Authority determines how far an agent may act without asking. Authority is not a switch but a ladder: from providing information only, to drafting a plan, confirming each trade, approving a plan, allowing a veto within a time limit, and finally acting on its own within an agreed scope. Which rung an operation can be placed on depends first on whether its consequences can be undone and whether its losses are capped. A financial account holds a special kind of risk: if it is not handled before a certain deadline, the loss becomes permanent, as with option expiry, forced liquidation of a margin position, and the deadlines of rights issues and tender offers. This irreversible exposure is exactly where an ongoing agent can help most, and exactly where authority must be tightest. The limits of authority must be enforced by the system at the execution layer, not left to the agent's own restraint; authority rises step by step with verifiable records and is withdrawn promptly after errors and changes.
Sheridan and Verplank (1978) divided the work between people and automated systems into ten levels. The version compiled by Parasuraman, Sheridan and Wickens (2000) runs from "the human makes all decisions" to "the system decides and acts on its own, ignoring the human". They further pointed out that the same system can sit at different levels in each of four stages: acquiring information, analysis, decision and execution. For financial agents, six levels are in common use: provide information only; draft a plan for the person to execute; execute after the person confirms each trade; the person approves a plan and the agent executes within it; the agent proposes an operation, which the person may veto within a set time, failing which it is executed; act on its own within an agreed scope and report afterwards. The higher the level, the less attention the user spends and the larger the exposure the user bears.
Which level an operation can be placed at depends first on two things: whether an error can be undone, and whether the loss is capped. Producing a summary or setting a reminder can be redone if wrong, at almost no loss. Adjusting a watchlist or changing a reminder's conditions can be reverted at any time. A scheduled regular investment has a limited amount per instalment, and an error can be corrected in the next period. Selling a holding, transferring funds out or accepting an offer cannot be undone once executed, and the loss may be large. Operations that are reversible and capped can be placed at high levels; operations that are irreversible or uncapped must stay at per-trade confirmation or lower.
Reversibility and loss caps set the ceiling on authority
Across: whether an error can be undone. Down: whether the loss is capped. Each cell shows the highest level of authority that can be granted. Tap an operation for an explanation.
Tap an operation.
An account holds a kind of risk that, if not handled before a certain deadline, turns into a permanent loss. An option loses all its time value after expiry, and an in-the-money option not handled according to the rules may be exercised automatically or lapse. When a margin account falls below its maintenance margin requirement, the broker issues a margin call, and if it is not met in time the position is liquidated; FINRA Rule 4210 sets the minimum maintenance margin. Rights issues and tender offers have clear deadlines, and missing them means accepting the default outcome. Some leveraged products are terminated early when they hit their call price. What these exposures share is a deadline, a clear trigger and an irreversible consequence. An agent's value here is to spot them in time, give advance reminders and explain the options; the step that actually changes holdings must be confirmed by the person.
Writing the limits of authority into the agent's instructions does not guarantee that they will be observed: the agent may misunderstand them, may be misled by content in its inputs, or may simply make a mistake. An old principle of computer security is that every program should hold only the least privilege needed for its task, and every access should be checked (Saltzer and Schroeder, 1975). For a financial agent, limits on instruments, per-transaction amount, cumulative amount, trading frequency and time window should be checked trade by trade by an execution layer independent of the agent; instructions that cross a limit are rejected outright, and a record is kept. The agent's instructions can help it make fewer mistakes; the execution layer's checks ensure that even when it does make one, it cannot cross the limits.
The level of authority need not be fixed once and for all. Like trust, it can be adjusted as evidence accumulates: the agent first runs at a low level on matters whose consequences are reversible and whose amounts are small, building a verifiable record of performance; once the record is sufficient, it moves up step by step to plan approval or time-limited veto. Authority must also be withdrawable promptly: after a single overreach or serious error, the level drops at once; when the agent changes model version, or the principal's circumstances change, authority must be confirmed again. Users should be able to see the current level of authority at all times and withdraw it with one tap.
Tap an event in the chart.
Per-trade confirmation is safest, but every confirmation takes the user's attention; full automation is easiest, but carries the most risk. Sheridan and Verplank's scale has a level between the two: the computer selects an action and informs the human early enough to allow time to stop it. Parasuraman et al. (2000) summarize it as "allows the human a restricted time to veto". The system proposes an operation and gives the person a set period in which to veto it; if the person does not object, it is executed. This level suits operations that have a deadline but whose consequences stay within an agreed range, such as rebalancing according to a plan. The length of the veto window is a trade-off: too short, and the user has no time to look; too long, and the deadline may be missed. As the deadline of an irreversible exposure approaches, the operation should return to per-trade confirmation, and reminders should become more insistent.
How long should the veto window be?
The agent proposes an operation with 24 hours to go before the deadline. On average the principal sees the message about 6 hours later. Drag the length of the veto window to see the probability that the principal sees it in time, and the risk of missing the deadline.
Assign authority by reversibility and loss cap; for irreversible exposure, remind in time and have the person confirm; enforce limits at the execution layer, raising authority with the record and withdrawing it on change.
Information, drafting, per-trade confirmation, plan approval, time-limited veto and autonomy within scope are set operation by operation, not by a single master switch.
Selling holdings, transferring funds, accepting offers and exercising options are confirmed by the person; small reversible matters can be handed to the agent.
For option expiry, margin calls, rights issue deadlines and the like, remind several times in advance; the step that changes holdings is taken by the person.
Check instrument, amount, total and authority for every trade and keep a record; lower the level after overreach, version updates and changes in circumstances, and allow withdrawal with one tap.