages the tickets. A staged order is a draft in the system: written there, but going nowhere until a person acts.
  6. Person: releases, amends or rejects. A named person reviews the batch and routes what they approve.
  7. Agent, depth 3: cancels its own stale tickets. When the expiry rule fires, the agent cancels the staged ticket and records why.
Step The agent does Depth Stays with a person
Role design Nothing None Deciding the role's reach
Read book Reads positions, limits, prices 1 None
Propose Calculates orders with reasons 2 Whether to trade at all
Pre-trade checks Runs and attaches results 1 Overriding a failed check
Stage Creates staged tickets 2 Release
Release Nothing None Routing to market
Expire Cancels its own stale tickets 3 Re-staging decisions

What does "release" mean here?

Release is the act that makes the trade real. Once an order is routed and filled, it cannot be called back; it can only be reversed with another trade at a different price. That is depth 4 territory, and in this design no agent holds it.

The release must be made by a person who is not the agent, under their own login, so the order system's audit records a human release. If your system allows the staging user to release, fix that in the role. Leaving it to the agent's brief is not enough.

Which controls apply, and what evidence do they leave?

Stage-only credential. The role cannot route. Evidence: the role definition, and refused verdicts for any attempt to call a release function.

Attached checks. Every staged ticket carries its check results. The releasing person sees a failed check before they can route, and any override is recorded with their name.

Provenance on prices and positions. Positions from your own books are sealed. Prices from a market data feed are external. If an order rests on an external price, the ticket shows it, and the releasing person can refresh before routing.

Expiry. A staged ticket expires after a set time or a set price move. Without expiry, a person may release a ticket that was right in the morning and wrong by the afternoon.

No splitting. If release thresholds depend on order size, the agent's bounds must test the parent order. Splitting a large order into small tickets to fall under a threshold is refused.

What should you measure?

  • Time from trigger to staged batch. The agent's main contribution.
  • Tickets released unchanged, amended or rejected. Rejections with reasons are your best training signal for the brief.
  • Tickets expired before release. Many expiries mean staging is running too far ahead of the people who release.
  • Failed pre-trade checks found at staging versus at routing. The agent should catch them first.
  • Release attempts by the agent's role. These should all be refused; any other outcome is an incident.

Where does it usually go wrong?

Rubber-stamp release. A person approving forty tickets in a minute is not reviewing them. Present the batch with the reasons and the checks first, and watch release speed as a warning sign.

The staging role can release. A configuration slip, often inherited from a dealer's role template. Test it before go-live by attempting a release and checking it is refused.

Stale rationale. The reason written at staging no longer holds at release. Expiry rules fix most of this.

Restricted list lag. The agent reads a copy that is hours old. Read the list from source at staging and again at release.

The same split between preparing and releasing runs through preparing outbound international wires for release. For the general limits on bank agents, see AI agents in banking.

Where Pryme fits

In Pryme, staging is certified as its own capability through Attest, replayed on your own past rebalances and breach cases. Release is not granted to any agent in this design, so the minimum chain refuses it at the role scope, and the verdict names that bound. Each ticket's facts carry their provenance class, and every verdict is filed in the append-only record.

Questions people ask

Can the agent stage trades for client portfolios as well as house accounts?

Yes, if the mandate data is in reach and the capability is certified on client cases. Suitability and mandate checks must run for each portfolio, not once for the batch.

What if the releasing person is unavailable?

Tickets wait, then expire. That is the intended outcome. A staged order that executes without a person defeats the point of staging.

Can the agent cancel orders a person has already released?

No. Once released, the order belongs to the dealing desk. The agent may flag a problem; a person decides whether to cancel.

Decide what your agents may do. Prove it. Enforce it.

Pryme Intelligence certifies each agent on your own cases, checks every action against five bounds at the moment it happens, and files every verdict.