A task must pass three gates before it can be handed to an AI agent. First, can the agent reach it? When the result of a task lands in information or in rights and obligations, the agent can act on it directly; when it lands in a person's body or in the physical world, the agent can only help indirectly. Second, may it be done on someone's behalf? Some things must be experienced, decided or attended in person, and the law may also require the person to act personally. Third, is it worth handing over? Explaining a task carries a fixed cost, and a small job done only once is not worth it. The share of the whole task that passes all three gates sets the ceiling on the amplification factor: if the agent can do only 30% of the work, then even if that 30% is done instantly, the whole task is only a little over 40% faster. This share is not an inherent property of the task. It is determined by the conditions the agent has: working only from the material in a conversation, being connected to data and tools, and being able to act on site raise the ceiling step by step.
What is handed to an agent comes in three scales. The first is a question: the principal asks and the agent answers, both are present at the same time, and an exchange usually lasts a few minutes; the agent works only when asked, and nothing further happens afterwards. The second is a piece of work with a deliverable: the principal states the goal and leaves, the agent runs for minutes or hours, and when it is done the principal comes back to accept the result. The third is an ongoing responsibility, such as looking after a set of assets over a long period: at agreed times, or when something changes outside, the agent checks, records and, where necessary, reports, over months or even years. Research on collaborative work distinguishes synchronous from asynchronous collaboration (Johansen, 1988): meetings and phone calls are synchronous, while email and shared documents are asynchronous. The three scales are, in turn, synchronous, bounded asynchronous and ongoing asynchronous.
As the scale grows, the principal's scarcest resource changes. When handing over a question, the scarce resource is waiting time, so the answer must be fast. When handing over a piece of work, the scarce resource is trust: what is handed over must get done. When handing over an ongoing responsibility, most contacts are initiated by the agent, and the scarce resource is the principal's attention. Ongoing asynchronous work therefore brings three requirements. Interruptions must be tiered, so that most checks are only recorded and do not interrupt. The past must be verifiable, with records stating the state of things at the time and the basis for each step. The future must be visible, so that the principal knows in advance when a decision will be required. The three scales nest within one another: a responsibility contains many pieces of work, and a piece of work contains many questions. A larger scale is not more advanced. A single query needs no long-term relationship; only things that need long-term care and a return under set conditions do.
A simple way to tell tasks apart is to ask what their result ultimately changes. By this standard, human affairs fall into five kinds: changing one's own body and mind, such as exercise and study; changing other people's bodies, minds and relationships, such as care and negotiation; changing things and the environment, such as cooking and repair; changing information, such as writing reports and doing analysis; and changing rights and obligations, such as signing contracts, transferring money and placing orders. Autor, Levy and Murnane (2003) classified work tasks as routine or non-routine and cognitive or manual in order to explain which work computers replaced. Classifying by where the result lands answers more directly whether an agent can act on the task.
The inputs and outputs of a large language model are information. With tools added, it can also submit forms, call interfaces and execute transactions on someone's behalf, that is, change rights and obligations. In tasks that change the body and the physical world, it can take part only indirectly: by giving advice, producing checklists and arranging schedules. Eloundou and colleagues (2023) assessed the work tasks of US occupations one by one and estimated that about 80% of workers have at least 10% of their tasks that large language models could speed up substantially, and about 19% have more than half of their tasks in this position; once callable tools are taken into account, the share of tasks that can be substantially sped up rises further. The reachable range widens as tools are added.
Whenever part of a task is beyond the agent, that part is still done by the principal. Let the share the agent can take on be a, and let it cut the effort for that share to 1/k of the original. The overall amplification factor is then 1 ÷ [(1 − a) + a ÷ k]. This has the same form as Amdahl's law (1967) in computer architecture: however large k is, the amplification factor never exceeds 1 ÷ (1 − a). When the agent can take on 30% of the work, the ceiling is about 1.4 times; when it can take on 90%, the ceiling is 10 times. Raising the applicable share often matters far more than making the agent faster.
The applicable share sets the ceiling
Drag the share the agent can take on and choose how many times it shortens that effort, to see the overall amplification factor. The dashed line is the ceiling if the agent finished instantly.
The share of a job's time that an agent can reach is called the reachable share. It is determined by the conditions the agent has, which fall into three tiers: a conversational model that answers only from the material given in the conversation, called Chat below; an agent connected to the data and systems of the principal and the organization, able to run code and to handle business with outside parties, called Agent below; and, beyond that, an embodied agent that can perceive and act on site. In addition, many tasks require access: authorization, licences and custody of assets determine whether the part an agent can reach may lawfully be carried out.
Using work-activity data from the US Occupational Information Network (O*NET) and the employment structure of EU industries, the author estimated the reachable share and the ceiling for each industry and occupation under the three tiers (author's estimate). With Chat alone, the reachable share across industries is 6% to 21%, and the ceiling for a whole job is only 1.1 to 1.3 times. As an Agent, it rises to 22% to 45%, averaging 36% across all industries, with a ceiling of 1.55 times; finance, professional services and information and communication gain the most, with ceilings of about 1.8 times. In construction, agriculture, forestry and fishing, and household production for own use, physical work and operation take up about half of working time, and the ceiling for a digital agent is only about 1.3 times; with embodied capability it rises to 2.7 to 2.8 times, overtaking the knowledge-intensive industries. The ceiling of the knowledge-intensive industries comes from elsewhere: communicating with, coordinating and persuading others takes up about 30% of their time, and the person's own judgements and decisions are also hard to hand over, so embodied capability helps them little. The differences between occupations are larger still: under the Agent tier, the ceiling for accountants and customer service representatives is about 2.8 times, for lawyers and general managers about 1.7 times, and for construction labourers and housekeeping cleaners only about 1.1 times.
Place 332 work activities on two axes: the horizontal axis is the activity's share of all working time, which answers whether it is worth doing; the vertical axis is the share an Agent can take on, which answers whether the conditions to do it exist (author's estimate). The 54 activities that are both worth doing and within reach together account for about 31% of working time, mainly recording, paperwork, providing information and communication, such as maintaining operational records, providing information to customers and creating visual designs. The 66 activities that are worth doing but lack the conditions account for about 38%. One group is physical operation such as cleaning, moving and maintenance, which lacks embodied capability; the other is directing organizational operations and supervising other people's work, which require the person to be present or to act personally.
The 92 activities that are within reach but small individually account for about 13%, and most finance-related activities are here: analysing business and financial data and analysing market and industry conditions have a reachable share of about 83%, and monitoring financial data and activities about 75%. In paid work they are concentrated in a few occupations, so each accounts for a small share; for individuals managing their own assets, however, they are the main effort.
Being able to do something does not mean it may be done on someone's behalf. Some things are valuable because the person experiences them: exercise, learning and companionship lose their meaning when done by someone else. Some things must be decided by the person: marriage, voting and major financial commitments. Some things require the person to be present: testifying in court and signing in person. Others must, by law, be done by the person: signing legally binding documents, answering questions truthfully during account opening and suitability assessment, and confirming that the risks of a product have been understood. In financial products, an agent can prepare, explain and execute, but authorization, confirmation and acknowledgement of risk must be left to the person.
Handing a task to someone else first incurs a fixed cost: explaining the goal and constraints, supplying the necessary information and confirming that the other party has understood. Coase (1937) used transaction costs to explain why firms exist: when getting something done through the market costs more than doing it in-house, people do it themselves. Handing work to an agent is the same. For a small job done only once, explaining it may cost more than doing it oneself; for something that recurs, one clear explanation can be used many times, and the fixed cost is spread out. Ongoing tasks are therefore the kind most worth handing over.
The fixed cost of explaining a task
The first time a task is handed to an agent, time must be spent explaining the goal, constraints and information needed; each time after that saves some time compared with doing it oneself. Drag the parameters to see after how many repetitions it starts to pay off.
Apply the three gates to the common tasks of an individual investor. Looking up data, explaining an event, researching a company, and screening and comparing all have results that land in information: the agent can reach them and may do them on the investor's behalf, so the applicable share is high. Ongoing tracking recurs and is the most worth handing over. Making a plan and placing orders, and managing a portfolio against goals, change rights and obligations: the agent can execute them, but every step needs the person's authorization and confirmation. The result of a review lands in the person's own understanding: the agent can organize the facts and break down the causes, but the conclusion must be the person's own. The tasks with the highest applicable share are precisely the information tasks that matter less for long-term results; allocation and execution, which matter most, are exactly the parts that need authorization.
Select a row.
Within the finance and insurance industry, the height of the ceiling depends on whether an occupation is close to processing information or close to deciding rights and obligations and facing clients (author's estimate). Review, recording and analysis occupations, such as credit analysts, bank tellers and quantitative analysts, have ceilings of about 3 to 3.7 times. Occupations that directly decide the disposal of assets or face clients, such as fund managers, financial advisers, actuaries and underwriters, have only 1.6 to 1.9 times. Even with embodied capability, the latter's ceiling barely changes: the limit comes from authorization and trust, not from tools. At the task level (task wording as in O*NET), the share that can be handed to an Agent is 0.82 to 0.90 for a securities broker's "Keep accurate records of transactions", "Monitor markets or positions" and "Report all positions or trading results", but only 0.25 for "Make bids or offers to buy or sell securities" and "Agree on buying or selling prices at optimal levels for clients"; for a fund manager's "Select specific investments or investment mixes for purchase by an investment fund" and "Select or direct the execution of trades", it is only 0.18 to 0.25.
Review, recording and analysis occupations have ceilings of about 3 to 3.7×; occupations that directly decide the disposal of assets or face clients, such as fund managers, financial advisers, actuaries and underwriters, have only 1.6 to 1.9×.
When investors manage their own assets, each investment task corresponds to a task of some professional: ongoing monitoring corresponds to the broker's "Monitor markets or positions", research to the fund manager's "Perform or evaluate research, such as detailed company or industry analyses", and review to the broker's "Report all positions or trading results". Using the author's assumed annual time spent on each task and converting with the reachable shares of the professionals' tasks, the overall ceiling is about 3.8 times for an allocation-oriented investor and about 4.8 times for a trader; on the same basis, fund managers, financial advisers and brokers have only 1.8 to 2.4 times (author's estimate).
Individual investors come out higher than professionals because the principal is the person themselves. A financial adviser spends about two-thirds of working time communicating with clients, explaining plans and recommending products, and an agent can hardly take this over; when investors manage their own assets, this work does not exist, and their effort goes almost entirely into watching the market, research, planning and review. Ongoing monitoring takes the largest share: under the author's assumptions, about 43% of an allocation-oriented investor's investment effort and about 73% of a trader's go into ongoing monitoring. Taking the trader as an example, if monitoring is sped up 10 times and the other tasks 2 times, the whole is about 4.8 times faster; the other way round, only about 2.6 times. Ongoing monitoring is also the kind of professional task an agent can best take over, provided the agent can keep running online under the agreed conditions.
First see which scale is being handed over; then look for tasks with a high applicable share that recur; supply the data, tools and access the agent needs; and, at the steps that require authorization, leave confirmation to the person.
Tiered interruptions, a verifiable past and a foreseeable future; a single query needs no long-term relationship.
An agent can directly change the results of these two kinds of task; for tasks involving the body and the physical world, it only assists.
Filling in, one by one, the steps of a task the agent cannot do raises the overall factor more than further speeding up production.
Account and execution data are context, trading channels are tools, and licences and custody are access; ongoing monitoring is where investors spend the most effort and what an agent can best take over.
The agent prepares and executes; how much to invest, whether to place an order and whether the risks are understood are decided by the person, with a record kept.
Let goals and constraints explained once serve for a long time; ongoing tracking tasks are the most worth handing over.