"Look after this money for me" is not a mandate that can be executed: it can be read as dozens of entirely different courses of action. What is handed to an agent should be a mandate that states at least four things: what the goal is and how it is measured; what must not be done; what to do in situations the mandate does not cover; and when the mandate needs to be reviewed. The goal should be measurable, but measurable goals get gamed: if the only test is whether the benchmark is beaten, an agent may quietly take on risks no one can see. Nor can a mandate ever be complete; situations it does not cover need a default rule: act conservatively, and stop to ask.
Professional asset management sets out the delegation relationship in an investment policy statement: the investor's return objective and risk objective, together with five kinds of constraint covering liquidity, time horizon, taxes, legal matters and personal preferences (Maginn et al., 2007; CFA Institute, 2010). The same holds when handing work to an AI agent. "Look after this money" could mean buying a money market fund, or holding a handful of growth stocks in concentration. Only when the goal, the constraints, the exceptions and the review are spelled out one by one do the feasible courses of action narrow to a few, and only then do the agent's choices have a basis on which they can be checked.
"Steady growth" is a wish, not a goal. It could mean a maximum drawdown of no more than 10%, or annualized volatility of no more than 8%, or never losing principal under any circumstances; the three measures lead to very different allocations. A goal must state both the return side and the risk side, and give the benchmark and the horizon for measuring it: over what period, compared with what, and by which measure success is judged. Once the measure is fixed, the agent's behaviour will optimize toward it.
One wish, three ways to measure it
The principal asks for "steady growth". Choose a measure and see the allocation the agent derives from it (illustrative).
Constraints mark out the range within which the agent may choose. Liquidity constraints say when how much money will be needed. Time-horizon constraints say what the money is ultimately for and when. Tax constraints affect when to sell and in which account to hold. Legal and regulatory constraints set out the instruments that may be held and the ways they may be traded. Personal preferences include industries and companies the user does not want to hold. Each added constraint removes some feasible portfolios; the more clearly the constraints are written, the less likely the agent is to make a choice the user cannot accept after the fact.
Each added constraint leaves fewer portfolios to choose from
Each dot is a candidate portfolio; the horizontal axis is risk and the vertical axis expected return (illustrative). Tick the principal's constraints one by one and see which portfolios remain.
Goodhart (1975), discussing monetary policy, observed that once a statistical indicator is used as a target for control, it no longer reliably reflects what it was meant to measure. AI researchers call this phenomenon "reward hacking" (Amodei et al., 2016): a system finds a way to improve the indicator that defeats its designers' intent. Finance has a ready-made example. Lo (2001) constructed a hypothetical fund that continually sells out-of-the-money put options: in calm years its returns are steady and beat the market, but when the market falls sharply it loses most of its capital. If the mandate says only "beat the benchmark", such a strategy fully meets the requirement. A goal must be paired with risk constraints, and the constraints must cover tail risks that cannot be seen.
No mandate, however detailed, can foresee every situation. Grossman and Hart (1986) called this the incompleteness of contracts: since not every situation can be specified in advance, what matters is who decides in the situations not covered, and by what rule. For an agent, a reasonable default rule is this: when a situation falls outside the mandate, first take the conservative course, such as pausing new trades and holding the current position; then put the situation and the available options to the principal for decision. Acting entirely on the agent's own judgment, and silently stopping to wait, are both poor defaults.
A mandate is written at a particular moment, according to the circumstances of that moment. Changes in income or family, or a goal drawing near, can make the original goals and constraints no longer apply; large moves in the market can also carry the original allocation far from what was agreed. Professional practice usually reviews the investment policy statement once a year, and at once when a major event occurs. An agent can do this more finely: it can keep checking whether the premises of the mandate still hold, and when they change, remind the principal to review, rather than continuing to execute a mandate that has already expired.
Treat the mandate as an object in the product: it has goals, constraints, exception rules and a period of validity, and every step the agent takes can be traced to one of its clauses.
Goals, constraints, exception rules and period of validity are each visible and editable, and every step the agent takes can be traced to one of them.
Translate wishes such as "steady" into measures that can be checked, and ask the principal to confirm them.
Test strategies under stress scenarios, not just against the benchmark.
In situations not covered, pause and put forward options; remind the principal to review when income, family or goals change.