Pricing Work When Production Is Cheap

A professional fee has always bundled three things that were never separated because they arrived together. The production: the model, the draft, the research pass, the deck. The judgement: knowing which answer is right, and that this one is subtly wrong. And carrying the consequence: a name on the work and an insurer behind it.

Machines have repriced the first of those and left the other two alone. Firms that cannot say which of the three a client is paying for will find the whole fee argued down at the speed of the cheapest component, because the client has already noticed that the first one got cheap.

The production component commoditises from the bottom#

Brynjolfsson, Li and Raymond followed 5,172 customer-support agents through a staggered rollout of a conversational assistant. Resolutions per hour rose about 15 per cent on average, 30 per cent for the less skilled and less experienced, and 36 per cent in the lowest skill quintile, with no significant gain for the most skilled.

One firm, one occupation, months of output, and the authors are clear that it says nothing about whether those novices became experts. The shape still describes what happens to a fee attached to production. The gap between a cheap provider's output and an expensive one's narrows from below, and it narrows fastest in the work a junior used to do at a senior rate.

The judgement component is local and does not transfer#

Dell'Acqua and colleagues gave 758 consultants tasks inside and just outside the model's competence. Inside the frontier, assisted consultants were markedly better and faster. Outside it, they did worse than consultants using no AI at all. The study cannot tell any firm where its own frontier runs. That matters for pricing: knowing which side of the line a particular client's problem sits on is local knowledge, acquired by doing the work, and now the thing being sold.

Vaccaro, Almaatouq and Malone found human and AI combinations performing worse on average than the better of the two alone across 106 experiments, with losses concentrated in decision-making. The benchmark is an oracle-selected best performer, and the studies predate current models. For a seller of expertise it carries one commercial implication: the claim that a human reviewed it is not automatically worth paying for, and a client who has read anything about this will eventually ask what the review consisted of.

Autor's argument runs the other way and deserves stating. He makes the case that AI's distinctive opportunity is to extend the reach of expertise, letting more people perform higher-stakes decision work currently reserved to a small elite. He is explicit that this is an argument about what is possible rather than a forecast, and no evidence yet shows it at scale. If it holds, the number of people who can sell judgement rises, which changes its price as surely as cheap production does.

The third component has not moved at all#

What a client buys with a signature is somebody to answer for the outcome. That has no machine substitute and no cheaper supplier.

Regulators have been clearer about this than most firms. The Financial Reporting Council's 2025 guidance brings machine learning and generative AI inside existing audit evidence and documentation standards without writing new rules, and declines to set an explainability threshold, saying what counts as appropriate varies by context. It is guidance rather than a finding about practice and creates no new requirements. The direction is unambiguous: the obligation to evidence the work stays with the professional whatever produced it.

Elish's work on moral crumple zones describes the failure mode at the other end. Responsibility can be misattributed to a human actor who had limited control over the behaviour of an automated system, and the analogy is deliberately asymmetric: a car's crumple zone protects the driver, while this arrangement protects the system at the person's expense. It is a concept paper working from selected cases and establishes no rate. As a description of what a badly structured engagement does to the named individual inside it, it is the sharpest thing in this base.

A fee that charges for production and gives away the consequence is mispriced in the direction that hurts most slowly.

How to separate them#

• Say which component each line is. Production, judgement, or carrying it. Clients argue about the first and rarely about the third, once it is visible.
• Do not bill a saved hour as a worked one. It is the fastest way to lose the argument about the other two, and the client usually finds out.
• Price the review by what it consists of. A review somebody could not have produced the work for is a different product from one they could. The capability test belongs in the engagement letter.
• Charge for the frontier knowledge explicitly. Knowing which parts of this problem the tools handle well is the output of years of doing it, and now a deliverable rather than background.
• Keep the evidence trail in your own format. The documentation obligation does not move to the supplier, so the record has to survive a change of tool.

What would settle it#

Fee data separated by component, before and after adoption, in firms that sell expertise. Nobody publishes it, and the firms holding it have reasons not to. The evidence above is about productivity, pairing and professional obligation; applying it to pricing is an argument made on this site.

Where this sits in my own argument#

The verifier's discount explains why the judgement component gets priced below the production component even where it takes more expertise. This page is the commercial consequence: a profession that has not separated the three parts of its fee will discover the market separating them instead.

Related SuperSkills research#

On the price of checking, the verifier's discount. On the professional duty question running the other way, could a professional be negligent for not using AI? On what thins when advice gets faster, decision quality. On what a firm can actually release, how much can one firm check?

Key sources

• Brynjolfsson, E., Li, D. and Raymond, L. (2023). Generative AI at Work.
• Dell'Acqua, F. et al. (2023). Field experimental evidence on the jagged technological frontier.
• Vaccaro, M., Almaatouq, A. and Malone, T. (2024). When combinations of humans and AI are useful.
• Autor, D. (2024). Applying AI to Rebuild Middle Class Jobs.
• Financial Reporting Council (2025). AI in audit: Illustrative example and documentation guidance.
• Elish, M. C. (2019). Moral Crumple Zones.

About this research#

Written by Rahim Hirji, author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company.

How this research works · Reviewed quarterly · Found an error? Tell me and it is corrected on the page.