Work handed to an agent still has to be explained and checked by the principal. If checking a result takes a quarter of the time it would take to do the work oneself, the amplification factor cannot exceed 4 times. The tasks best handed over are therefore those that are much cheaper to check than to produce: code can be tested, calculations can be rerun, and citations can be opened at the source. Research conclusions and investment judgements, by contrast, take almost as much effort to check as to redo. There are three ways to lower the cost of checking: write down the acceptance criteria in advance, change the unit of display from long reasoning to key actions and evidence, and allocate checking effort according to risk. Checking also has a hidden weakness: when results are mostly right, the people checking gradually grow lax.
After handing work over, the principal still has to spend effort on two things: making clear what is wanted, and confirming that the result returned is correct. Even if the agent finished instantly and never erred, neither could be avoided. If checking a result takes r times the time it would take to do the work oneself, the amplification factor cannot exceed 1 ÷ r. In the GDPval evaluation, Patwardhan and colleagues (2025) had experts from various industries review model outputs; review took on average about a quarter of the time the experts needed to do the work themselves. On this alone, the ceiling on the amplification factor is only about 3.7 times. When results are hard to verify, the principal has only two choices. Check carefully, and the checking time enters the denominator, pulling the factor down towards 1. Or do not check, and the factor looks high, but the quality risk is not priced in, and when something goes wrong, trust collapses with it.
Checking time caps the amplification factor
Take a task that would take 100 minutes to do oneself, and assume the agent finishes instantly and never errs. Drag the checking time as a share of the time to do it oneself, to see the ceiling on the amplification factor.
How much a principal is willing to hand over depends on the judgement of whether the agent can deliver and will act diligently. What determines that judgement is not the value of the task but two things: how serious the consequences of an error are, and whether they can be reversed; and whether a single result can be judged right or wrong in time. The author judged each of 332 work activities on these two dimensions (author's estimate), which yields four zones. The hands-off zone holds the most activities, 110, accounting for about 39% of working time: consequences can mostly be remedied, results can be checked on the spot or item by item, and trust builds quickly through a large number of verifiable deliveries. Designing computer systems and applications and troubleshooting computer faults are here, since their results can be tested on the spot. The item-by-item zone has 69 activities, about 21%: consequences are larger, but results can be checked before they take effect, and authority can be extended step by step as the delivery record grows. The process-evidence zone has 41 activities, about 11%: consequences are larger, and results take a long time to judge. The remaining perception zone accounts for about 12%: consequences are lighter, but right and wrong are hard to judge, so trust rests on perception and is easily overestimated. Activities that the two coders placed on different sides of a boundary account for about 17% and are not counted in any zone.
Finance-related activities span these zones. Analysing market and industry conditions has lighter consequences, but its correctness takes a long time to judge, so it falls in the perception zone. Monitoring financial data and activities can be checked item by item. Executing financial transactions has serious, irreversible consequences, but the elements of an order can be judged on the spot before it takes effect, so it falls in a corner of the item-by-item zone. Providing financial advice and making financial plans fall in the process-evidence zone. Trust is therefore built in sequence: start by accumulating a delivery record with lower-consequence analysis and monitoring; confirm trade execution item by item before it takes effect; then move into advice and discretionary management, raising authority step by step on the basis of prior agreement, process evidence recording what each action was based on at the time, and aligned interests. This matches the division in trust theory: Mayer, Davis and Schoorman (1995) divide trustworthiness into ability, benevolence and integrity. When results can be verified, ability can be tested directly; when they are hard to verify, the principal can only judge by integrity (whether the agreement is kept) and benevolence (whether interests are aligned).
For some tasks, finding the answer is hard but confirming it is easy: solving a sudoku takes repeated trial, while checking a completed one takes only a pass through each row and column. Computer science describes such problems through the gap in difficulty between verifying and solving (Cook, 1971). What an agent can amplify most is exactly this kind of task, where checking is far cheaper than producing: code can be tested, numbers can be recomputed by a program, and citations can be opened and compared with the source. By contrast, checking whether a research conclusion is sound or an investment judgement is right often takes as much effort as redoing it, and investment results take a long time to emerge. The same agent achieves a far higher amplification factor on the first kind of task than on the second.
Select an item.
The effort of explaining is also in the denominator, and the cost of an unclear explanation multiplies later. Boehm and Basili (2001) summarized the experience of software engineering: fixing a requirements error found after delivery often costs tens or even hundreds of times as much as fixing it at the requirements stage. Handing work to an agent is the same. Say only "Take a look at how this fund is doing", and the result is likely not what was wanted, needing several rounds back and forth. Write down in advance what to compare it with, the period, the measures and what counts as done, and an acceptable result can come back in one go, with a clear checklist for checking.
An agent can show its reasoning, but long passages of reasoning are hard to check: they take a long time to read, and they may not reflect what the agent actually relied on. In experiments by Turpin and colleagues (2023), when a prompt implied a certain bias, models were swayed by it, yet the reasoning they wrote almost never mentioned it (only 1 of 426 explanations did). A unit that is easier to check is key actions and evidence: which data were queried, where they came from and as of when; which calculations were done and what the results were; what was changed and why. Each item can be opened and verified on its own, and the checker needs to examine only the few points relevant to the conclusion.
Not every result needs to be checked with the same care. Auditing and quality control have long used stratified sampling (Neyman, 1934): objects are first divided into strata by risk, high-risk ones are all checked, and low-risk ones are sampled at random. Work handed to an agent can have its checking effort allocated in the same way: results that involve moving funds, changing positions or going to outside parties are confirmed item by item, while collations and summaries for internal reference are sampled in proportion. Limited checking time is then concentrated where errors cost most.
The more often an agent's results are right, the more easily the people checking them grow lax. Parasuraman and Manzey (2010) call this automation bias and complacency: people tend to accept the recommendations of automated systems, do less independent checking, and fail to notice when the system occasionally errs. The remedy comes from experience in safety inspection: regularly insert test items with known answers to measure whether checkers can still catch errors, and record the results of checking so that the checking itself can be checked.
Lowering the cost of checking raises the amplification factor more than speeding up production does: choose tasks that are easy to check, write down acceptance criteria in advance, show key actions and evidence, allocate checking by risk, and guard against checkers growing lax.
Take on first the tasks that can be checked by testing, recomputing and comparing with the source.
When results are hard to verify, either count checking as part of the effort or state clearly the risk of leaving them unchecked.
Hand off first what can be verified on the spot; confirm trade execution item by item; base advice and discretionary management on the agreement and process records.
Put the objects, period, measures and format in a checklist, and accept the result item by item against it.
Every figure and conclusion can be opened at its source, and the checker looks only at the few points relevant to the conclusion.
Confirm item by item results that involve funds and positions, and sample the rest; mix in test items regularly and record the results of checking.