Activities, not occupations

Two benchmarks now define what “economically valuable” means for AI agents. GDPval takes the sectors that contribute more than 5% of US GDP, the five best-paid digital occupations in each, and thirty tasks per occupation, 1,320 in all, each graded by blinded experts against a professional’s deliverable. Agents’ Last Exam clusters the digital occupations of the federal taxonomy into 13 domains and 55 sub-fields, sources 1,490 task instances from 250 practitioners, and scores each artifact against a verifiable reference. Both pick tasks by where work is paid, and both grade the deliverable. Neither weights a result by what the work is worth.

AERead asks a different question, so it has to pick differently. We benchmark agents’ readiness for value delivery: when an agent acts for a principal against other actors who have their own interests, does the outcome it reaches deliver value to the principal? That question is not organised by occupation. A purchasing agent, a sales manager and a lawyer sit at the same table when a supply contract is signed. It is organised by activity: what is done, with whom, and what depends on it.

A map of the economy by what is done

We built one. Every occupation in the US wage bill (BLS, May 2023: 151.9 million jobs, $9.94 trillion) is assigned to one of ten activities that firms, households and governments perform with other actors, or to “others” when the work transforms or delivers rather than deals. Shares of GDP follow, with the non-labour half of value added allocated in the same proportions; on labour alone the ten activities are 18.9%, which is the floor.

Hours and decisions

Those shares are shares of what people are paid, and for the activities we care about they understate the stakes by two orders of magnitude, because the value of a deal is not in the hours spent on it. Procurement is 0.5% of GDP as an activity; the purchases it decides come to $23 trillion a year, three-quarters of GDP: $20 trillion of inputs bought by businesses, which GDP leaves out by construction because it counts only value added, and $3.1 trillion of public procurement. Hiring is 1.0% and sets the $15.7 trillion paid in compensation.

An acquisition makes the same point in one deal. Advisory fees run near 1% of deal value ($15.4 billion worldwide in the first half of 2024). The premium negotiated at the table averaged 32 to 33% of the target’s market value for US targets between $1 and $10 billion in 2023 and 2024, and 40 to 44% below $1 billion. And the academic record says where that premium goes: at announcement the combined gain to both companies averages +1.8% of their value, targets gain 16% and acquirers lose 0.7%, so on average the buyer hands the whole anticipated gain to the seller. That is a failure of price discipline, not of the quality of the model or the deck.

There is a general reason the hours lens stays small. When the tasks of a business are complements, a chain that is only as strong as its weakest link, making any one task infinitely productive raises output by no more than that task’s share of it; Chad Jones puts it as “software is about 2% of GDP, if we had infinite software we’d be 2% richer.” Automating transaction work is bounded by its 12.4% share in the same way. A better decision is bounded by something else: the misallocation in the flow it governs.

This is the part of the economy to optimise

These flows are where a better decision has the most to work on. The $23 trillion of purchases, the $15.7 trillion of compensation and the $5.6 trillion of investment are decided by a few hundred billion dollars of judgment, so a dollar of judgment here acts on more value than a dollar of judgment anywhere else in the economy. The value is realised through allocation.

Every one of these markets leaves value on the table because the parties cannot see each other’s information. A buyer who cannot tell a good supplier from a bad one pays for the average, and the good suppliers leave. An acquirer who cannot see the target’s books overpays and later writes the asset down. A worker and a job that fit never meet because neither side could verify the other. A better decision moves the outcome toward what the market would reach if everyone knew everything: the right supplier gets the contract, the right worker the job, the asset the owner who can use it best, the risk the party able to bear it. The distance closed is the value, and it is realised through four channels. Three create it; the fourth, price, moves it.

We count the value at two levels. The market-level score is the share of the full-information optimum that the market realises; it is immune to transfers by construction and is the headline. The principal-level score is the surplus the agent obtained against the reference, within its mandate; it captures the transfer, because the principal does care who gets the premium. They are reported side by side and never added.

Each case decomposes the gap between what was realised and the optimum and names the stage where it sits, which is how m_c is measured. The sign of r_c is not automatic, since an agent that is better at extracting or withholding can raise its principal’s take while lowering what the market realises, and one that is too eager or too timid lowers both; that is why the market-level score leads and a share of every task has no deal as the right answer.

What a world can score

We keep an activity when three things hold. Its value is decided rather than produced, so the flow it governs is large against what it costs to perform. There is a counterparty with its own interest and private information, so readiness means bargaining, matching, commitment, protection and refusal rather than output quality. And the result can be scored inside a world at the end of the episode against a computable reference: surplus, cost against the optimum, feasibility, a constraint that was violated.

12.4% of GDP passes all three: procurement, the negotiated part of selling, financing terms, hiring, contracting, risk transfer, brokerage and advice. By wage bill, 82% of that is business judgment and 18% is legal and compliance work, and that is the point rather than a problem. In our worlds judgment is not graded; its consequence is measured. A model is scored by what it obtained for its principal, not by whether an expert would nod at its reasoning.

How the tasks are distributed

Tasks are weighted by the flow a class of transaction decides, not by the hours spent on it, with each transaction counted once and its two seats sharing the weight. Five classes result, and they are also the five headline categories: we do not publish an undecomposed score.

Household consumption is counted at the part a household actually deals over. Of the $20.9 trillion spent in 2025, 55% is bought at posted prices, where the decision is the seller’s; 12% is the imputed rent of owner-occupied homes, with no transaction at all; 17% is health care priced between providers and insurers. What remains, $3.2 trillion, is negotiated (leases, vehicles) or chosen on terms and claimed on (financial services and insurance). A firm’s spend is treated differently on purpose: it is decided at the contract and the automated orders execute it, so the whole flow counts.

Inside each class, five rules:

• Every class has tasks at the four stages a deal passes through: finding and qualifying the counterparty, price, structure and terms, and claims or renegotiation after signing, starting at 20, 35, 25 and 20%.
• A quarter to a third of tasks have refusing, walking away or escalating as the best move. In our own runs one model signed 264 of its 1,601 leases at zero rent; another, at its default settings, walked away from deals that were good for its principal. A set with only deal-optimal tasks rewards eagerness, and one with only traps rewards timidity.
• Every task carries the principal’s objective and authority limits, so escalation is a scoreable move rather than a way out.
• Every task records the seat the agent holds and whether the principal is a firm, a household or the state, so results decompose by seat and principal as well as by class and stage.
• Confirmatory runs face scripted, verified counterparties; the harness never judges a move.

Three ways to pick tasks